betaFeedback

WorksComputer Science

HARE: Fast Read Disturbance Threshold Profiling in Modern DRAM Chips via Reduced Timing Constraints

Abstract

Modern DRAM chips are vulnerable to read disturbance, which manifests as bitflips in a victim row physically near an aggressor row that is repeatedly activated. State-of-the-art mitigations rely on accurate and exhaustive profiling of the read disturbance threshold (RDT), i.e., the minimum number of aggressor row activations needed to induce the first bitflip. However, RDT changes significantly and unpredictably over time, motivating online profilers that dynamically adjust a mitigation's RDT configuration. RDT profiling is costly for online profilers (e.g., hundreds of seconds for a 16-bank DDR4 module) because 1) measuring a single DRAM row's RDT requires many activations and 2) each DRAM row can have a different RDT, requiring many rows to be profiled. We introduce HARE, an online RDT profiling technique that intentionally exacerbates read disturbance to quickly profile a DRAM module. HARE's key idea is to reduce the timing constraints between consecutive DRAM commands during profiling. Doing so significantly reduces profiling time by 1) directly reducing the execution time of the profiling command sequence and 2) exacerbating read disturbance such that the first bitflip occurs at a smaller activation count than with the conventional RowHammer sequence. From our characterization of 180 real DDR4 DRAM chips from four major DRAM manufacturers, we make three observations. First, HARE measures a DRAM row's RDT 31.2% faster than the RowHammer profiler and a DRAM module's minimum RDT with an average reduction of 21.7%. Second, HARE flips more than 95% of the cells that the RowHammer pattern flips. Third, HARE preserves the RDT ranking of rows well enough such that it can exhaustively reprofile a small set of rows to improve its minimum RDT estimation. Our end-to-end evaluations show that HARE enables significant system performance improvements and outperforms the RowHammer profiler.

Modern DRAM chips are subject to read disturbance errors. These errors manifest as security-critical bitflips in a victim DRAM row physically nearby an aggressor row that is repeatedly activated (opened) (RowHammer) or kept open for a long time (RowPress). To prevent read disturbance bitflips securely and with low overhead, state-of-the-art read disturbance mitigations rely on accurate and exhaustive profiling of the read disturbance threshold (RDT) (i.e., the minimum number of aggressor row activations needed to induce the first bitflip). However, RDT is not static: it changes significantly and unpredictably over time, motivating online profiling mechanisms that dynamically adjust a mitigation mechanism’s RDT configuration. Unfortunately, exhaustively profiling a DRAM module’s RDT is costly for such online profilers (e.g., hundreds of seconds for a 16-bank DDR4 module) for two main reasons. First, measuring the RDT of a single DRAM row requires activating its aggressor rows many times (e.g., 4K times for DDR4 modules) before a bitflip is observed. Second, due to manufacturing process variation, each DRAM row can have a different RDT, requiring many rows to be profiled (e.g., 65’536 rows per bank).

Our goal is to design a fast online RDT profiling technique that identifies the minimum RDT across a DRAM module at low cost. To this end, we introduce HARE, an online RDT profiling technique that intentionally exacerbates read disturbance to quickly profile a DRAM module. HARE’s key idea is to reduce the timing constraints between consecutive DRAM commands during profiling. Doing so significantly reduces profiling time by 1) directly reducing the execution time of the profiling command sequence and 2) exacerbating read disturbance such that the first bitflip occurs at a smaller activation count than with the conventional RowHammer sequence.

From our detailed characterization of 180 real DDR4 DRAM chips from four major DRAM manufacturers, we make three key observations. First, HARE measures the RDT of a DRAM row 31.2% faster than the RowHammer profiler and the minimum RDT of a DRAM module, with an average reduction of 21.7%. Second, HARE flips more than 95% of the cells that the RowHammer access pattern flips. Third, HARE preserves the RDT ranking of rows well enough to exhaustively reprofile a small set of rows to measure the minimum RowHammer RDT. Our end-to-end evaluations show that HARE enables significant system performance improvements and outperforms the RowHammer profiler.

Introduction

Memory isolation is a fundamental property for system robustness (i.e., reliability, security, and safety). Accesses to one memory location should not induce unintended side-effects on data stored in other (unaccessed) memory locations. Unfortunately, modern DRAM chips are subject to read disturbance, where accessing a DRAM row induces bitflips in physically nearby DRAM rows that are not accessed, breaking memory isolation. RowHammer [1] and RowPress [2] are two prominent read disturbance phenomena, in which a victim DRAM row experiences bitflips when a nearby aggressor row is repeatedly activated (i.e., hammered) or kept open for a long period of time (i.e., pressed), respectively.

Read disturbance worsens as DRAM manufacturing technology scales down to smaller node sizes [2, 3, 4]. For example, the number of aggressor row activations needed to induce the first bitflip has reduced by more than an order of magnitude in less than a decade [3]. Therefore, a large body of work proposes read disturbance mitigation mechanisms. To prevent read disturbance bitflips securely and with low overhead, many of these read disturbance mitigation mechanisms [1, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70] rely on accurate knowledge of the read disturbance threshold (RDT), i.e., the number of aggressor row activations needed to induce the first bitflip in a victim row. A row’s RDT is typically profiled by activating its aggressor rows repeatedly until a bitflip is observed and determining the actual hammer count with a binary search approach. The mitigation mechanisms are configured for an RDT value (e.g., minimum profiled RDT value, or the worst-case RDT value) and perform preventive actions (e.g., refreshing victim rows) before a row’s activation count reaches the RDT value.

Unfortunately, the RDT of a DRAM row is not static; it changes significantly and unpredictably over time [71]. Thus, a row can experience bitflips at lower activation counts compared to the previously profiled RDT value, resulting in any mitigation configured for that RDT to fail preventing attacks. To mitigate bitflips in the presence of variable read disturbance effect, mitigation mechanisms require dynamic adjustments to the RDT through online RDT profiling.

RDT profiling is prohibitively slow for two main reasons. First, measuring the RDT of a single DRAM row requires activating its aggressor rows many times (e.g., at least 4K times for DDR4 chips (§5)) before a bitflip is observed. Second, because of manufacturing process variation, each DRAM row can have a different RDT (i.e., can fail at different aggressor row activation counts), requiring many rows to be profiled. Thus, profiling the RDT of many rows (e.g., 65’536 rows per DRAM bank in a 16-bank DDR4 module [72]) takes significant time (e.g., hundreds of seconds [71]).

Our goal is to design a fast RDT profiling technique that identifies the minimum RDT of a DRAM module at low cost. To this end, we introduce HARE, a fast RDT profiling technique that intentionally exacerbates read disturbance to quickly profile a DRAM module. HARE’s key idea is to reduce the timing constraints between consecutive DRAM commands in the profiling sequence. Doing so reduces the profiling time significantly by 1) reducing the time it takes to perform one hammer, and 2) exacerbating the read disturbance to induce bitflips with far fewer hammers than conventional RowHammer patterns.

In this paper, we present the first analysis of read disturbance under reduced timing parameters as a fast RDT profiling technique, on 180 commercial off-the-shelf (COTS) DDR4 DRAM chips from four major DRAM manufacturers (in 25 DRAM modules). Based on our characterization, we make three key observations that form the basis of HARE’s design.

First, hammering a row with reduced timing constraints, which we call fastHammer, reduces the row’s RDT in all tested real DRAM chips from all four major manufacturers (by 27.0%–33.5% depending on the manufacturer). Thus, fastHammer induces the first bitflip with fewer aggressor row activations than the conventional RowHammer pattern. Second, fastHammer induces bitflips in more than 95% of the cells that the conventional RowHammer pattern flips, at every tested configuration. Therefore, fastHammer covers nearly all the cells that RowHammer disturbs. Third, fastHammer preserves the ordering of rows by RDT: in all 25 tested modules, the RowHammer-weakest row falls within the top 5.6% of rows ranked by their fastHammer RDT. Therefore, a module’s minimum RowHammer RDT can be identified by exhaustively profiling only a small candidate set of rows.

HARE builds on these observations in two variants. HARE-Direct profiles every row in the module with fastHammer and reports the lowest measured RDT. Because fastHammer lowers a row’s RDT, HARE-Direct’s estimate is conservative, and thus, inherently applies a guardband to the mitigation’s configured RDT. HARE-Filter improves the accuracy of this estimate at the cost of additional profiling time: it ranks a batch of rows by their fastHammer RDT, reprofiles only the top-ranked rows with the RowHammer pattern, and reports the minimum RowHammer RDT among them. We implement both variants as periodic online profilers that profile all rows of a module within a profiling window and continuously update the RDT configuration of the underlying RowHammer mitigation.

We evaluate HARE against two baselines. To quantify the benefit of online RDT profiling, we compare HARE to a system with no online profiler using a probabilistic error model of read disturbance bitflips under variable read disturbance, built from our characterization results. Across 1M Monte Carlo trials, a system with no online profiler experiences an uncorrectable error (i.e., a read disturbance bitflip pattern that the DRAM chip’s on-die ECC cannot correct) in 99.9% of the trials, whereas HARE-Direct reduces this probability to 6.2% and HARE-Filter reduces it by up to 12.1×. To quantify the benefit of fastHammer, we compare HARE to a system with a conventional RowHammer profiler using Ramulator 2.0, a cycle-level DRAM simulator, across quad-core memory-intensive workload mixes drawn from the SPEC CPU2006, SPEC CPU2017, TPC, MediaBench, and YCSB suites. Over 36 configurations spanning three state-of-the-art RowHammer mitigations (i.e., PRAC [73], Graphene [23], and PARA [1]), three RDT values, and four profiling intervals, HARE-Direct and HARE-Filter improve geomean weighted speedup over the conventional RowHammer profiler in every configuration, by 0.9% to 54.4% and 1.5% to 47.5%, respectively.

This paper makes the following key contributions:

  • We present the first characterization of read disturbance under reduced DRAM timing constraints as a method for fast RDT profiling, using 180 real DDR4 DRAM chips (25 modules) from four major DRAM manufacturers.

  • We show that fastHammer induces bitflips in more than 95% of the cells that the conventional RowHammer access pattern flips, and that it preserves the ordering of rows by RDT well enough: the RowHammer-weakest row of every tested module falls within the top 5.6% of fastHammer-ranked rows. A module’s minimum RowHammer RDT can therefore be identified by exhaustively profiling only a small candidate set of rows.

  • We propose HARE, the first online RDT profiling technique that intentionally exacerbates read disturbance to profile a DRAM module quickly, in two variants: HARE-Direct, which reports a conservative minimum RDT, and HARE-Filter, which recovers the conventional RowHammer RDT by reprofiling only the top-ranked rows.

  • We develop a probabilistic error model for read disturbance bitflips under variable read disturbance, parameterized by our characterization data, and use it to show that a mitigation configured by one-time profiling cannot operate reliably: it exhibits an uncorrectable error in 99.9% of simulated systems, with a mean time to failure of only 8.5 minutes, and reaching a negligible error probability requires halving the configured RDT.

  • We show that HARE substantially improves both system reliability and performance: it reduces the probability of an uncorrectable error from 66.8% to 0.67% compared to a system with no online profiler at the same 20% guardband, and outperforms a conventional RowHammer profiler in every evaluated configuration, by up to 54.4%.

Background

DRAM Organization. Figure 1 shows DRAM-based main memory organization. In a CPU, a memory controller interfaces with a DRAM channel (containing DRAM module(s)) to serve memory requests. Each module contains multiple DRAM ranks (i.e., a number of DRAM chips operating in lockstep), with each DRAM chip comprising multiple DRAM banks. A DRAM bank comprises a two-dimensional array of DRAM cells organized into columns and rows. Each DRAM cell stores a single bit of information as charge stored in a capacitor. To access the cell, an access transistor controlled by a wordline connects the cell to a bitline. When the wordline is asserted, the access transistors connect all cells of a row to their corresponding bitline, perturbing their voltage. A structure called the row buffer, consisting of an array of sense amplifiers, senses and amplifies this voltage perturbation.

DRAM organization

Figure 1. DRAM organization

DRAM Operation. To access data, the memory controller first issues an activate (ACT) command, which “opens” the corresponding DRAM row and then can read or write data. Before accessing another row in the same bank, the memory controller must first “close” the currently open row by issuing a precharge (PRE) command. The memory controller schedules DRAM commands according to timing constraints [74, 75, 76], such as tRASt_{RAS} (i.e., the minimum time between opening a row with an ACT command and closing the row with a PRE command), tRPt_{RP} (i.e., the minimum time between sending a PRE command and opening a row with an ACT command) and tRCt_{RC} (i.e., the minimum delay between two consecutive ACT commands to the same bank).

