EasyRider: Mitigating Power Transients in Datacenter-Scale Training Workloads

Overview

This document is a preprint by Dillon Jensen and colleagues, published on arXiv in 2026 (arXiv:2604.15522) from Stanford University. It presents EasyRider, a rack-level power-delivery architecture for high-voltage DC AI racks that conditions training power before it reaches upstream distribution.

The problem is framed as follows: for an NVIDIA H100, communication phases cause a near-instantaneous 80% power reduction, and a 50,000-GPU job might draw 35 MW when computing and drop to 7 MW during communication. High-ramp gas turbines can change output by about 10-20 MW per minute, needing several minutes to go from idle to maximum. The speed of an AI training load can thus exceed the ramp capability of the grid side.

EasyRider combines three elements: a passive LC filter for high-frequency swings, an actively controlled rack battery and DC regulator for low-frequency swings and ramp limiting, and a software controller that holds battery state of charge mid-band. A 10 kW / 400 VDC prototype tested with normalised cluster traces and real GPU training shows rack swings reaching the grid as gradual changes within grid ramp and frequency limits, with no change to training code.

In the Titan X test EasyRider used 19% less energy than the software-burn approach. On a measured H100 trace the cells cycle at only about 0.05C, predicting roughly 60 years to 80% capacity; deeper swings raise this toward 0.26C and about a decade. The prototype bill of materials was about $3,500 ($0.35/W). Limit: it is a 10 kW prototype; large-scale field application and the ageing model are not validated.

Key points

1. For an NVIDIA H100, communication phases cause a near-instantaneous 80% power reduction; a 50,000-GPU job might draw 35 MW when computing and drop to 7 MW during communication.
2. High-ramp gas turbines can change output by about 10-20 MW/min, needing several minutes to go from idle to maximum.
3. In the Titan X test EasyRider used 19% less energy than the software-burn approach and had a lower peak power.
4. On a measured H100 trace the cells cycle at only about 0.05C, predicting roughly 60 years to 80% capacity; deeper swings raise this toward 0.26C and about a decade; the prototype battery uses less than 1% of its capacity.
5. The 10 kW prototype bill of materials was about $3,500 ($0.35/W), which scaled to a GB200 rack is under 1.25% of an estimated $3.7 million rack cost.
6. Without software control SoC drifted to approximately 62% after a few hours; the controller returned it to mid-band within approximately 20 minutes.
7. Combines a passive LC filter, an active rack battery and DC regulator and a software controller; an arXiv preprint.

Using it in design

• Offer rack-level smoothing storage (filter + battery/supercap + DC regulator) as a catalog item for HVDC AI racks, sized by peak power swing and grid ramp limit rather than backup energy.
• In the electrical model, place it behind the rack PDU so upstream UPS and transformers see a smoothed load.
• Compare it against GPU power-floor/burn options using an energy-overhead calculation.
• Include battery chemistry (LFP), C-rate and SoC management as parameters for lifetime estimation.

Who should read it

Engineers designing the power architecture of high-density AI training facilities and those planning the grid interface.

Publication details

Authors

Dillon Jensen; Grant Wilkins; Obi Nnorom; Hugo Budd; Ram Rajagopal; Juan Rivas-Davila; Phil Levis

Publisher

arXiv preprint (Stanford University)

Year

2026

Document number

arXiv:2604.15522

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2604.15522

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation