Search arXivSearch

arXiv subjects

Hugo Budd

Publications and source records attributed to Hugo Budd.

2 recordsLinked to original sources

EasyRider: Mitigating Power Transients in Datacenter-Scale Training Workloads

Large-scale AI model training workloads use thousands of GPUs operating in tightly synchronized loops. During synchronous communication, start-up, shut-down, and checkpointing, GPU power consumption can swing from peak to idle within milliseconds. Such steep power ramp rates induce reactive power transients, leading to voltage and frequency shifts that can damage transformers, generators, and protection equipment on the broader power grid. To solve this problem, we introduce EasyRider, a power architecture to mitigate power fluctuations at the rack level. EasyRider uses passive and active hardware components to attenuate rack power swings and rack-scale energy storage for the large amounts of energy needed to smooth high-power racks. A software system continually monitors the energy storage system to maximize its lifetime in the presence of frequent charge/discharge cycles. EasyRider filters rack power variations to be within grid safety requirements without requiring software modifications to AI training frameworks or wasting energy. We evaluate EasyRider on a 10kW/400VDC-rated rack-scale prototype system, demonstrating its effectiveness across heterogeneous power levels and workload power profiles.

cs.AR

Active Filter Design for Buffering Datacenter-Scale Power Fluctuations from Training AI

Training large AI models requires thousands of processors working in parallel. This synchronized load poses a challenge for traditional power delivery architectures, because IT power ramps much faster than grid hardware can compensate. This paper presents a hardware solution to this problem, introducing a novel power delivery architecture which automatically detects changes in load power and compensates using on-rack energy storage, giving time for the grid to respond to the varying load. The architecture is validated experimentally, powering a training load while limiting the grid-side power ramp rate such that it stays within pre-specified range. The prototype presented is rated to deliver up to 10 kW of buffered power in a 400 VDC system, with a bill of materials cost of $3,500 USD. The system design is described in detail. A key advantage of this architecture versus other approaches is that it reliably buffers power fluctuations without requiring any changes to system software, making it compatible with any training workload.

eess.SY