betaFeedback

WorksPhysics and Astronomy

Optimizing an RFSoC LLRF Controller for a Low-PRF, Short Pulse System

Abstract

The Advanced Light Source (ALS) LINAC operates with 3 GHz RF pulses at 0.7 Hz pulse repetition frequency (PRF) and 1.2 microsecond duration, imposing severe limitations on traditional PID and adaptive feedforward control schemes. An RFSoC-based controller prototype was developed and presented in prior work to upgrade the legacy LLRF systems for the LINAC and Subharmonic Buncher. This paper presents sampling scheme optimizations and interleaving calibration techniques for RFSoC based control of pulsed RF systems. We detail bench characterization methodologies to isolate component-level phase noise across individual ADC, DAC, and clock components in the Xilinx ZCU208 hardware ecosystem. Finally, we explore translating open-loop controller bench measurements into controller-added closed-loop LLRF performance predictions.

Introduction

ALS is currently upgrading the LINAC and Subharmonic Buncher Low Level RF control systems. The RFSoC-based platform was described in previous work [1] and the PYNQ-based firmware design is available on GitHub [2]. Initial RF characterizations and clocking optimizations indicated performance comparable to other digital LLRF platforms for normally conducting cavities [3]. Recent work has focused on developing measurement metrics to evaluate if the RFSoC-based platform can achieve the performance required for the LINAC application.

ALS LINAC System

The two 3 GHz LINAC RF stations are pulsed at 0.7 Hz, 1.2 µs pulses. The 400 W pre-amplifier and the 17 MW klystron are currently run in saturation with an open loop legacy LLRF analog amplitude and phase control. The very low Pulse Repetition Frequency (PRF) caused the Thales klystrons to arc during initial commissioning. System optimizations required using the cathode pulse end to control the tail of the RF pulse to widen the pulse length without arcing. Future upgrades may include increasing the cathode PRF to reduce arcing and adding a downstream SLAC Energy Doubler (SLED) to widen the RF pulse.

The RFSoC controller prototype was used to characterize the existing RF waveforms in the Accelerating Sections (AS) cavities 1 and 2. From these initial measurements we observed a highly non-linear, but repetitive phase profile through the pulse, confirming the need for improved phase control. The new system will provide open loop drive and closed loop phase control to avoid interfering with the existing high power RF system tuning.

ALS LINAC Accelerating Sections 1 and 2 FWD and REV Waveforms

Figure 1. ALS LINAC Accelerating Sections 1 and 2 FWD and REV Waveforms

LINAC LLRF Performance Requirements

The project requested 0.1° RMS RF field stability in the 3 GHz normal conducting cavities. Future work will quantify open-loop performance of the high-power RF system. Due to the low PRF and short pulses, the control system’s ability to correct for high-power RF errors is limited, so the LLRF platform inherent noise performance is critical. The system front end RF characterization showed a negligible affect on system linearity or stability [1], so current effort is focused on characterizing the ZCU208 platform, RF-identical to the modified ’LBL208’ platform that will be used in production. Among other minor optimizations, the LBL208 swaps the XCZU48DR-2FSVG1517 chip for the cheaper XCZU47DR-1FSVG1517 and is currently in production for the ALS-U BPM systems. [4].

Project Status

The team is manufacturing chassis hardware and implementing the required custom python software on top of the PYNQ framework and our custom IP design, utilizing our standard 250 MHz LLRF DSP blocks. Commissioning is planned before the upcoming ALS Dark Time for the ALS 2.0 Upgrade, October 2027.

ADC Interleaving Spur Calibration

Background

The RFSoC achieves GS/s sampling rates by interleaving multiple ADCs into tiles. Imperfections and mismatches between the sub ADCs manifest as spurs at k∗Fs/Nk * F_s/N, where FsF_s is the composite sampling rate of the RF-ADC, NN is the number of sub-RF-ADCs interleaved together, and k=0…N−1k = 0 \ldots N - 1. The RFSoC provides calibration algorithms to correct for DC offsets, gain differences, and sampling clock time skew between the sub ADCs.

The foreground calibration runs as part of the RF-ADC initialization state machine sequence to correct the DC offsets, so there should be no signal energy present at k∗Fs/Nk * F_s/N during ADC startup.

After the RF-ADC initializes, the background calibration updates the corrections as the environment changes. The background calibration process requires an input signal of -40 dBFS, and blocks require 7 μ7\ \mus to change state. For our 1.2 μ1.2\ \mus pulse length application, we will calibrate the ADCs with a test signal and then freeze these calibration weights, disabling the background calibration.

Interleaving Spur Locations

ImperfectionSpur Location
DC offsetk∗Fs/Nk * F_s/N
Gain+/−fin+(k/N)∗Fs+/-f_{\mathrm{in}} + (k/N) * F_s
Time skew+/−fin+(k/N)∗Fs+/-f_{\mathrm{in}} + (k/N) * F_s

Table 1. RFSoC Interleaving Spurs.

Digital downconversion shifts the residual gain and timing skew spurs to baseband.

In the RFSoC, the FIR filter and the rate-reduction sample decimation are implemented as a polyphase filter which performs FIR filtering on the samples that will be preserved through decimation. This limits the unfiltered in-band content to −fs2M-\frac{f_s}{2M} to fs2M\frac{f_s}{2M}, and if fs2M≤∣k∗fsN∣\frac{f_s}{2M} \le\left|k * \frac{f_s}{N}\right|, where MM is the decimation factor, all of the gain and timing skew spurs will be filtered before decimations folds them into baseband.

With the background calibration enabled, the timing and gain skew spurs were suppressed below the noise floor in our measurements.

If the fRF=A∗fs2Mf_{RF} = \frac{A*f_s}{2M}, where AA is an integer, then the filtered timing and gain skew spurs will fold directly onto DC in a zero-IF architecture, preventing the background calibration from detecting the skew between sub-adcs. Uncalibrated, the filtered but uncalibrated skew spurs measured 60-80 dBc near 3 GHz (7.5-75 fsf_s at 3 GHz). In these schemes the background calibration can be performed with a test signal near fRFf_{RF}.

With our 4 GS/s fsf_s and 3 GHz fRFf_{RF} scheme, performing a background calibration with the NCO offset by a few MHz suppresses the gain and timing skew spurs below 90 dBc for nominal operation: a residual 1.7 fsf_s phase error. Future work will move to a 5 GS/s fsf_s to mitigate this issue.

Characterization Measurements

ADC 3 GHz Spectrum

With the frozen, offset interleaving calibration in place, the ADC spectrum can be evaluated: The 62.5 MHz spurs will be investigated for wideband diagnostic performance improvement in future work.

Baseband spectrum, 3 GHz RF sampled at 4 GS/s

Figure 2. Baseband Spectrum, 3GHz RFINRF_{\mathrm{IN}}, 4 GS/s fsf_s

DAC 3 GHz Spectrum

DAC spectrum, 3 GHz RF output, 4 GS/s sampling rate

Figure 3. DAC Spectrum, 3GHz RFOUTRF_{\mathrm{OUT}}, 4 GS/s fsf_s, with 70 dBc 759 kHz modulating spurs

Sampling Clock Phase Noise

The Xilinx CLK104 board was modified to the ’CLK104A’: adding 0Ω\Omega jumpers to connect the J11 ’external clock input’ connector to the LMK04828 to be used as a buffer and a divider for clock generation [3]. For the 4 GS/s LLRF design, the LMX2594 synthesizer was optimized for low jitter. The ALS BPM team is also using the CLK104A, configured differently to allow the recovered clock from the timing link to be used as the time base reference [4].

Simplified CLK104A diagram

Figure 4. Simplified CLK104A Diagram

The CLK104A and DAC phase noise measurements were made with the Rhode and Schwartz FSWP8 Phase Noise Analyzer, calibrated to achieve <1<1 fs added jitter. The LMX1204 is a very low phase noise clock divider, <1<1 fs as a 3 GHz buffer, which combined with the FSWP enables an additive phase noise measurements across frequency.

CLK104A Phase Noise Measurement Setup

Figure 5. CLK104A Phase Noise Measurement Setup

CLK104A Additive PN, 500MHz RFIN, 4GS/s fs

Figure 6. CLK104A Additive PN, 500MHz RFINRF_{\mathrm{IN}}, 4GS/s fsf_s 61.8 fs [1 Hz, 1 MHz]

The CLK104A sampling clock generation impacts the phase noise performance but is required for the current ALS MO architecture, which distributes 500 MHz RF over coax to control stations.

DAC Phase Stability

Sampling clock generation dominates the DAC phase noise performance.

DAC + CLK104A Phase Noise Measurement Setup

Figure 7. DAC + CLK104A Phase Noise Measurement Setup

DAC + CLK104A 500MHz RFIN, 4GS/s fs, 3GHz RFOUT

Figure 8. DAC + CLK104A 500MHz RFINRF_{\mathrm{IN}}, 4GS/s fsf_s, 3GHz RFOUTRF_{\mathrm{OUT}} 60.9 fs [1 Hz, 1 MHz]

DAC Phase Noise Measurement Setup

Figure 9. DAC Phase Noise Measurement Setup

DAC Additive PN, 4GS/s fs, 4GHz RFOUT

Figure 10. DAC Additive PN, 4GS/s fsf_s, 4GHz RFOUTRF_{\mathrm{OUT}} 13.5 fs [1 Hz, 1 MHz]

ADC Phase Stability - Measurement Background

ADC characterization is limited by the capture depth and rate. Current RFSoC firmware allows a capture depth of 32,768 samples (131.07 μ\mus with the 250 MHz DSP clock) and a maximum capture rate of ∼37\sim37 Hz. Therefore these ADC characterization measurements of a CW source are pulsed performance results and must be characterized by the intrapulse and interpulse performance. Intrapulse performance measures the stability of a single pulse, and can be evaluated as the mean of the standard deviations of the pulses.

1τpulse≤f≤BWbaseband\frac{1}{\tau_{\mathrm{pulse}}} \le f \le\mathrm{BW}_{\mathrm{baseband}}

where τpulse\tau_{\mathrm{pulse}} is the pulse duration and BWbaseband\mathrm{BW}_{\mathrm{baseband}} is the baseband bandwidth.

Interpulse measures the stability between pulses, and is characterized by the standard deviation of the pulse average values. Averaging within the pulse suppresses the intrapulse noise by the number of samples per capture.

Averaging over a gate width τgate\tau_{\mathrm{gate}} acts as a boxcar filter in the time domain, a sinc filter response in the frequency domain with a 3dB cutoff of ≈0.443τgate\approx\frac{0.443}{\tau_{\mathrm{gate}}}. Frequency content up to fcapture2\frac{f_{\mathrm{capture}}}{2} is directly measured and fcapture2\frac{f_{\mathrm{capture}}}{2} to 0.443τgate\frac{0.443}{\tau_{\mathrm{gate}}} is aliased into the measurement.

Therefore the interpulse measurement describes the noise spectra of:

PRFNcaptures≤f≤0.443τgate\frac{P_{\mathrm{RF}}}{N_{\mathrm{captures}}} \le f \le\frac{0.443}{\tau_{\mathrm{gate}}}

For these ALS capture measurement sets of 131.07μ\mus pulses, 350 captures at 35 Hz, the interpulse jitter captures: 0.1Hz ≤f≤\le f \le 3.38kHz and the intrapulse jitter: 7.63kHz ≤f≤\le f \le 125MHz Future firmware upgrades will increase the capture rate by streaming to the DDR4, enabling complete characterization of the ADC phase noise across frequency offsets.

ADC Phase Stability - Measurements

The ADC measures φADC=φfs−φfRF\varphi_{\mathrm{ADC}} = \varphi_{f_s} - \varphi_{f_{\mathrm{RF}}}, so the ADC and CLK104A noise can be characterized by using the same reference signal to generate the sampling clock and the RF input. Since the CLK104A synthesizer PLL loop is ∼\sim250 kHz, the CLK104A added noise contributes primarily to the intrapulse noise, with a small broadband jitter increase from the buffer.

The ADC noise is further evaluated by determining the uncorrelated noise between two channels fed the same RF signal, a technique used to characterize the fundamental receiver noise floor [5]:

ADC Uncorrelated Noise Measurement Setup

Figure 11. ADC Uncorrelated Noise Measurement Setup

ADC Uncorrelated Noise

Figure 12. ADC Uncorrelated Noise

The Pearson cross-correlation coefficient predicts the shared noise between the ADCs:

ρ12=Cov⁡(y1,y2)σ1σ2=σshared2σ1σ2\rho_{12} = \frac{\operatorname{Cov}(y_1,y_2)}{\sigma_1\sigma_2} = \frac{\sigma_{\mathrm{shared}}^2}{\sigma_1\sigma_2}

And the variance of the normalized difference between the two ADCs predicts the ADC noise contribution:

σuncorr=Var⁡(ADC1−ADC2)2\sigma_{\mathrm{uncorr}} = \sqrt{\frac{\operatorname{Var}(\mathrm{ADC}_1-\mathrm{ADC}_2)}{2}}
ParameterInterpulseIntrapulseTotal
ADC 1 σ1\sigma_144 fs73 fs85 fs
ADC 2 σ2\sigma_241 fs76 fs86 fs
Correlation Coefficient (ρ12\rho_{12})+94%+67%—
Residual ADC Noise (σuncorr\sigma_{\mathrm{uncorr}})10 fs43 fs44 fs
Sampling Clock Jitter (σclk\sigma_{\mathrm{clk}})——61 fs
Weighted ADC + CLK43 fs75 fs86 fs

Table II. ADC Noise for 3 GHz Input, 4 GS/s fsf_s

The alignment between the measured ADC + CLK104A phase noise and the correlation-weighted sum of the individual measurements provides some validation of the result.

Phase Noise Summary

ComponentJitterEvaluation Freq
CLK104A62 fs4 GHz
DAC + CLK104A61 fs4 GHz
DAC13 fs4 GHz
ADC + CLK104A86 fs3 GHz
ADC44 fs3 GHz

Table III. RFSoC Phase Noise Components

The CLK104A and DAC were evaluated at 4 GHz, the current fsf_s, to enable the FSWP additive measurements. The ADC was evaluated at 3 GHz, the LINAC RF input.

Control Schemes

PI Intrapulse Feedback

The classic Low Level RF control system use Proportional-Integral (PI) feedback to track the amplitude and phase of the RF cavity voltage and adjust the RF drive signal to correct for errors [6]. The system latency limits the bandwidth and response speed of the PI loop.

Delay SourceValue
Phase lag due to delay360∘∗BWloop∗latency360^\circ* BW_{\mathrm{loop}} * latency
Phase PI integrator90∘90^\circ
Phase lag cavityatan(BWloop/(BWcavity/2))∗180/πatan(BW_{\mathrm{loop}}/(BW_{\mathrm{cavity}}/2)) * 180/\pi

Table IV. Phase Delays.

The phase margin of the PI loop is 180∘−delaysum180^\circ- delay_{\mathrm{sum}} and should be >45∘> 45^\circ to ensure loop stability and maximize the loop bandwidth. For our estimated 450 ns total system latency (290 ns from the RFSoC [1]) and 0.4μ\mu cavity fill time, this limits our maximum loop bandwidth to ∼\sim100 kHz. The settling time for this loop is 3-4τloop\tau_{\mathrm{loop}}, where

τloop=12π⋅100 kHz≈1.6 μs\tau_{\mathrm{loop}} = \frac{1}{2\pi\cdot100\ \mathrm{kHz}} \approx1.6\ \mu\mathrm{s}

Therefore this loop would require approximately 4.8 to 6.4 μ\mus to regulate and is not useful for this short pulse case.

Static Feedforward, and Adaptive Feedforward

A combination of static feedforward (SFF) and adaptive feedforward (AFF) can reduce repetitive, synchronous errors in short pulse systems.

SFF is generated from a one-time calibration and is as the starting pulse shape applied by the arbitrary waveform generating (AWF) functionality of the DAC. This pulse shape is the inverse of the typical intrapulse error pattern of the pulse.

The AFF loop operates pulse-by-pulse to adjust the pulse shape as the error profile mutates due to slow changes. Stability requires a Q filter that performs intrapulse time averaging to limit the AFF’s response to high-frequency noise:

Un+1(t)=Un(t)+(q∗En)(t)U_{n+1}(t) = U_n(t) + (q * E_n)(t)

Where UnU_n is the drive waveform and EnE_n is the error waveform. To ensure that the AFF only attempts to correct unaliased errors, the BW of the Q filter should be limited to frequency variations that fall below the pulse-to-pulse Nyquist limit (half the pulse repetition rate, fPRF/2f_{\mathrm{PRF}}/2).

This limits the AFF correction feature size to 0.35/BWQ0.35/BW_Q. For ALS at 0.7 Hz : 1s, limiting the AFF to DC corrections of the pulse average value.