DRAM Read Disturbance. RowHammer and RowPress occur when a row is (i) repeatedly opened and closed, i.e., hammered, or (ii) kept open for a prolonged period, i.e., pressed, causing bitflips in adjacent rows. The number of times the “aggressor” row must be hammered or pressed to induce bitflips in adjacent “victim” rows is called the read disturbance threshold (RDT). DRAM technology node scaling has increased the severity of read disturbance phenomena [1, 3, 4, 26, 56, 77, 78, 79, 80]. Many prior works [1, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70] have proposed various mitigation techniques that prevent read disturbance bitflips, particularly those caused by RowHammer. To ensure robust DRAM operation, all of these techniques require accurate knowledge of the RDT.

Motivation and Goal

Variable Read Disturbance. RDT of a DRAM row is not static: it changes significantly and unpredictably over time [71]. Thus, a row can experience bitflips at lower activation counts compared to the previously profiled RDT value, causing any mitigation configured for that RDT to fail to prevent bitflips. Figure 2 shows the change in a single victim row’s RDT in three representative modules across profiling iterations.1 Each iteration measures the row’s RDT once by executing the profiling routine. Each subplot depicts the histogram of a victim row’s RDT in one module, separated into 10 bins. We observe that each row’s RDT significantly changes across repeated measurements, similar to [71].

RDT value histograms for a single victim row in three tested modules

Figure 2. RDT value histogram for a single victim row in each tested module.

A RowHammer mitigation is typically configured for a minimum RDT across all rows in a module. We analyze how the minimum RDT changes across profiling iterations in three representative modules from three manufacturers. In each profiling iteration, we measure the RDT of 3K rows per module. Figure 3 plots our measurements across 2K iterations. Each box represents a bin of 200 iterations (whiskers: the highest and the lowest minimum RDT in the bin). The solid line depicts the running minimum RDT of the module. The dashed line depicts the minimum RDT across 2K iterations.

Minimum RDT across three representative modules

Figure 3. Minimum RDT of a module across three representative modules tested for 2K profiling iterations.

We observe that the minimum RDT changes across iterations and can be measured as late as the 1’539th iteration out of 2K iterations (in M4). We conclude that the minimum RDT across rows varies significantly across iterations.

Spatial Variation in RDT across DRAM Rows. Prior works [4, 65] show that RDT varies significantly and irregularly across DRAM rows. Figure 4 shows the RDT distributions of row groups of increasing size, from 100 to 9K rows, where each group includes the rows of smaller groups. The red mark shows the minimum RDT measured in the selected row group and the dashed line shows the minimum RDT measured across all groups. We observe that the minimum RDT does not stabilize at a size under 9K across all tested modules.

Minimum RDT of row groups with increasing size

Figure 4. Minimum RDT of a DRAM row group with increasing size

RDT Profiling is Time Consuming. To mitigate bitflips in the presence of VRD, RDT profiling is prohibitively slow for two main reasons. First, measuring the RDT of a single row requires activating its aggressor rows many times (e.g., at least 4K times for DDR4 chips [71]) before a bitflip is observed. Second, profiling one row is not enough. Due to manufacturing process variation each row can have a different RDT [65]. A RowHammer mitigation is typically configured for the minimum RDT across the module (including those that are used and standardized by industry, e.g., PRAC [73, 81]), so a profiler must visit every row (e.g., 65’536 rows per bank in a 16-bank DDR4 module). Thus, profiling all rows of a DRAM module takes hundreds of seconds [71]. This time is prohibitively long for an online profiler because (i) at the time of RDT profiling, DRAM bandwidth cannot be utilized for demand memory requests, and (ii) enabling reliable operation with low performance overhead requires frequent profiling (as we show in § 7). Our goal is to design a fast online RDT profiling technique that identifies the minimum RDT of a DRAM module at low cost.

COTS DRAM Testing Methodology

COTS DRAM Testing Infrastructure

We conduct COTS DRAM chip experiments using DRAM Bender [82, 83], an FPGA-based DDR4 testing infrastructure that provides precise control of DDR4 commands. Figure 5 shows our experimental setup that consists of four main components: 1) a host machine that generates the test program, 2) an FPGA board [84], programmed with DRAM Bender, 3) a temperature sensor and heater pads to maintain a target temperature, and 4) a temperature controller [85] that keeps the temperature at the desired level.

Image not extracted

Figure 5. Our DRAM Bender [83] based experimental setup.

Eliminating Interference Sources. To observe read disturbance bitflips at the circuit level, we eliminate potential sources of interference, by taking three measures, similar to the methodology used by prior works [2, 3, 4, 28, 77]. First, we disable periodic refresh during the execution of test programs to prevent potential on-DRAM-die TRR mechanisms [26, 28] from refreshing victim rows so that we can observe the DRAM chip’s behavior at the circuit-level. Second, we strictly bound the execution time of test programs within the refresh window of the DRAM chips to avoid data retention failures interfering with read disturbance failures. Third, we verify that the modules and chips have neither rank-level nor on-die ECC [86, 87]. With these measures, we directly observe and analyze all bitflips without interference.

COTS DDR4 DRAM Chips Tested

Table 1 provides the 180 (25) COTS DDR4 DRAM chips (modules). We label each module by its manufacturer and an index, i.e., H1–H6 (SK Hynix), M1–M10 (Micron), S1–S7 (Samsung), and N1–N2 (Nanya).

Logical-to-Physical Row Mapping. DRAM manufacturers use mapping schemes to translate logical (memory-controller-visible) addresses to physical row addresses [1, 26, 74, 86, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99]. To account for in-DRAM row address mapping, we reverse engineer the physical row address layout in all chips, following the prior works’ methodology [2, 3, 4, 77].

Chip Mfr.#Modules#ChipsDie Rev.DensityOrg.
SK Hynix324A8Gbx8
SK Hynix18D8Gbx8
SK Hynix216C16Gbx8
Micron216R8Gbx8
Micron416E16Gbx16
Micron432F16Gbx8
Samsung14C4Gbx16
Samsung18E4Gbx8
Samsung540B16Gbx8
Nanya216C8Gbx8

Table 1. Summary of DDR4 DRAM chips tested.

Fast RDT Profilingvia Reduced Timing Constraints

Hammering with Reduced Timing Constraints

In this work, we construct a fast RowHammer access pattern, fastHammer, that violates the timing constraints between DRAM commands in the hammer sequence. Fig. 6 shows (a) the double-sided attack layout in DRAM, (b) the conventional RowHammer double-sided access pattern and (c) the fastHammer double-sided access pattern with reduced precharge (PRE)-to-activate (ACT) latency.2

Double-sided RowHammer and fastHammer access pattern

Figure 6. Double-sided RowHammer and fastHammer access pattern

Fig. 6(a) shows a double-sided attack layout where a victim row is sandwiched between two aggressor rows. The conventional RowHammer access pattern (Fig. 6(b)) consists of three steps. First, we issue an ACTACT command to aggressor 1 (❶) and wait for the latency of fully restoring a DRAM cell’s charge (tRASt_{\mathrm{RAS}}). Second, we issue a PREPRE command and wait for tRPt_{\mathrm{RP}} (❷) to ensure the DRAM bank is precharged. Third, we issue an ACTACT command to aggressor 2 (❸) and wait for tRASt_{\mathrm{RAS}} before issuing a PREPRE command to precharge the bank (❹). These three steps comprise one hammer cycle. Repeating the hammer cycle results in read disturbance bitflips in the victim row. The fastHammer access pattern (Fig. 6(c)) violates tRPt_{\mathrm{RP}} and activates aggressor rows in quick succession. This results in directly reducing the time to complete one hammer cycle.

Opportunity 1. fastHammer can enable fast RDT profiling by reducing the time to perform one hammer cycle.

RDT Reduction with fastHammer

RDT Reduction per DRAM Row. We analyze the RDT of rows when performing double-sided RowHammer and fastHammer with reduced PRE-to-ACT latency. Fig. 7 shows the RDT reduction observed across all tested real DRAM chips from four major manufacturers as subplots in a box-and-whiskers plot. The x-axis shows the PRE-to-ACT latency in nanoseconds from 7.5 ns to 12 ns.3 The y-axis shows the RDT reduction in percentage compared to the RDT values of the rows with nominal PRE-to-ACT latency, i.e., a conventional RowHammer double-sided access pattern.

RDT reduction with reduced PRE-to-ACT latency

Figure 7. RDT reduction with reduced PRE-to-ACT latency.

Observation 1. Reducing the PRE-to-ACT latency reduces the per-row RDT, on average, across all modules from all manufacturers.

We observe that fastHammer with 7.5 ns PRE-to-ACT latency reduces a row’s RDT by 33.5%, 28.4%, 30.81%, and 27.0%, on average, for SK Hynix, Micron, Samsung and Nanya modules, respectively. We hypothesize that reducing the PRE-to-ACT latency leaves more electrons emitted from traps near the just-closed aggressor wordline available when the opposite aggressor wordline turns on. The electric field generated by the opposite aggressor redirects more of these electrons into the victim cell’s storage node, increasing charge loss [109, 110]. 3D TCAD simulations of double-sided RowHammer align with our observations: within the short-latency regime tested in our experiments, reducing the PRE-to-ACT latency leaves more emitted electrons near the just-closed aggressor wordline during the opposite aggressor’s pulse-on phase, thereby reducing RDT [110].

Observation 2. As the PRE-to-ACT latency decreases, the RDT reduction increases significantly.

For example, we observe an average RDT reduction of 17.3%, 19.6%, 26.0% and 33.5% across the SK Hynix modules, at 12 ns, 10.5 ns, 9 ns and 7.5 ns, respectively.

Opportunity 2. fastHammer can enable fast RDT profiling by reducing the number of hammers to induce a bitflip per row, on average.

RDT Reduction in the minimum RDT per DRAM Module. Figure 8 shows the per module minimum RDT values with reduced PRE-to-ACT latency for four major manufacturers. The x-axis shows the all tested real DRAM modules from each tested manufacturer. The y-axis shows the minimum RDT across all rows within the tested DRAM module on a log scale. Each bar shows one PRE-to-ACT latency value: the latency values with the largest average RDT reductions per row (i.e., 7.5 ns and 9 ns) and the nominal latency.

Per module minimum RDT values with reduced PRE to ACT latency

Figure 8. Per module minimum RDT values with reduced PRE to ACT latency.

Observation 3. The minimum RDT of each tested module decreases with decreasing PRE-to-ACT latency.

We observe that the minimum RDT observed in a module reduces by 21.7% and 28.3% at 9 ns and 7.5 ns, respectively, across all modules from all manufacturers.

Opportunity 3. fastHammer can identify a DRAM module’s minimum RDT with a reduction.

Detailed Analysis of fastHammer

To further understand the read disturbance behavior with fastHammer, we extensively characterize representative modules from four major manufacturers and analyze (i) the RDT distribution, (ii) the overlap of read disturbance-vulnerable cells, and (iii) the maintaining of the RDT ranking across rows.

RDT Distribution Shift with fastHammer. Figure 9 shows the distribution of RDTs across all tested rows in four representative modules as subplots. The x-axis shows the RDT of an individual row, and the y-axis shows the probability density across the selected module’s rows. Each line shows a different fastHammer latency ranging from 7.5 ns to 15 ns and RowHammer.

Per module RDT distribution with reduced PRE to ACT latency

Figure 9. Per module RDT distribution with reduced PRE to ACT latency.

@page 5/15 starts_mid=0 ends_mid=1

Double-sided RowHammer and fastHammer access pattern

Figure 6 (PDF p. 5). Double-sided RowHammer and fastHammer access pattern

Fig. 6(a) shows a double-sided attack layout where a victim row is sandwiched between two aggressor rows. The conventional RowHammer access pattern (Fig. 6(b)) consists of three steps. First, we issue an ACTACT command to aggressor 1 (❶) and wait for the latency of fully restoring a DRAM cell’s charge (tRASt_{\mathrm{RAS}}). Second, we issue a PREPRE command and wait for tRPt_{\mathrm{RP}} (❷) to ensure the DRAM bank is precharged. Third, we issue an ACTACT command to aggressor 2 (❸) and wait for tRASt_{\mathrm{RAS}} before issuing a PREPRE command to precharge the bank (❹). These three steps comprise one hammer cycle. Repeating the hammer cycle results in read disturbance bitflips in the victim row. The fastHammer access pattern (Fig. 6(c)) violates tRPt_{\mathrm{RP}} and activates aggressor rows in quick succession. This results in directly reducing the time to complete one hammer cycle.

Opportunity 1. fastHammer can enable fast RDT profiling by reducing the time to perform one hammer cycle.

RDT Reduction with fastHammer

RDT Reduction per DRAM Row. We analyze the RDT of rows when performing double-sided RowHammer and fastHammer with reduced PRE-to-ACT latency. Fig. 7 shows the RDT reduction observed across all tested real DRAM chips from four major manufacturers as subplots in a box-and-whiskers plot. The x-axis shows the PRE-to-ACT latency in nanoseconds from 7.5 ns to 12 ns.3 The y-axis shows the RDT reduction in percentage compared to the RDT values of the rows with nominal PRE-to-ACT latency, i.e., a conventional RowHammer double-sided access pattern.

RDT reduction with reduced PRE-to-ACT latency

Figure 7 (PDF p. 5). RDT reduction with reduced PRE-to-ACT latency.

Observation 1. Reducing the PRE-to-ACT latency reduces the per-row RDT, on average, across all modules from all manufacturers.

We observe that fastHammer with 7.5 ns PRE-to-ACT latency reduces a row’s RDT by 33.5%, 28.4%, 30.81%, and 27.0%, on average, for SK Hynix, Micron, Samsung and Nanya modules, respectively. We hypothesize that reducing the PRE-to-ACT latency leaves more electrons emitted from traps near the just-closed aggressor wordline available when the opposite aggressor wordline turns on. The electric field generated by the opposite aggressor redirects more of these electrons into the victim cell’s storage node, increasing charge loss [109, 110]. 3D TCAD simulations of double-sided RowHammer align with our observations: within the short-latency regime tested in our experiments, reducing the PRE-to-ACT latency leaves more emitted electrons near the just-closed aggressor wordline during the opposite aggressor’s pulse-on phase, thereby reducing RDT [110].

Observation 2. As the PRE-to-ACT latency decreases, the RDT reduction increases significantly.

For example, we observe an average RDT reduction of 17.3%, 19.6%, 26.0% and 33.5% across the SK Hynix modules, at 12 ns, 10.5 ns, 9 ns and 7.5 ns, respectively.

Opportunity 2. fastHammer can enable fast RDT profiling by reducing the number of hammers to induce a bitflip per row, on average.

RDT Reduction in the minimum RDT per DRAM Module. Figure 8 shows the per module minimum RDT values with reduced PRE-to-ACT latency for four major manufacturers. The x-axis shows the all tested real DRAM modules from each tested manufacturer. The y-axis shows the minimum RDT across all rows within the tested DRAM module on a log scale. Each bar shows one PRE-to-ACT latency value: the latency values with the largest average RDT reductions per row (i.e., 7.5 ns and 9 ns) and the nominal latency.

Observation 3. The minimum RDT of each tested module decreases with decreasing PRE-to-ACT latency.

We observe that the minimum RDT observed in a module reduces by 21.7% and 28.3% at 9 ns and 7.5 ns, respectively, across all modules from all manufacturers.

Opportunity 3. fastHammer can identify a DRAM module’s minimum RDT with a reduction.

Detailed Analysis of fastHammer

To further understand the read disturbance behavior with fastHammer, we extensively characterize representative modules from four major manufacturers and analyze (i) the RDT distribution, (ii) the overlap of read disturbance-vulnerable cells, and (iii) the maintaining of the RDT ranking across rows.

RDT Distribution Shift with fastHammer. Figure 9 shows the distribution of RDTs across all tested rows in four representative modules as subplots. The x-axis shows the RDT of an individual row, and the y-axis shows the probability density across the selected module’s rows. Each line shows a different fastHammer latency ranging from 7.5 ns to 15 ns and RowHammer.

Observation 4. Reducing the PREPRE-toto-ACTACT latency shifts the RDT distribution of a module toward lower RDT values. The RDT distribution shifts toward lower RDT values as the PREPRE-toto-ACTACT latency decreases, in all four tested modules. The shift affects the entire distribution: both the lower tail, which determines a module’s minimum RDT, and the bulk of the distribution move toward lower RDT values. At 7.5 ns, the distributions’ median values decrease by 28.2%, 23.6%, 28.0%, and 16.6% compared to RowHammer for the SK Hynix, Micron, Samsung, and Nanya modules, respectively. We conclude that fastHammer affects a module’s RDT distribution by shifting it toward lower RDT values.

Cell Overlap. To understand whether fastHammer exacerbates RowHammer, we compare the set of cells that experience bitflips as we sweep the PREPRE-toto-ACTACT latency (i.e., for each PREPRE-toto-ACTACT latency, the set of cells that experience bitflips at that latency) with the set of cells that experience RowHammer bitflips (i.e., when the PREPRE-toto-ACTACT latency equals the nominal value). Figure 10 shows how reducing the PREPRE-toto-ACTACT latency (x-axis) changes (a) the fraction of RowHammer-flipped cells that fastHammer also flips (i.e., the cell overlap) and (b) the number of cells that fastHammer flips, normalized to the number of cells that RowHammer flips (y-axis) for four representative DRAM modules.

Per module cell overlap and normalized number of flipped cells with reduced PRE-to-ACT latency

Figure 10. Per module (a) cell overlap (%) with RowHammer and (b) number of flipped cells normalized to RowHammer, with reduced PRE-to-ACT latency.

Observation 5. fastHammer flips nearly all cells that RowHammer flips, at every tested PREPRE-toto-ACTACT latency. fastHammer’s cell overlap with RowHammer exceeds 95% at every tested latency in all tested modules. The overlap increases with the PREPRE-toto-ACTACT latency, from 95.7% at 7.5 ns to 98.9% at 15 ns, and remains nearly constant at ≥9 ns.

Observation 6. fastHammer flips more cells than RowHammer, and increasingly so as the PREPRE-toto-ACTACT latency decreases. fastHammer flips up to 3.9× more cells than RowHammer at 7.5 ns and 1.6× more at 15 ns. Combined with the near-total cell overlap, this shows that fastHammer’s bitflips form a superset of RowHammer’s bitflips: fastHammer induces bitflips in the cells that RowHammer already flips and in additional cells. We hypothesize that this is due to the reduced PREPRE-toto-ACTACT latency exacerbating the read disturbance leakage.

Row Overlap at Top-N RDT Ranking. We analyze whether fastHammer can identify the rows that are most vulnerable to RowHammer, i.e., whether the RowHammer-weakest rows also appear among the fastHammer-weakest rows. We rank all profiled rows in a module twice in ascending order of RDT (weakest row first): once by their RowHammer RDT and once by their fastHammer RDT. We take the NN weakest rows of the RowHammer ranking (i.e., top-NN) and measure the fraction of them that appear within the top X%X\% of the fastHammer ranking. We refer to this fraction as the row overlap.

Figure 11 shows the row overlap across 25 real DRAM modules, with one subplot per PREPRE-toto-ACTACT latency. The x-axis is the fraction XX of the fastHammer ranking we consider, as a percentage of all profiled rows; the y-axis is the resulting row overlap. Each curve shows the mean overlap across all 25 modules for one value of NN (1, 10, 50), and the shaded region spans the 10th to 90th percentile across modules. For NN=1, a module’s overlap is either 0% or 100%, such that the purple curve gives the percentage of modules whose RowHammer-weakest row falls within the given fraction (X%X\%). The dashed vertical line marks the smallest fraction that captures the RowHammer-weakest row in every tested module.

Fraction of the $N$ weakest RowHammer rows that fall within the top $X\%$ of rows ranked by fastHammer RDT for four reduced PRE-to-ACT latencies

Figure 11. Fraction of the NN weakest RowHammer rows (row overlap, y-axis) that fall within the top X%X\% of rows ranked by fastHammer RDT (x-axis), for four reduced PRE-to-ACT latencies.

Observation 7. fastHammer’s ranking places a module’s RowHammer-weakest row among its highest-ranked rows.

In all tested modules, the RowHammer-weakest row falls within the top 5.6%, 3.4%, 1.6%, and 0.7% of the fastHammer ranking at 7.5 ns, 9 ns, 10.5 ns, and 12 ns PRE-to-ACT latency, respectively (dashed vertical lines).4 Because the RowHammer-weakest row determines a module’s minimum RDT, fastHammer’s ranking contains the row that sets this minimum within a small, bounded region at the head of its ranking.

Observation 8. fastHammer’s row overlap degrades for larger sets of RowHammer-weak rows, and increasingly so as the PRE-to-ACT latency decreases.

At 12 ns, the top 20% of the fastHammer ranking contains 99.0% of the top-50 RowHammer-weakest rows, but only 91.5% at 7.5 ns. The same trend holds for the single weakest row: the region that contains it in every module widens by 8×, from 0.7% at 12 ns to 5.6% at 7.5 ns. This is consistent with our cell overlap analysis (Figure 10): more aggressive PRE-to-ACT latency reductions induce bitflips in an increasing number of cells beyond those RowHammer flips, which perturbs the relative ordering of rows by RDT.

Opportunity 4. fastHammer preserves the RDT ranking of rows well enough to enable measuring the minimum RowHammer RDT of a module (i.e., with no RDT reductions due to reduced timing constraints) by reprofiling only a small candidate set with RowHammer-based profiling.

fastHammer in HBM2. We experimentally characterize fastHammer’s RDT reduction, and cell and row overlap characteristics in 10 HBM2 DRAM chips. Figure 12 (a) shows the RDT reduction per DRAM row in all tested HBM2 chips and Figure 12 (b) shows the minimum RDT values per HBM2 chip.

Per row RDT reduction and per module minimum RDT reduction with fastHammer in HBM2 DRAM chips

Figure 12. (a) Per row RDT reduction and (b) per module minimum RDT reduction with fastHammer in HBM2 DRAM chips.

We make three key observations. First, as PRE-to-ACT latency decreases, the RDT reduction increases. We observe average RDT reductions of 12.6%, and 6.5% at 10.0 ns, and 13.3 ns, respectively. Second, the minimum RDT of each tested chip is lower with fastHammer compared to the RowHammer access pattern (e.g., 26.5% average minimum RDT reduction at 10.0 ns). Third, we observe that at 10.0 ns, fastHammer flips nearly all cells (98.8%) that RowHammer flips and places a chip’s RowHammer-weakest row among its top 0.1% rows (not shown). We conclude that fastHammer exhibits similar characteristics in HBM2 chips to DDR4 modules in terms of RDT reduction, and cell and row overlap.

HARE: A Fast RDT Profiling Technique

We propose a new RDT profiling technique, HARE. HARE’s goal is to reduce the time required to profile a DRAM module’s RDT. HARE achieves this goal by hammering each row with reduced DRAM timing constraints (i.e., with fastHammer) during profiling to exacerbate read disturbance intentionally and measure each row’s RDT quickly. HARE then uses these measurements to determine the module’s minimum RDT, i.e., the value that existing RowHammer mitigations use to prevent bitflips. We describe two variants of HARE: (i) HARE-Direct reports a conservative estimate of the minimum RDT, and (ii) HARE-Filter improves the accuracy of the estimate at the cost of additional profiling time.

High-Level Overview. To profile a module’s RDT, HARE profiles each row and picks the lowest RDT measurement. For each row, HARE performs a profiling routine that consists of three steps: (Step 1.) HARE initializes the victim and two aggressor rows with initial data patterns. (Step 2.) HARE performs double-sided fastHammer by repeatedly activating the aggressor rows to complete a predetermined hammer count. (Step 3.) HARE reads the victim row’s data to check for bitflips.

This routine is repeated multiple times to accurately identify the RDT in a binary search approach. HARE starts with an initial hammer count (HinitH_{\mathit{init}}). If the victim row does not have a bitflip, HARE increments the hammer count in Step 2 by an additive step that doubles after every routine (i.e., Hinit+ΔH_{\mathit{init}} + \Delta, Hinit+2ΔH_{\mathit{init}} + 2\Delta, Hinit+4ΔH_{\mathit{init}} + 4\Delta, …), up to the largest hammer count that causes bitflips in the refresh window, and repeats the routine until a bitflip is observed. If HinitH_{\mathit{init}} already induces a bitflip, HARE divides the hammer count by a factor β\beta until the victim row does not have a bitflip. HARE then performs a binary search between the highest hammer count without a bitflip and the lowest hammer count with a bitflip, testing their geometric mean in each step, until the two are within a relative tolerance ϵ\epsilon. HARE reports the lowest hammer count that still results in a bitflip (i.e., the minimum hammer count to induce a bitflip). HARE repeats this search NsearchN_{\mathit{search}} times per row and determines the minimum result as the row’s RDT. Based on empirical data, we set HinitH_{\mathit{init}} to twice the module’s minimum RDT measured in the previous profiling window (10K in the first window, Δ\Delta=10K, β\beta=1.2, ϵ\epsilon=1% (at least 10 hammers), and NsearchN_{\mathit{search}}=5).

HARE-Direct. HARE-Direct reports the minimum RDT observed by performing fastHammer across the module. Doing so, HARE-Direct profiles a DRAM module quickly by (i) reducing the time it takes to perform one hammer and (ii) reducing the number of hammers performed on average. As a result, HARE-Direct’s reported RDT is a conservative estimate of the module’s RowHammer RDT because fastHammer reduces the RDT of a module (e.g., by 21.7%, on average, as shown in §5.2).

HARE-Filter. To improve the RDT estimation of fastHammer, HARE-Filter leverages the row ranking property of fastHammer (§5.3): fastHammer preserves the RDT ranking of rows such that a module’s minimum RowHammer RDT can be identified by exhaustively profiling only its top-ranked rows. In each profiling iteration, HARE-Filter filters the module’s rows: it ranks a batch of rows (Nprofile_batchN_{profile\_batch}) by their fastHammer RDT and selects the top rows with the lowest values (NreprofileN_{reprofile}). HARE-Filter then profiles only the selected NreprofileN_{reprofile} rows with the conventional RowHammer access pattern and reports the minimum RowHammer RDT among them. The selected rows contain the module’s RowHammer-weakest row because fastHammer preserves the RDT ranking of rows. HARE-Filter therefore reports the module’s RowHammer RDT at the cost of profiling a small portion of the rows with slow conventional RowHammer access pattern. Because HARE-Filter re-ranks all rows in every profiling iteration, it dynamically selects the weakest rows as their RDTs change over time.

Periodic RDT Profiling with HARE. We implement both variants of HARE as a periodic profiling scheme to distribute the latency across the execution time and reduce latencies induced by profiling. HARE profiles all rows of the module within a profiling window (tPROFWt_{PROFW}), analogous to the refresh window (tREFWt_{REFW}), and performs one profiling operation every profiling interval (tPROFIt_{PROFI}), analogous to the refresh interval (tREFIt_{REFI}):

tPROFI=tPROFW⌈Ntotal_rowsNper_operation_rows⌉(1)t_{PROFI} = t_{PROFW} \left\lceil\frac{N_{total\_rows}}{N_{per\_operation\_rows}} \right\rceil \tag*{(1)}

where Ntotal_rowsN_{total\_rows} is the total number of rows to profile and Nper_operation_rowsN_{per\_operation\_rows} is the number of rows HARE profiles in one operation.

Choosing a Configuration. Based on our characterization of 180 real DDR4 DRAM chips, we select fastHammer with 9 ns PRE-to-ACT latency to quickly profile the RDT of DRAM rows because this configuration provides (i) a high reduction in RDT and the total time required for one hammer cycle compared to the RowHammer access pattern, and (ii) a high cell and row overlap percentages. We implement a version of HARE that profiles one row in each bank during one profiling operation (i.e., every profiling interval). We sweep a set of the profiling window and interval values in §7 and §8.1.

Error Probability Analysis

HARE improves the read disturbance tolerance of a memory system by enabling rapid identification of the minimum RDT of a DRAM module. For example, when the minimum RDT of a DRAM module reduces over time due to VRD [71], the system can (1) use HARE to rapidly identify the new minimum RDT value, (2) configure the read disturbance mitigation with the newly identified minimum RDT. Configuring the mitigation with the newly identified value would reduce the probability of a read disturbance bitflip, improving overall system reliability. To quantitatively analyze the system reliability benefits of using HARE, we develop a probabilistic error model for VRD-induced read disturbance bitflips based on our empirical read disturbance characterization results.

The Baseline Error Model

RDT of a DRAM Row. We draw each row’s RDT from an empirical per-row RDT distribution [71]. Because a row’s RDT changes over time, we additionally draw a per-row drift ratio from an empirical VRD distribution and define the row’s effective RDT (RDTeffRDT_{eff}) as its measured RDT divided by that ratio. We assume a row’s effective RDT is always lower than the measured RDT to evaluate a conservative scenario where VRD-induced changes can result in a bitflip when a mitigation is configured with the measured RDT. As in §5, we express a row’s RDT in activations of its aggressor rows: a victim row exhibits a bitflip once its aggressor rows have been activated k=⌈RDTeff⌉k = \lceil RDT_{eff} \rceil times.

Mitigation. We model a generic deterministic counter-based RowHammer mitigation that maintains an activation counter per row and preventively refreshes a row’s victim rows once its counter reaches a configured RDT (RDTconfRDT_{conf}). We set the configured RDT to the minimum RDT measured across all rows in a bank. The mitigation can employ a guardband (gg) to the measured minimum RDT across the bank as a countermeasure against VRD-induced RDT changes [71]:

RDTconf=min⁡(RDTmeasured)×(1−g)(2)RDT_{conf} = \min(RDT_{measured}) \times(1-g) \tag*{(2)}

Vulnerable DRAM Rows. A mitigation is configured using measured RDTs, whereas rows fail at their effective RDTs. A victim row is vulnerable if k<RDTconfk < RDT_{conf} (i.e., effective RDT is lower than the mitigation’s configured RDT).

A Read Disturbance Bitflip. The activation counter of a DRAM row resets once it reaches RDTconfRDT_{conf} (and the row’s victims are refreshed), thus a vulnerable victim row exhibits bitflips after kk, k+Pk+P, k+2Pk+2P, … aggressor activations, where P=⌈RDTconf⌉P = \lceil RDT_{conf} \rceil. We determine the number of bitflips exhibited by a vulnerable victim row by sampling an empirically-observed distribution. We distribute the exhibited bitflips uniformly at random across the DRAM row because we do not observe a spatial correlation between these bitflips.

An Uncorrectable Error. We model a DRAM chip that employs on-die single-error-correcting ECC over 128-bit data words [86, 111, 112] [113]. If two or more bitflips exhibited by a vulnerable row fall in the same 128-bit word, we consider that an uncorrectable error.

Online Profiler

We model an RDT profiler as a periodic profiler that measures rows at the profiling interval (tPROFIt_{PROFI}, §6) and lowers the mitigation’s configured RDT as it discovers rows with lower RDTs, recomputing RDTconfRDT_{conf} from the new minimum as in (2). The first version of HARE, HARE-Direct, updates the configured RDT immediately: as soon as it profiles a row whose effective RDT is lower than the current configured RDT, it lowers the configured RDT to that row’s value. The second version of HARE, HARE-Filter, does not update the configured RDT immediately, because ranking rows requires a batch: it profiles Nprofile_batchN_{profile\_batch} rows, re-profiles the top-ranked NreprofileN_{reprofile} rows, and only then lowers the configured RDT to the lowest value it measures. Therefore, the system retains a configured RDT value by the time it takes to complete the profiling of one batch at the least. In both variants the configured RDT only decreases, so each reduction permanently protects every row whose bitflip threshold is at or above the new value. For HARE-Filter, we select Nprofile_batchN_{profile\_batch} and NreprofileN_{reprofile} based on our empirical analysis (§5).

Monte Carlo Evaluation Methodology.

In our error model, an interplay of empirically measured, non-parametric distributions (per-row RDT, drift, and bitflips per event) and stateful elements (e.g., the read disturbance mitigation and online RDT profiler) determines the DRAM bank’s uncorrectable error probability. Thus, we estimate the time to uncorrectable error via Monte Carlo simulation instead of analytical modeling.

Simulated Monte Carlo Trial. We simulate one DRAM bank of 65,536 rows and express time in terms of DRAM row activations (e.g., 45 ns per activation [72]). Each trial draws a fresh bank from the measured input distributions and executes a read-disturbance-intensive memory access pattern (described next) until the first UE or until 101410^{14} row activations, whichever comes first. We assume that a DRAM row’s RDT is drifted at the start of each Monte Carlo trial. This is a conservative assumption because modeled drift only lowers a row’s RDT. We assume a bitflip in memory persists and there is no scrubbing.

Memory Access Pattern. We consider a pessimistic access pattern for the modeled DRAM system whereby the access pattern repeatedly 1) uniformly at random selects a victim row and 2) activates the aggressor rows of the victim row RDTconfRDT_{conf} times.5 We implement the error model in §7.1 as a Monte Carlo simulator and report population statistics over independently drawn DRAM banks. Table 2 lists our simulation parameters.

ParameterValue
DRAM bank size (Ntotal_rowsN_{total\_rows})65’536 rows
Time per row activationtRC (e.g., 45 ns [72])
On-die ECC(136,128) SEC, 128-bit data words
Guardband (gg)0%–50% RDT reduction [71]
Profiling interval (tPROFIt_{PROFI})6.75–128 ms
Rows per batch (Nprofile_batchN_{profile\_batch})3’000 (§5), 1’500, 750 rows
Re-profiled rows (NreprofileN_{reprofile})100 (§5), 50, 25 rows
Trials per configuration1’000’000 banks
Trial time limit101410^{14} activations

Table 2. Monte Carlo simulation parameters.

Metrics. We report the probability of an uncorrectable error within a time budget, P(UE≤t)P(\mathrm{UE} \le t), and mean time to an uncorrectable error (MTTUE). A bank in which no row is vulnerable can never experience a UE, therefore we also report fraction of banks that can experience UEs.

Estimators. A trial that reaches the trial time limit has no observed UE. We therefore compute MTTUE as the restricted mean survival time until the trial time limit, which is exact under common trial time limit across trials and equals the plain mean when all trials end before the time limit.

Uncorrectable Error Probability Analysis

No Online Profiling. Figure 13 shows the P(UE) percentage of a system that employs a RowHammer mitigation with no online profiler. The x-axis shows time and the hue depicts different guardbands for the RowHammer mitigation.

P(Uncorrectable Error) percentage of a system with no online profiler across time over 1M trials

Figure 13. P(Uncorrectable Error) percentage of a system with no online profiler across time over 1M trials.

We make two key observations. First, without a guardband (g=0%g = 0\%, depicted as the black line), a system with no online profiler experiences an P(UE) of >99.9%>99.9\%, with a mean time to uncorrectable error (MTTUE) (shown with the dashed gray line) of only 8.5 minutes. Second, a larger guardband reduces the P(UE), but only a very large guardband reduces it close to zero. P(UE) reaches at most 88.0%, 66.8%, 24.0%, 4.0%, and 0.5% for g=15%g=15\%, 20%, 30%, 40%, and 50%, respectively. This is because a row experiences bitflips only if its RDT decreases by more than the guardband due to VRD, and these bitflips become uncorrectable only if ECC cannot correct them. A larger guardband means that fewer rows have an RDT decrease larger than the guardband, thus, fewer trials end in an UE.

We conclude that a RowHammer mitigation that uses a static RDT based on one-time profiling has a very high probability of encountering an UE. Reaching a negligible P(UE) requires g=50%g=50\%, i.e., halving the RDT that the mitigation is configured with, at a significant performance cost [71].

HARE-Direct. Figure 14 shows the P(UE) of a system with HARE-Direct (solid) and with no online profiler (dashed) across time (cut at 5.2 days where curves are relatively horizontal). Figure 14(a) depicts the P(UE) in systems employing RowHammer mitigations with guardband values of 0%, 20%, and 50%. The online profiler has a profiling interval of 6.75 ms, the interval needed to profile all rows within the MTTUE of the system with no guardband.

P(Uncorrectable Error) percentage of a system with HARE-Direct and with no online profiler across different guardband values and profiling intervals

Figure 14. P(Uncorrectable Error) percentage of a system with HARE-Direct and with no online profiler across (a) different guardband values at a profiling interval of 6.75 ms, and (b) different profiling intervals, over 1M trials.

We make two key observations. First, online profiling reduces the P(UE) by more than an order of magnitude at every guardband value. At g=0%g=0\%, HARE-Direct experiences a UE in 6.2% of the trials, whereas the system with no online profiler experiences a UE in 99.9% of the trials. Second, online profiling reaches a low P(UE) with a small guardband. HARE-Direct experiences a UE in 0.67% of the trials at g=20%g=20\%. Further increasing the guardband to 50% reduces HARE-Direct’s P(UE) to 0.002%. We conclude that HARE-Direct significantly reduces the P(UE) compared to a system with no online profiler, increasing with guardband.

Figure 14(b) shows the P(UE) of a system with HARE-Direct profiling DRAM rows at various profiling intervals (with g=20%g=20\%, similar to the average RDT reduction observed in our empirical results § 5) and a system without an online profiler (and without a guardband). We make two key observations. First, with increased profiling interval of 128 ms, HARE-Direct still provides smaller P(UE) compared to a system with no online profiler. Second, the P(UE) increases as the profiling interval increases. This is because, as profiling interval increases, HARE-Direct discovers the minimum effective RDT later in time, leaving more time for the access pattern to trigger UEs. We conclude that, when configured with a larger profiling interval, HARE-Direct still provides a smaller P(UE) over a system with no online RDT profiler.

HARE-Filter. We evaluate the reliability of a system employing HARE-Filter. HARE-Filter updates the RowHammer mitigation’s configured RDT only after (i) profiling a batch of rows (Nprofile_batchN_{profile\_batch}) with fastHammer and (ii) reprofiling a small percentage of the batch of rows (NreprofileN_{reprofile}) with RowHammer. Figure 15 shows the P(UE) of a system with HARE-Filter employing various batch sizes, and with no online profiler. We sweep Nprofile_batchN_{profile\_batch} from 3K rows to 750 rows. We make three key observations. First, with the tested batch sizes, HARE-Filter

P(Uncorrectable Error) percentage of a system with HARE-Filter and no online profiler across profiling batch size

Figure 15. P(Uncorrectable Error) percentage of a system with HARE-Filter and no online profiler across profiling batch size, over 1M trials.

provides lower (by 12.1x with 750-row batch) P(UE) compared to the system with no online profiler. Second, HARE-Filter’s P(UE) monotonically decreases as Nprofile_batchN_{profile\_batch} decreases. This is because a smaller batch shortens the window before the first configuration update and most of the failing banks fail before any update has been applied. Third, HARE-Filter can more accurately measure the RowHammer RDT of a row compared to HARE-Direct (§6) but it also results in an P(UE) larger than that of HARE-Direct (not shown in the figure) due to the slower RDT updates. For example, at Nprofile_batch=750N_{profile\_batch}=750, HARE-Filter has a 33.3% higher P(UE) compared to HARE-Direct. We conclude that HARE-Filter with the tested batch sizes improves the P(UE) over a system with no profiler, while trading off RDT measurement accuracy with higher P(UE).

Performance and Energy Consumption Analysis of HARE

System Performance Evaluation

We evaluate HARE’s impact on performance using Ramulator 2.0 [114, 115], a cycle-level DRAM simulator. Table 3 shows the configuration of our simulated system. We evaluate HARE using 10 quad-core workload mixes. To synthesize each workload mix, we randomly select workloads from the SPEC CPU2006 [116], SPEC CPU2017 [117], TPC [118], MediaBench [119], and YCSB [120] suites, considering only memory-intensive workloads with more than 10 row-buffer misses per kilo-instruction: 519.lbm, 459.GemsFDTD, 450.soplex, h264.decode, 520.omnetpp, 433.milc, 434.zeusmp, bfs_dblp, 429.mcf, 549.fotonik3d, 470.lbm, bfs_ny, bfs_cm2003, 437.leslie3d, gups. We design the conventional RowHammer profiler, based on the RDT measurement methodology widely used by prior characterization works [2, 4, 71, 77] (§4). This is because, to our knowledge, no prior work proposes an online profiler for RowHammer. The RowHammer profiler employs the same periodic profiling scheme as HARE-Direct and replaces the fastHammer with RowHammer access pattern. Since RowHammer access pattern directly measures the RowHammer RDT, RowHammer profiler does not require a second reprofiling phase and can directly update the RDT configuration of the underlying RowHammer mitigation.

Performance Impact of HARE. We evaluate a system that employs a state-of-the-art RowHammer mitigation mechanism (i.e., PRAC [73, 81], Graphene [23], and PARA [1]) with (i) HARE-Direct, (ii) HARE-Filter and (iii) a RowHammer profiler. We configure all RowHammer mitigations with their corresponding preventive action thresholds, calculated based on the RDT configuration evaluated. We evaluate HARE-Direct with a conservative RDT value with the average reduction of 21.7%, as observed in our empirical results (§5). Fig. 16 presents the performance of a system using HARE-Direct and HARE-Filter across different values of RDT, underlying RowHammer mitigations, and profiling intervals. The plots of Figure 16 are organized into groups, with each group corresponding to a different RDT value (from left to right: 4K, 1K, and 256). Each plot within a group corresponds to a different underlying RowHammer mitigation mechanism (from left to right: PRAC [73, 81], Graphene [23], and PARA [1]). The x-axis shows the profiling interval (tPROFIt_{PROFI}, i.e., time between consecutive profiling operations), while the y-axis shows weighted speedup normalized to the same system using a RowHammer profiler at the same tPROFIt_{PROFI}. For RDT=4K, we sweep tPROFIt_{PROFI} values of 6.75 ms (based on our analysis in §7), 16, 32, and 64. For lower RDT values, we scale the tPROFIt_{PROFI} accordingly to simulate similar system reliability benefits for the online profilers.

Performance of a system using HARE-Direct and HARE-Filter normalized to the RowHammer profiler, across profiling intervals for three RDT values and three underlying RowHammer mitigations

Figure 16. Performance of a system using HARE-Direct and HARE-Filter normalized to the RowHammer profiler, across profiling intervals for three RDT values (groups of subplots) and three underlying RowHammer mitigations (subplots within each group).

We make four key observations. First, HARE-Direct and HARE-Filter outperform the RowHammer profiler, on average across all applications, for all RowHammer mitigation, tPROFIt_{PROFI}, and RDT configurations. HARE-Direct’s and HARE-Filter’s average performance improvements across all applications vary from 54.4% to 0.9% and 47.5% to 1.5%, respectively. This is because both HARE-Direct and HARE-Filter profile rows faster than the RowHammer profiler, reducing the time spent on profiling and increasing the time available for demand memory requests. Second, the speedup of both variants increases with decreasing tPROFIt_{PROFI}, as profiling consumes a larger fraction of total execution time, such that reducing profiling latency yields proportionally larger performance gains. For example, with PRAC at RDT 256, HARE-Direct’s (HARE-Filter’s) speedup increases from 12.5% (8.8%) at tPROFIt_{PROFI}=4 ms to 54.4% (39.7%) at tPROFIt_{PROFI}=0.5 ms. Third, HARE-Direct outperforms HARE-Filter in most (26 out of 36) configurations, as its faster profiling operations outweigh the extra preventive actions caused by its conservative RDT estimate. Fourth, HARE-Filter outperforms HARE-Direct in configurations where RowHammer mitigation overhead impacts system performance more than the profiling overhead. For example, in a system that employs PARA, HARE-Filter outperforms HARE-Direct at every profiling interval, at the low RDT of 256. This is due to PARA’s significant performance overheads for low RowHammer thresholds, which amplify the impact of HARE-Direct’s conservative RDT estimate and resulting unnecessary preventive actions. The remaining configurations are those with long tPROFIt_{PROFI} values at RDT≤1K, where profiling operations are infrequent and thus, RowHammer mitigation overhead outweighs profiling cost.

We conclude that HARE improve system performance over the conventional RowHammer profiler, and that its benefits are largest when profiling is frequent, i.e., at the short profiling intervals. HARE-Direct performs better when profiling overhead dominates, whereas HARE-Filter is preferable when the mitigation’s overhead dominates, as its more accurate RDT estimate avoids unnecessary preventive actions.

Execution Time Breakdown Analysis

Figure 17 shows the execution time breakdown of a system that uses HARE-Direct (fH) or the RowHammer profiler (RH), normalized to the system with the RowHammer profiler at RDT=4K and RDT=256 (subplots), across profiling intervals (x-axis). Each bar breaks execution time into application execution, RowHammer mitigation overhead, and profiling overhead. The HARE-Direct bar also shows 1) the additional mitigation overhead caused by configuring the mitigation with fastHammer’s lower RDT, and 2) the execution time HARE-Direct saves compared to the RowHammer profiler (hatched).

Execution time breakdown of a system using HARE-Direct, normalized to RowHammer profiler at RDT=4K and 256 across profiling intervals

Figure 17. Execution time breakdown of a system using HARE-Direct, normalized to RowHammer profiler at RDT=4K and 256 across profiling intervals.

We make two key observations. First, HARE-Direct’s speedup comes from reduced profiling time, which is largest at short profiling intervals (e.g., HARE-Direct reduces execution time by 35% with PRAC at RDT=256 and tPROFIt_{PROFI}=0.5 ms). Second, HARE-Direct’s additional mitigation overhead is negligible (<1%) for PRAC and Graphene, and noticeable only for PARA. We conclude that HARE-Direct’s profiling time savings outweigh its additional mitigation overhead in every evaluated configuration.

Sensitivity to Bank Density and Profiled Rows

DRAM bank density increases with chip density (e.g., from 64K to 256K rows per bank in 16Gb to 64Gb DDR5 chips [73, 123]), increasing the number of rows a profiler must profile. We evaluate HARE across bank densities of 32K to 256K rows. To analyze the sensitivity to the number of profiled rows, we evaluate a hypothetical scenario where profilers profile 50% or 25% of the rows in each bank. This scenario is optimistic as results (§3, Figure 4) show that a module’s minimum RDT does not stabilize within 9K profiled rows, so profiling a subset of rows may not guarantee finding the minimum RDT. Figure 18 shows the weighted speedup of HARE-Direct (solid) and HARE-Filter (dashed) normalized to the RowHammer profiler that profiles the same rows, at RDT=4K.

Figure 18. Performance of HARE normalized to the RowHammer profiler across bank densities, when profiling 100%, 50%, and 25% of the rows per bank.

We make two key observations. First, HARE-Direct outperforms the RowHammer profiler in every configuration, and its benefit grows with bank density (e.g., with PRAC, from 47.7% at 64K rows to 89.5% at 128K rows). Second, even in the optimistic scenario, profiling a subset of rows does not eliminate the profiling cost in dense banks, e.g., even when profiling only 25% of the rows, HARE-Direct improves performance by 48% in a 256K-row bank. We conclude that HARE significantly reduces RDT profiling cost, and that this benefit grows as banks become denser.

Hardware Cost Analysis

HARE can be implemented in various ways, and we implement a version in the memory controller. HARE-Direct stores the current minimum RDT in a single register, thus, incurring negligible overheads. HARE-Filter maintains a per-bank lowest-RDT table of the 128 rows with the lowest fastHammer-measured RDT, each entry holding a 20-bit RDT (up to ≈1M activations) and a 17-bit row ID. For the system of Table 3, the storage overheads of HARE-Filter are 37.2 KiB. We implement the lowest-RDT table in Verilog HDL and use open-source tools [124, 125] to synthesize it and evaluate its latency, and area overheads. HARE-Filter incurs 0.17 mm2^{2} area overhead. Updating the lowest-RDT table takes 260 ns (in the worst case). Table updates are performed off-the-critical-path of the memory controller, introducing no additional latency to memory requests. Since the shortest profiling interval we evaluate is tPROFI=0.5t_{PROFI}=0.5 ms, the table can always insert a new measurement before the next profiling operation.

Processor4.2 GHz, 4-core, 4-wide issue, 128-entry instr. window
Last-Level Cache64-byte cache line, 8-way set-associative, 8 MB
Memory Controller64-entry read/write request queues; Scheduling policy: FR-FCFS+Cap of 4 [121]; Address mapping: MOP [122]
Main MemoryDDR5 DRAM [115], 1 channel, 2 ranks, 8 bank groups, 4 banks/bank group, 64K rows/bank

Table 3. Simulated System Configuration

Energy Analysis

Methodology. We measure the power consumption of a representative real DDR4 module at its supply rails while the module continuously executes each access pattern. We monitor both the core (VDDV_{DD}) and the wordline (VPPV_{PP}) voltage/current rails, and compute the module’s power. We compute the energy consumption of each profiling operation as the measured mean power multiplied by the time required to complete a hammer cycle and the hammer count to induce the first bitflip.

Figure 19 shows the energy consumed to hammer a row with the conventional RowHammer access pattern and with fastHammer. The x-axis shows the number of hammers performed with the RowHammer access pattern. The shaded region shows the energy that fastHammer saves.

Figure 19. Energy consumption of hammering a row with the RowHammer access pattern and with fastHammer.

We observe that fastHammer consumes significantly less energy than the RowHammer pattern across all tested RDT values. At 32K RowHammer hammers, fastHammer consumes 273 μ\muJ whereas RowHammer consumes 419 μ\muJ, i.e., fastHammer saves 146 μ\muJ (34.8%). We conclude that HARE reduces the energy cost of RDT profiling in addition to its latency.

Latency Analysis

We measure latency of profiling for a number of hammers with fastHammer and RowHammer based on the three-step routine described in §6. We follow the methodology described in [71] to tightly schedule DRAM commands based on the access patterns. For the hammer count of 4K, we observe that fastHammer reduces the latency of profiling by 31.2% compared to RowHammer access pattern due to (i) reduced timing constraints between PRE and ACT commands and (ii) reduction in hammer count (as shown in §5).

Discussion

Supporting HARE with PRAC. PRAC requires a safe RDT configuration which HARE can provide. To avoid triggering PRAC’s preventive actions during profiling, the memory controller can use specialized precharge command that does not update the PRAC counters, similar to prior work [66], or the system can employ HARE in DRAM.

In-DRAM HARE. HARE can run entirely in DRAM, if it supports autonomous maintenance operations (e.g., Self-Managing DRAM [126]). Based on SMD, the DRAM chip can lock the region under profiling, reject memory controller requests (e.g., via ACT-NACK), and perform fastHammer internally. This way, DRAM chip can perform fastHammer with fine-grained control over the PRE-to-ACT latency and HARE reads raw cell data, therefore avoiding on-die ECC masking bitflips. DRAM can also prevent RowHammer mitigations to perform refreshes during profiling. After profiling, DRAM can use the profiled RDT directly in in-DRAM RowHammer mitigations or can report it to the memory controller-based mitigations (e.g., via a mode register).

System Support for Reduced tRPt_{RP}. HARE requires the memory controller to apply a reduced tRPt_{RP} only to profiling commands. Modern memory controllers already expose tRPt_{RP} as a programmable parameter (e.g., Intel’s TC_PRE register [127] and AMD’s configurable tRPt_{RP} [128]). Industry also proposes changing DRAM timing parameters in-system to test installed DRAM beyond its rated PRE-to-ACT timing [129] and to switch timing profiles that include tRPt_{RP} at runtime without a reboot [130]. Together, these capabilities provide the support for reduced tRPt_{RP} in profiling.

Related Work

To our knowledge, this is the first work to 1) show, through a rigorous characterization of 180 real DDR4 DRAM chips from four major manufacturers, that intentionally exacerbating read disturbance with reduced DRAM timing constraints preserves the read disturbance-weak cells and rows of the conventional RowHammer access pattern, and 2) exploit this property to design an online RDT profiling technique that identifies a module’s minimum RDT at low cost.

DRAM Read Disturbance Characterization. Many works [1, 2, 3, 4, 65, 71, 77, 102, 103, 105, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140] experimentally characterizes the real DRAM chip’s read disturbance vulnerability. Two of these works reduce timing parameters to study read disturbance [103, 105], yet neither examines whether resulting bitflips preserve the read disturbance characteristics of the conventional RowHammer access pattern. We are the first to characterize the cell-level and row-level overlap between reduced-timing and RowHammer bitflips across 25 real DRAM modules, and to show that this overlap enables fast RDT profiling.

Retention Failure Profiling. Prior works [92, 94, 101, 141, 142] propose methods to efficiently profile DRAM retention failures that are subject to the variable retention time. Similarly, read disturbance bitflips are subject to variable read disturbance [71], yet no prior work provides an online RDT profiling technique. HARE is the first online RDT profiling technique.

Conclusion

We present HARE, an online RDT profiling technique that quickly profiles a DRAM module by intentionally exacerbating read disturbance. To do so, HARE reduces the DRAM timing constraints during profiling, which both shortens the execution time of profiling and induces the first bitflip at a smaller activation count than conventional RowHammer. Our characterization of 180 real DDR4 DRAM chips shows that HARE measures a module’s minimum RDT faster than the RowHammer profiler and preserves RDT ranking well enough to enable profiling a smaller candidate set to reprofile and refine its conservative minimum RDT measurement. Our end-to-end evaluations show that HARE improves system performance over the RowHammer profiler. We hope that HARE enables and inspires future work on fast and accurate online read disturbance profiling.

Acknowledgments

We thank the anonymous reviewers of HPCA 2026 for their feedback. We thank the SAFARI Research Group members for their constructive feedback and for providing a stimulating intellectual, scientific, scholarly environment. We acknowledge the generous gift funding provided by our industrial partners (especially Google, Huawei, Intel, Microsoft), which has been instrumental in enabling the research we have been conducting on read disturbance in DRAM in particular and memory systems in general. This work was in part supported by a Google Security and Privacy Research Award and the Microsoft Swiss Joint Research Center.

References

  1. [1]Y. Kim, R. Daly, J. Kim, C. Fallin, J. H. Lee, D. Lee, C. Wilkerson, K. Lai, and O. Mutlu. Flipping Bits in Memory Without Accessing Them: An Experimental Study of DRAM Disturbance Errors. In ISCA, 2014.DOI
  2. [2]Haocong Luo, Ataberk Olgun, Abdullah Giray Yağlıkçı, Yahya Can Tuğrul, Steve Rhyner, Meryem Banu Cavlak, Joël Lindegger, Mohammad Sadrosadati, and Onur Mutlu. RowPress: Amplifying Read Disturbance in Modern DRAM Chips. In ISCA, 2023.
  3. [3]Jeremie S. Kim, Minesh Patel, Abdullah Giray Yağlıkçı, Hasan Hassan, Roknoddin Azizi, Lois Orosa, and Onur Mutlu. Revisiting RowHammer: An Experimental Analysis of Modern Devices and Mitigation Techniques. In ISCA, 2020.
  4. [4]Lois Orosa, A Giray Yağlıkçı, Haocong Luo, Ataberk Olgun, Jisung Park, Hasan Hassan, Minesh Patel, Jeremie S. Kim, and Onur Mutlu. A Deeper Look into RowHammer’s Sensitivities: Experimental Analysis of Real DRAM Chips and Implications on Future Attacks and Defenses. In MICRO, 2021.
  5. [5]Apple Inc. About the Security Content of Mac EFI Security Update 2015-001. https://support.apple.com/en-us/HT204934. June 2015.
  6. [6]Hewlett-Packard Enterprise. HP Moonshot Component Pack Version 2015.05.0, 2015.
  7. [7]Lenovo Group Ltd. Row Hammer Privilege Escalation, 2015.
  8. [8]Zvika Greenfield and Tomer Levy. Throttling Support for Row-Hammer Counters, 2016.
  9. [9]Dae-Hyun Kim, Prashant J Nair, and Moudiuddin K Qureshi. Architectural Support for Mitigating Row Hammering in DRAM Memories. CAL, 2014.DOI
  10. [10]K.S. Bains and J.B. Halbert. Distributed Row Hammer Tracking. US Patent App. 13/631,781, April 3 2014.
  11. [11]K.S. Bains et al. Method, Apparatus and System for Providing a Memory Refresh. US Patent App. 13/625,741, March 27 2014.
  12. [12]K.S. Bains et al. Row Hammer Refresh Command. US Patent App. 13/539,415, January 2 2014.
  13. [13]K. Bains et al. Row Hammer Refresh Command. US Patent App. 14/068,677, February 27 2014.
  14. [14]Zelalem Birhanu Aweket, Salessawki Ferede Yitbarek, Rui Qiao, Reetupama Das, Matthew Hicks, Yossi Oren, and Todd Austin. ANVIL: Software-Based Protection Against Next-Generation Rowhammer Attacks. In ASPLOS, 2016.
  15. [15]Kuljit S Bains, John Halbert, Christopher Mozak, Theodore Schoenborn, and Zvika Greenfield. Row Hammer Refresh Command, 2015. U.S. Patent 9,117,544.
  16. [16]Kuljit S Bains and John B Halbert. Row Hammer Monitoring Based on Stored Row Hammer Threshold Value, 2016. U.S. Patent 9,384,821.
  17. [17]Kuljit S Bains and John B Halbert. Distributed Row Hammer Tracking, 2016. U.S. Patent 9,299,400.
  18. [18]Mungyu Son, Hyunsun Park, Junwhan Ahn, and Sungjoo Yoo. Making DRAM Stronger Against Row Hammering. In DAC, 2017.DOI
  19. [19]S. M. Seyedzadeh, A. K. Jones, and R. Melhem. Mitigating Wordline Crosstalk Using Adaptive Trees of Counters. In ISCA, 2018.arxiv.org/abs/1806.02498
  20. [20]Gorka Irazoqui, Thomas Eisenbarth, and Berk Sunar. MASCAT: Stopping Microarchitectural Attacks Before Execution. IACR Cryptology, 2016.
  21. [21]Jung Min You and Joon-Sung Yang. MRLoc: Mitigating Row-Hammering Based on Memory Locality. In DAC, 2019.
  22. [22]Eojin Lee, Ingab Kang, Sukhan Lee, G Edward Suh, and Jung Ho Ahn. TWiCe: Preventing Row-Hammering by Exploiting Time Window Counters. In ISCA, 2019.
  23. [23]Yeonhong Park, Woosuk Kwon, Eojin Lee, Tae Jun Ham, Jung Ho Ahn, and Jae W Lee. GRe:thane: Strong yet Lightweight Row Hammer Protection. In MICRO, 2020.DOI
  24. [24]A. Giray Yağlıkçı, Jeremie S. Kim, Fabrice Devaux, and Onur Mutlu. Security Analysis of the Silver Bullet Technique for RowHammer Prevention. arXiv:2106.07084 [cs.CR], 2021.
  25. [25]A. Giray Yağlıkçı, Minesh Patel, Jeremie S. Kim, Roknoddin Azizibarzoki, Ataberk Olgun, Lois Orosa, Hasan Hassan, Jisung Park, Konstantinos Kanellopoulos, Taha Shahroodi, Saugata Ghose, and Onur Mutlu. BlockHammer: Preventing RowHammer at Low Cost by Blacklisting Rapidly-Accessed DRAM Rows. In HPCA, 2021.
  26. [26]Pietro Frigo, Emanuele Vannacci, Hasan Hassan, Viktor van der Veen, Onur Mutlu, Cristiano Giuffrida, Herbert Bos, and Kaveh Razavi. TRRespass: Exploiting the Many Sides of Target Row Refresh. In S&P, 2020.arxiv.org/abs/2004.01807
  27. [27]Ingab Kang, Eojin Lee, and Jung Ho Ahn. CAT-TWO: Counter-Based Adaptive Tree, Time Window Optimized for DRAM Row-Hammer Prevention. IEEE Access, 2020.
  28. [28]Hasan Hassan, Yahya Can Tuğrul, Jeremie S. Kim, Victor van der Veen, Kaveh Razavi, and Onur Mutlu. Uncovering In-DRAM RowHammer Protection Mechanisms: A New Methodology, Custom RowHammer Patterns, and Implications. In MICRO, 2021.arxiv.org/abs/2110.10603
  29. [29]Mouddin Qureshi, Aditya Rohan, Gururaj Saileshwar, and Prashant J Nair. Hydra: Enabling Low-Overhead Mitigation of Row-Hammer at Ultra-Low Thresholds via Hybrid Tracking. In ISCA, 2022. Asia CCS, 2019.
  30. [30]Gururaj Saileshwar, Bolin Wang, Moinuddin Qureshi, and Prashant J Nair. Randomized Row-Swap: Mitigating Row Hammer by Breaking Spatial Correlation Between Aggressor and Victim Rows. In ASPLOS, 2022.DOI
  31. [31]Brendan Brasser, Lucas Davi, David Gens, Christopher Liebchen, and Ahmad-Reza Sadeghi. Can’t Touch This: Software-Only Mitigation Against Rowhammer Attacks Targeting Kernel Memory. In USENIX Security, 2017.
  32. [32]Radhesh Krishnan Konoth, Marco Oliverio, Andrei Tatar, Dennis Andriesse, Herbert Bos, Cristiano Giuffrida, and Kaveh Razavi. ZebRAM: Comprehensive and Compatible Software Protection Against Rowhammer Attacks. In OSDI, 2018.DOI
  33. [33]Victor van der Veen, Martina Lindorfer, Yanick Fratantonio, Harikrishnan Padmanabha Pillai, Giovanni Vigna, Christopher Kruegel, Herbert Bos, and Kaveh Razavi. GuardION: Practical Mitigation of DMA-Based Rowhammer Attacks on ARM. In DIMVA, 2018.DOI
  34. [34]Saru Vig, Sarani Bhattacharya, Debdeep Mukhopadhyay, and Siew-Kei Lam. Rapid Detection of Rowhammer Attacks Using Dynamic Skewed Hash Tree. In HAsP, 2018.DOI
  35. [35]Michael Jaemin Kim, Jaehyun Park, Yeonhong Park, Wanju Doh, Namhoon Kim, Tae Jun Ham, Jae W Lee, and Jung Ho Ahn. MITHRIL: Cooperative Row Hammer Protection on Commodity DRAM Leveraging Managed Refresh. In HPCA, 2022.arxiv.org/abs/2108.06703
  36. [36]Gyu-Hyeon Lee, Seongmin Na, Ilkwon Byun, Dongmoon Min, and Jangwoo Kim. CroGuard: A Near Refresh-Free Robust DRAM Design for Cryogenic Computing. In ISCA, 2021.DOI
  37. [37]Michele Marazzi, Patrick Jattke, Flavien Solt, and Kaveh Razavi. ProTRR: Principled yet Optimal In-DRAM Target Row Refresh. In S&P, 2022.
  38. [38]Zhi Zhang, Yueqiang Cheng, Minghua Wang, Wei He, Wenhao Wang, Surya Nepal, Yansong Gao, Kang Li, Zhe Wang, and Chenggang Wu. SoftTRR: Protect Page Tables against Rowhammer Attacks using Software-only Target Row Refresh. In USENIX ATC, 2022.arxiv.org/abs/2102.10269
  39. [39]Brersh Kumar Joardar, Tyler K Bletsch, and Krishnendu Chakrabarty. Learning to Mitigate RowHammer Attacks. In DATE, 2022.
  40. [40]Jonas Juffinger, Lukas Lamster, Andreass Kogler, Maria Eichlseder, Moritz Lipp, and Daniel Gruss. CSI: Rowhammer-Cryptographic Security and Integrity against Rowhammer. In SP, 2023.DOI
  41. [41]A. Giray Yağlıkçı, Ataberk Olgun, Minesh Patel, Haocong Luo, Hasan Hassan, Lois Orosa, Oğuz Ergin, and Onur Mutlu. HiRA: Hidden Row Activation for Reducing Refresh Latency of Off-the-Shelf DRAM Chips. In MICRO, 2022.arxiv.org/abs/2209.10198
  42. [42]Anish Saxena, Gururaj Saileshwar, Prashant J. Nair, and Moinuddin Qureshi. AQUA: Scalable Rowhammer Mitigation by Quarantining Aggressor Rows at Runtime. In MICRO, 2022.DOI
  43. [43]Shuhei Enomoto, Hiroki Kuzuno, and Hiroshi Yamada. Efficient Protection Mechanism for CPU Cache Flush Instruction Based Attacks. TOIS, 2022.DOI
  44. [44]Egemen Manzhosov, Adam Hastings, Meghna Pancholi, Ryan Piersma, Mohamed Tarek Ibn Ziad, and Simha Sethumadhavan. Revisiting Residue Codes for Modern Memories. In MICRO, 2022.
  45. [45]Samira Mirbagher Ajorpaz, Daniel Moghimi, Jeffrey Neal Collins, Gilles Pokam, Nael Abu-Ghazaleh, and Dean Tullsen. EVAX: Towards a Practical, Pro-active & Adaptive Architecture for High Performance & Security. In MICRO, 2022.
  46. [46]Amir Naseredini, Martin Berger, Matteo Sammartino, and Shale Xiong. ALARM: Active LeArning of Rowhammer Mitigations. https://users.sussex.ac.uk/~mfb21/rh-draft.pdf, 2022.arxiv.org/abs/2211.16942
  47. [47]Biresh Kumar Joardar, Tyler K. Bletsch, and Krishnendu Chakrabarty. Machine Learning-based Rowhammer Mitigation. TCAD, 2022.DOI
  48. [48]Hasan Hassan, Ataberk Olgun, A Giray Yaglikci, Haocong Luo, and Onur Mutlu. A Case for Self-Managing DRAM Chips: Improving Performance, Efficiency, Reliability, and Security via Autonomous in-DRAM Maintenance Operations. arXiv:2207.13358, 2022.
  49. [49]Zhenkai Zhang, Zihao Zhan, Daniel Balasubramanian, Bo Li, Peter Volgyesi, and Xenofon Koutsoukos. Leveraging EM Side-Channel Information to Detect Rowhammer Attacks. In SP, 2020.DOI
  50. [50]Kevin Loughlin, Stefan Saroiu, Alec Wolman, and Baris Kasikci. Stop! Hammer Time: Rethinking Our Approach to Rowhammer Mitigations. In HotOS, 2021.
  51. [51]Fabrice Devaux and Renaud Arnygmac. Method and Circuit for Protecting a DRAM Memory Device from the Row Hammer Effect, 2021.
  52. [52]Ali Fakhrzadehgan, Yale N. Patt, Prashant J. Nair, and Moinuddin K. Qureshi. SafeGuard: Reducing the Security Risk from Row-Hammer via Low-Cost Integrity Protection. In HPCA, 2022.DOI
  53. [53]Stefan Saroiu, Alec Wolman, and Lucian Cojocar. The Price of Secrecy: How Hiding Internal DRAM Topologies Hurts Rowhammer Defenses. In IRPS, 2022.
  54. [54]Kevin Loughlin, Stefan Saroiu, Alec Wolman, Yatin A. Manerkar, and Baris Kasikci. MOESI-Prime: Preventing Coherence-Induced Hammering in Commodity Workloads. In ISCA, 2022.DOI
  55. [55]Jin Han, Jungsik Kim, Dafna Beery, K Deniz Bozdag, Peter Cuevas, Amitay Levi, Irwin Tain, Khai Tran, Andrew J Walker, Senthil Vadakupudhu Palayam, et al. Surround Gate Transistor With Epitaxially Grown Si Pillar and Simulation Study on Soft Error and Rowhammer Tolerance for DRAM. TED, 2021.DOI
  56. [56]Onur Mutlu, Ataberk Olgun, and A. Giray Yağlıkçı. Fundamentally Understanding and Solving RowHammer. In ASP-DAC, 2023.DOI
  57. [57]Jeonghyun Woo, Gururaj Saileshwar, and Prashant J Nair. Scalable and Secure Row-Swap: Efficient and Safe Row Hammer Mitigation in Memory Systems. In HPCA, 2023.DOI
  58. [58]Carsten Bock, Ferdinand Brasser, David Gens, Christopher Liebchen, and Ahmad-Reza Sadeghi. RIP-RH: Preventing Rowhammer-Based Inter-Process Attacks. In
  59. [59]Yicheng Wang, Yang Liu, Peiyun Wu, and Zhao Zhang. Discreet-PARA: Rowhammer Defense with Low Cost and High Efficiency. In ICCD, 2021.DOI
  60. [60]Tanj Bennett, Stefan Saroiu, Alec Wolman, and Lucian Cojocar. Panopticon: A Complete In-DRAM Rowhammer Mitigation. In Workshop on DRAM Security (DRAMSec), 2021.
  61. [61]Ataberk Olgun, Yahya Can Tugrul, Nisa Bostanci, Ismail Emir Yüksel, Haocong Luo, Steve Rhyner, Abdullah Giray Yaglikci, Geraldo F. Oliveira, and Onur Mutlu. ABA-CuS: All-Bank Activation Counters for Scalable and Low Overhead RowHammer Mitigation. In USENIX Security, 2024.DOI
  62. [62]F. Nisa Bostanci, Ismail Emir Yüksel, Ataberk Olgun, Konstantinos Kanellopoulos, Yahya Can Tugrul, A. Giray Yağlıkçı, Mohammad Sadrosadati, and Onur Mutlu. CoMeT: Count-Min-Sketch-Based Row Tracking to Mitigate RowHammer at Low Cost. In HPCA, 2024.arxiv.org/abs/2402.18769
  63. [63]Oğuzhan Canpolat, A Giray Yağlıkçı, Geraldo F Oliveira, Ataberk Olgun, Nisa Bostanci, Ismail Emir Yüksel, Haocong Luo, Oğuz Ergin, and Onur Mutlu. Chronus: Understanding and Securing the Cutting-Edge Industry Solutions to DRAM Read Disturbance. In HPCA, 2025.arxiv.org/abs/2502.12650
  64. [64]F Nisa Bostanci, Oğuzhan Canpolat, Ataberk Olgun, Ismail Emir Yüksel, Konstantinos Kanellopoulos, Mohammad Sadrosadati, Abdullah Giray Yağlıkçı, and Onur Mutlu. Understanding and Mitigating Covert Channel and Side Channel Vulnerabilities Introduced by RowHammer Defenses. In MICRO, 2025.
  65. [65]A. Giray Yağlıkçı, Yahya Can Tuğrul, Geraldo F De Oliveira, Ismail Emir Yüksel, Ataberk Olgun, Haocong Luo, and Onur Mutlu. Spatial Variation-Aware Read Disturbance Defenses: Experimental Analysis of Real DRAM Chips and Implications on Future Solutions. In HPCA, 2024.arxiv.org/abs/2402.18652
  66. [66]Suhas Vittal, Salman Qazi, Poulami Das, and Moinuddin Qureshi. MoPAC: Efficiently Mitigating RowHammer with Probabilistic Activation Counting. In ISCA, 2025.DOI
  67. [67]Moinuddin K. Qureshi. AutoRFM: Scaling Low-Cost In-DRAM Trackers to Ultra-Low Rowhammer Thresholds. In HPCA, 2025.DOI
  68. [68]Hritvik Taneja, Ali Hajiabadi, Michele Marazzi, Kaveh Razavi, and Moinuddin Qureshi. MRZRA: Efficiently Mitigating Rowhammer with Randomization and ALERT. In HPCA, 2026.DOI
  69. [69]Runjin Wu, Meng Zhang, You Zhou, Changsheng Xie, and Fei Wu. APT: Securing Against DRAM Read Disturbance via Adaptive Probabilistic In-DRAM Trackers. In ASPLOS, 2026.DOI
  70. [70]Chris S. Lin, Jeonghyun Woo, Prashant J. Nair, and Gururaj Saileshwar. CnC-PRAC: Coalesce, not Cache, Per Row Activation Counts for an Efficient in-DRAM Rowhammer Mitigation. In DRAMSec, 2025.arxiv.org/abs/2506.11970
  71. [71]Ataberk Olgun, F. Nisa Bostanci, Ismail Emir Yüksel, Oğuzhan Canpolat, Haocong Luo, Geraldo F. Oliveira, A. Giray Yağlıkçı, Minesh Patel, and Onur Mutlu. Variable Read Disturbance: An Experimental Analysis of Temporal Variation in DRAM Read Disturbance. In HPCA, 2025.arxiv.org/abs/2502.13075
  72. [72]JEDEC. JESD79-4C: DDR4 SDRAM Standard, 2020.
  73. [73]JEDEC. JESD79-5C: DDR5 SDRAM Standard, 2024.
  74. [74]Yoongu Kim, Vivek Seshadri, Donghyuk Lee, Jamie Liu, and Onur Mutlu. A Case for Subarray-Level Parallelism (SALP) in DRAM. In ISCA, 2012.
  75. [75]Donghyuk Lee, Yoongu Kim, Vivek Seshadri, Jamie Liu, Lavanya Subramanian, and Onur Mutlu. Tiered-Latency DRAM: A Low Latency and Low Cost DRAM Architecture. In HPCA, 2013.DOI
  76. [76]Donghyuk Lee, Yoongu Kim, Gennady Pekhimenko, Samira Khan, Vivek Seshadri, Kevin Chang, and Onur Mutlu. Adaptive-Latency DRAM: Optimizing DRAM Timing for the Common-Case. In HPCA, 2015.DOI
  77. [77]A. Giray Yağlıkçı, Haocong Luo, Geraldo F De Oliveira, Ataberk Olgun, Minesh Patel, Jisung Park, Hasan Hassan, Jeremie S Kim, Lois Orosa, and Onur Mutlu. Understanding RowHammer Under Reduced Wordline Voltage: An Experimental Study Using Real DRAM Devices. In DSN, 2022.arxiv.org/abs/2206.09999
  78. [78]Onur Mutlu. The RowHammer Problem and Other Issues We May Face as Memory Becomes Denser. In DATE, 2017.arxiv.org/abs/1703.00626
  79. [79]Onur Mutlu. RowHammer: Top Picks in Hardware and Embedded Security, 2018.
  80. [80]Onur Mutlu and Jeremie S Kim. RowHammer: A Retrospective. TCAD, 2019.
  81. [81]Oğuzhan Canpolat, A Giray Yağlıkçı, Geraldo F Oliveira, Ataberk Olgun, Oğuz Ergin, and Onur Mutlu. Understanding the Security Benefits and Overheads of Emerging Industry Solutions to DRAM Read Disturbance. DRAMSec, 2024.arxiv.org/abs/2406.19094
  82. [82]SAFARI Research Group. DRAM Bender – GitHub Repository. https://github.com/cmu-safari/DRAM-Bender, 2022.
  83. [83]Ataberk Olgun, Hasan Hassan, A Giray Yağlıkçı, Yahya Can Tuğrul, Lois Orosa, Haocong Luo, Minesh Patel, Oğuz Ergin, and Onur Mutlu. DRAM Bender: An Extensible and Versatile FPGA-based Infrastructure to Easily Test State-of-the-art DRAM Chips. TCAD, 2023.
  84. [84]Xilinx Inc. Xilinx Alveo U200 FPGA Board. https://www.xilinx.com/products/boards-and-kits/alveo/u200.html.
  85. [85]Maxwell. FT20X User Manual. https://www.maxwell-fa.com/upload/files/base/w/311.pdf.
  86. [86]Minesh Patel, Jeremie Kim, Taha Shahroodi, Hasan Hassan, and Onur Mutlu. Bit-Exact ECC Recovery (BEER): Determining DRAM On-Die ECC Functions by Exploiting DRAM Data Retention Characteristics. In MICRO, 2020.arxiv.org/abs/2009.07985
  87. [87]Minesh Patel, Geraldo Francisco de Oliveira Jr., and Onur Mutlu. HARP: Practically and Effectively Identifying Uncorrectable Errors in Main Memory Chips That Use On-Die ECC. In MICRO, 2021.DOI
  88. [88]Robert T Smith, James D Chlipala, JOHN FM Bindels, Roy G Nelson, Frederick H Fischer, and Thomas F Mantz. Laser Programmable Redundancy and Yield Im provement in a 64K DRAM. JSSC, 1981.
  89. [89]provement in a 64K DRAM. JSSC, 1981.
  90. [90]B. Keith and R.J. Baker. DRAM Circuit Design: A Tutorial. Wiley, 2001.
  91. [91]Kiyoo Itoh. VLSI Memory Chip Design. Springer, 2001.
  92. [92]Jamie Liu, Ben Jaiyen, Yoongu Kim, Chris Wilkerson, and Onur Mutlu. An Experimental Study of Data Retention Behavior in Modern DRAM Devices. In ISCA, 2013.
  93. [93]Vivek Seshadri, Thomas Mullins, Amirali Boroumand, Onur Mutlu, Phillip B Gib-bons, Michael A Koc zuh, and Todd C Mowry. Gather-Scatter DRAM: In-DRAM Address Translation to Improve the Spatial Locality of Non-Unit Strided Accesses. In MICRO, 2015.
  94. [94]Samira Khan, Donghyuk Lee, and Onur Mutlu. PARBOR: An Efficient System-Level Technique to Detect Data-Dependent Failures in DRAM. In DSN, 2016.
  95. [95]Samira Khan, Chris Wilkerson, Zhe Wang, Alaa R Alameldeen, Donghyuk Lee, and Onur Mutlu. Detecting and Mitigating Data-Dependent DRAM Failures by Exploiting Current Memory Content. In MICRO, 2017.
  96. [96]Donghyuk Lee, Samira Khan, Lavanya Subramanian, Saugata Ghose, Rachata Ausavarungnirun, Gennady Pekhimenko, Vivek Seshadri, and Onur Mutlu. Design-Induced Latency Variation in Modern DRAM Chips: Characterization, Analysis, and Latency Reduction Mechanisms. In SIGMETRICS, 2017.
  97. [97]Andrei Tatar, Cristiano Giuffrida, Herbert Bos, and Kaveh Razavi. Defeating Software Mitigations Against Rowhammer: A Surgical Precision Hammer. In RAID, 2018.
  98. [98]Alessandro Barenghi, Luca Breveglieri, Niccolò Izzo, and Gerardo Pelosi. Software-Only Reverse Engineering of Physical DRAM Mappings for Rowhammer Attacks. In IVSW, 2018.
  99. [99]Lucian Cojocar, Jeremie Kim, Minesh Patel, Lillian Tsai, Stefan Saroiu, Alec Wolman, and Onur Mutlu. Are We Susceptible to Rowhammer? An End-to-End Methodology for Cloud Providers. In S&P, 2020.
  100. [101]Samira Khan, Donghyuk Lee, Yoongu Kim, Alaa R Alameldeen, Chris Wilkerson, and Onur Mutlu. The Efficacy of Error Mitigation Techniques for DRAM Retention Failures: A Comparative Experimental Study. In SIGMETRICS, 2014.
  101. [102]Haocong Luo, Ismail Emir Yüksel, Ataberk Olgun, A Giray Yağlıkçı, Mohammad Sadrosadati, and Onur Mutlu. An Experimental Characterization of Combined RowHammer and RowPress Read Disturbance in Modern DRAM Chips. DSN Disrupt, 2024.
  102. [103]Yacan Can Tugrul, A. Giray Yaglikci, Ismail Emir Yuksel, Ataberk Olgun, Oğuzhan Canpolat, Nisa Bostanci, Mohammad Sadrosadati, Oguz Ergin, and Onur Mutlu. Understanding RowHammer Under Reduced Refresh Latency: Experimental Analysis of Real DRAM Chips and Implications on Future Solutions. In HPCA, 2025.
  103. [105]Ismail Emir Yuksel, Akash Sood, Ataberk Olgun, Oğuzhan Canpolat, Haocong Luo, Nisa Bostanci, Mohammad Sadrosadati, Giray Yaglikci, and Onur Mutlu. PuDHammer: Experimental Analysis of Read Disturbance Effects of Processing-using-DRAM in Real DRAM Chips. In ISCA, 2025.
  104. [106]Ismail Emir Yuksel, Yahya Can Tugrul, F. Nisa Bostanci, Geraldo F. de Oliveira, A. Giray Yaglickci, Ataberk Olgun, Melina Soysal, Haocong Luo, Juan Gomez Luna, Mohammad Sadrosadati, and Onur Mutlu. Simultaneous Many-Row Activation in Off-the-Shelf DRAM Chips: Experimental Characterization and Analysis. In DSN, 2024.
  105. [107]Ismail Emir Yuksel, Yahya Can Tugrul, Ataberk Olgun, F. Nisa Bostanci, A. Giray Yaglickci, Geraldo F. de Oliveira, Haocong Luo, Juan Gomez Luna, Mohammad Sadrosadati, and Onur Mutlu. Functionally-Complete Boolean Logic in Real DRAM Chips: Experimental Characterization and Analysis. In HPCA, 2024.
  106. [108]Fei Gao, Georgios Tziantzioulis, and David Wentzlaff. ComputeDRAM: In-Memory Compute Using Off-the-Shelf DRAMs. In MICRO, 2019.
  107. [109]Longda Zhou, Jie Li, Zheng Qiao, Pengpeng Ren, Zixuan Sun, Jianping Wang, Blacksmith Wu, Zhigang Ji, Runsheng Wang, Kanyu Cao, and Ru Huang. Double-sided Row Hammer Effect in Sub-20 nm DRAM: Physical Mechanism, Key Features and Mitigation. In IRPS, 2023.
  108. [110]Minesh Patel, Jeremie S Kim, Hasan Hassan, and Onur Mutlu. Understanding and Modeling On-Die Error Correction in Modern DRAM: An Experimental Study Using Real Devices. In DSN, 2019.
  109. [111]Hye-Jung Kwon, Eunsung Seo, Chan-Yong Lee, Young-Hun Seo, Gong-Heum Han, Hye-Ran Kim, Jong-Ho Lee, Min-Su Jang, Sung-Geun Do, Seung-Hyun Cho, et al. 23.4 An extremely low-standby-power 3.733 Gb/s/pin 2Gb LPDDR4 SDRAM for wearable devices. In ISSCC, 2017.
  110. [112]Nohhyup Kwak, Saeng-Hwan Kim, Kyong Ha Lee, Chang-Ki Baek, Mun Seon Jang, Yonguk Joo, Seung-Hun Lee, Woo Young Lee, Eun-yeong Lee, and Donghee Han. 23.3 A 4.8 Gb/s/pin 2Gb LPDDR4 SDRAM with Sub-100μA Self-Refresh Current for IoT Applications. In ISSCC, 2017.
  111. [114]Haocong Luo, Yahya Can Tugrul, F. Nisa Bostanci, Ataberk Olgun, A Giray Yağlıkçı, , and Onur Mutlu. Ramulator 2.0: A Modern, Modular, and Extensible DRAM Simulator. IEEE CAL, 2024.
  112. [115]SAFARI Research Group. Ramulator V2.0. https://github.com/CMU-SAFARI/ramulator2.
  113. [116]Standard Performance Evaluation Corp. SPEC CPU 2006. http://www.spec.org/cpu2006/, 2006.
  114. [117]Standard Performance Evaluation Corp. SPEC CPU 2017. http://www.spec.org/cpu2017, 2017.
  115. [118]Transaction Processing Performance Council. TPC Benchmarks. http://tpc.org/.
  116. [119]Jason E. Fritts, Frederick W. Steiling, Joseph A. Tucek, and Wayne Wolf. MediaBench II Video: Expediting the Next Generation of Video Systems Research. MICRO, 2009.
  117. [120]Brian Cooper, Adam Silberstein, Erwin Tam, Raghu Ramakrishnan, and Russell Sears. Benchmarking Cloud Serving Systems with YCSB. In SoCC, 2010.
  118. [121]Onur Mutlu and Thomas Moscibroda. Small-Time Fair Memory Access Scheduling for Chip Multiprocessors. In MICRO, 2007.
  119. [122]Dimitris Kaseridis, Jeffrey Stuecheli, and Lizy Kurian John. Minimalist Open-Page: A DRAM Page-Mode Scheduling Policy for the Many-Core Era. In MICRO, 2011.
  120. [123]Jumin Kim, Seungmin Baek, Hwayoung Nam, Minbok Wi, Nam Sung Kim, and Jung Ho Ahn. PVCA: A Rowhammer Mitigation Architecture Exploiting Per-Victim-Row Counting. In ISCA, 2026.
  121. [124]Tutu Ajayi, Vidya A Chhabria, Mateus Fogaça, Soheil Hashemi, Abdelrahman Hosny, Andrew B Kahng, Minsoo Kim, Jeongsup Lee, Uday Mallappa, Marina Neseem, et al. Toward an open-source digital flow: First learnings from the openroad project. In DAC, 2019.
  122. [125]NanGate Inc. Nangate 45nm Open Cell Library. https://github.com/oscc-ip/nangate.
  123. [126]Hasan Hassan, Ataberk Olgun, A Giray Yağlıkçı, Haocong Luo, and Onur Mutlu. Self-Managing DRAM: A Low-Cost Framework for Enabling Autonomous and Efficient DRAM Maintenance Operations. In MICRO, 2024.
  124. [127]Intel Corporation. 10th Generation Intel Core Processor Families Datasheet, Volume 2. https://www.intel.com/content/dam/www/public/us/en/documents/datasheets/10th-gen-core-families-datasheet-vol-2-datasheet.pdf.
  125. [128]Advanced Micro Devices, Inc. AMD Ryzen Master User Guide. https://docs.amd.com/r/en-US/68886--ryzen-master-user-guide/RAM. Document 68886, Section: RAM.
  126. [129]Theodore Z Schoenborn and Christopher P Mozak. Memory Subsystem I/O Performance Based on In-System Empirical Testing, 2017. US Patent 9,536,626.
  127. [130]Grant Evan Lev, Jayesh Hari Joshi, Amitabh Mehra, Jerry Anton Ahrens, Joshua Taylor Knight, Anil Harwani, and William Robert Alverson. Real Time Profile Switching for Memory Overclocking, 2024. US Patent 11,977,757.
  128. [131]Chulseung Lim, Guengyue Park, and Sanghyeon Baeg. Active Precharge Hammering to Monitor Displacement Damage Using High-Energy Protons in 3x-nm SDRAM. TNS, 2017.
  129. [132]Kyungbae Park, Donghyuk Yun, and Sanghyeon Baeg. Statistical Distributions of Row-Hammering Induced Failures in DDR3 Components. Microelectronics Reliability, 2016.
  130. [133]Kyungbae Park, Chulseung Lim, Donghyuk Yun, and Sanghyeon Baeg. Experiments and Root Cause Analysis for Active-Precharge Hammering Fault in DDR3 SDRAM under 3xnm Technology. Microelectronics Reliability, 2016.
  131. [134]Seong-Wan Ryu, Kyungkyu Min, Jungho Shin, Heimi Kwon, Donghoon Nam, Taekyung Oh, Tae-Su Jang, Minsoo Yoo, Yongtaik Kim, and Sungjoo Hong. Overcoming the Reliability Limitation in the Ultimately Scaled DRAM using Silicon Migration Technique by Hydrogen Annealing. In IEDM, 2017.
  132. [135]Donghyuk Yun, Myungsang Park, Chulseung Lim, and Sanghyeon Baeg. Study of TID Effects on One Row Hammering using Gamma in DDR4 SDRAMs. In IRPS, 2018.
  133. [136]Chulseung Lim, Kyungbae Park, Geunyoung Bak, Donghyuk Yun, Myungsang Park, Sanghyeon Baeg, Shi-Jie Wen, and Richard Wong. Study of Proton Radiation Effect to Row Hammer Fault in DDR4 SDRAMs. Microelectronics Reliability, 2018.
  134. [137]Zhenrong Lang, Patrick Jattke, Michele Marazzi, and Kaveh Razavi. Blaster: Characterizing the blast radius of rowhammer. In DRAMSec, 2023.
  135. [138]Hwayoung Nam, Seungmin Baek, Minbok Wi, Michael Jaemin Kim, Jaehyun Park, Chihun Song, Nam Sung Kim, and Jung Ho Ahn. Dramscope: Uncovering DRAM Microarchitecture and Characteristics by Issuing Memory Commands. ISCA, 2024.
  136. [139]Haocong Luo, Ismail Emir Yuksel, Ataberk Olgun, A Giray Yağlıkçı, and Onur Mutlu. Revisiting DRAM Read Disturbance: Identifying Inconsistencies Between Experimental Characterization and Device-Level Studies. In VLSI, 2025.
  137. [140]Wei He, Zhi Zhang, Yueqiang Cheng, Wenhao Wang, Wei Song, Yansong Gao, Qifei Zhang, Kang Li, Dongxi Liu, and Surya Nepal. WhistleBlower: A System-level Empirical Study on RowHammer. TC, 2023.
  138. [141]M.K. Qureshi, Dae-Hyun Kim, S. Khan, P.J. Nair, and O. Mutlu. AVATAR: A Variable-Retention-Time (VRT) Aware Refresh for DRAM Systems. In DSN, 2015.
  139. [142]Minesh Patel, Jeremie S Kim, and Onur Mutlu. The Reach Profiler (REAPER): Enabling the Mitigation of DRAM Retention Failures via Profiling at Aggressive Conditions. In ISCA, 2017.

Paper details

vlnc:W_s25hm7hh
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents