betaFeedback

WorksEngineering

AI Inference-Aware Optimal Power Flow: Converting Token Requirements into Grid Flexibility

Abstract

The rapid growth, temporal variability, and geographic concentration of electricity demand from artificial intelligence (AI) inference can aggravate transmission congestion near data centers (DCs). The spatial and temporal flexibility of inference workloads offers an opportunity to alleviate this congestion: requests can be routed across DCs or deferred within users' quality-of-service (QoS) requirements. However, inference routing primarily considers graphics processing unit (GPU) availability, leaving this flexibility underused in grid operation. Using GPU measurements, this paper models inference latency and DC power consumption to quantify the flexibility available under users' time-to-first-token (TTFT) and inter-token latency (ITL) requirements. These models are then incorporated into an AI inference-aware optimal power flow (OPF) that coordinates inference routing and power system operation among DC operators, the transmission system operator, and inference users. By linking user requests, token generation at DCs, and electricity delivery through the transmission network, the formulation schedules inference workloads subject to both users' QoS requirements and transmission constraints. Simulations on the PJM system show that, compared with computing-based routing and with electricity price-based routing, coordinating inference routing with network operation enables growing DC electricity demand deliverable to be met at lower cost.

Paper details

vlnc:W_edjzdfxg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

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

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…