Bounding Relevant Noise Sources

Tx Noise

Noise in the Tx chain will be injected into the cavity within its bandwidth and then considered as part of the error signal in a feedback system. With a PI loop, this sets the lower bound of relevant Tx noise to PIENBWPI_{\mathrm{ENBW}}. The upper bound is set by the ENBW of the RF cavity, 1/2tf1/2t_f. For the ALS LINAC open-loop case,

0 Hz≤f≤1.25 MHz0\ \mathrm{Hz} \leq f \leq1.25\ \mathrm{MHz}

The high-power RF system with a 1-2 MHz BW is sensitive to narrowband Tx spurs, and the normal conducting RF cavity applies ∼\sim6 dB of suppression to the closest spurs. Through the direct modulation of the drive signal, AM-PM generation in the klystron, and IM3 products from the klystron, these spurs add ∼\sim18 fs to the Tx error.

Noise in the Rx chain is an error source overlaid on the signal from the cavity, indistinguishable from true error. In our DC-only interpulse scheme, the Rx noise is suppressed by averaging of the waveforms used for SFF calibration and intrapulse averaging of the feedforward DC error.

Conclusion

ParameterPhase ErrorTiming Jitter
DAC Phase Noise0.046∘0.046^\circ RMS43 fs RMS
TX Spurs0.020∘0.020^\circ RMS18 fs RMS
Open-Loop Phase Error (RSS)0.050∘0.050^\circ RMS46 fs RMS
Diagnostic Phase Error (RSS)0.093∘0.093^\circ RMS86 fs RMS

Table V. RFSoC LLRF Added Error Estimates

For our open-loop controller case, the RFSoC added noise is 46 fs, 25% of our cavity field error budget. The 4 GS/s waveform capture currently achieves 86 fs, <0.1∘<0.1^\circ added phase error. In addition to the 3 GHz LINAC case described here, this controller will be used to upgrade the 125 MHz and 500 MHz Subharmonic Bunchers with longer RF pulses enabling intrapulse feedback [1]. Operational data from these installed systems is expected next year.

Acknowledgment

This work is supported by the Office of Science, Office of Basic Energy Sciences, of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231. Qiang Du led the LLRF RFSoC platform development with Shree Murthy, who optimized the system clocking architecture. Mike Chin designed the LBL208 and generously provided test hardware, ADC insights, and lab camaraderie. Our director Ben Flugstad drives us to improve the ALS through innovation, and our manager Angel Jurado leads the LINAC Upgrade project.

References

  1. [1]Q. Du, S. Murthy, V. Moore, A. Jurado Lopez, M. Chin, S. S. Basak, D. Nett, and B. Flugstad, “RFSoC based LLRF system design at ALS,” arXiv preprint arXiv:2510.13711, 2025.
  2. [2]Q. Du and S. Murthy, “RFSoC based MIMO System,” https://github.com/berkeleylab/lbnl-dsllrf.
  3. [3]S. Murthy, V. Moore, Q. Du, A. Jurado, K. Penney, M. Chin, D. Nett, and B. Flugstad, “Comparative Evaluation of Xilinx RFSoC Platform for Low-Level RF Systems,” in Low Level RF workshop 2025, Newport News, VA, USA, October 2025.arxiv.org/abs/2510.13711
  4. [4]M. J. Chin, L. M. Russo, N. Ly, V. Moore, S. Murthy, G. Portmann, and J. M. Weber, “Beam Position Monitor Electronics for ALS-U Using RFSoC,” in Proc. Int. Beam Instrum. Conf. (IBIC), paper MORT07. [Online]. Available: https://indico.jacow.org/event/113/contributions/14530/
  5. [5]L. Doolittle, G. Bell, C. Serrano, Q. Du, and G. Huang, “LLRF system design and performance for LCLS-II,” in Proc. SRF’17, ser. International Conference on RF Superconductivity, no. 18. JACoW Publishing, Geneva, Switzerland, Oct. 2017, paper FRBA02, pp. 923–928. [Online]. Available: https://proceedings.jacow.org/srf2017/papers/frba02.pdfDOI
  6. [6]S. Simrock and M. Grecki, Low-Level Radio Frequency Systems. Springer Cham, 2022.DOI

Paper details

vlnc:W_swvt4nn7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 4

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents