Chroma — leaderboard

Grid Guard smooths GPU power spikes in AI data centers by 80% via workload staggering

Empromptu AI has introduced Grid Guard, a software application aimed at reducing GPU-driven power volatility in AI data centers by predicting and smoothing rapid load swings before they hit facility power infrastructure.

The problem Grid Guard targets is the step-change behavior that can happen when thousands of GPUs execute synchronized operations—starting training batches, synchronizing across a cluster, or processing large inputs. Empromptu describes these events as causing power demand to swing by tens of megawatts in milliseconds, a rate of change that can stress turbines and generators and push operators toward expensive overbuilds.

Grid Guard’s core control lever is time: it staggers GPU workload execution by 50 to 200 ms so sharp spikes become ramps. In early deployments, the company reports an average 80% reduction in power volatility, which it says comes without meaningful impact on AI workload performance.

Empromptu groups Grid Guard’s functionality into four areas. It predicts near-term workload-driven power swings, smooths swings by phasing and scheduling workloads (including staggering batch starts, breaking large inputs into smaller processing steps, and reusing cached compute results), and sends forward demand signals to batteries, generators, and “smart power systems” so equipment can prepare. It also “prices” volatility by converting load behavior into an economic signal intended to make the cost of inefficient load patterns visible.

Software-level load shaping is an appealing knob because it can reduce the size of the problem upstream, rather than relying entirely on batteries and generators to absorb step loads after the fact. But operators will still want to validate how any scheduler-level staggering interacts with cluster utilization targets and job completion times, especially at scale where small delays can compound across pipelines.

“Grid Guard is how we make that buildout sustainable, by making the software smarter about when and how workloads fire,” said Shanea Leven, CEO and co-founder of Empromptu.

Grid Guard is Empromptu’s first data center operations application and is already engaged with a major power company ahead of a live deployment. Empromptu says Grid Guard is available now.

Source: Empromptu AI

Chroma — below posts

Get Data Center Engineering News In Your Inbox:

By subscribing, you agree to our Privacy Policy and Terms of Use. You can unsubscribe at any time.

Chroma — tower

Popular Posts:

Johnson-Controls-has-launched-an-Absorption-Chiller-Reference-Design-Guide-aimed-at-data-centers-running-on-on-site-power-generation
Johnson Controls' new absorption chiller guide targets 44% lower cooling power in AI data centers
trm-DC4-landing-page-banner
CPC says its full-flow AI cooling connectors cut pressure drop up to 90%
39D1CAB6-7D1C-4CA7-A2CF-D0919F6EA523_1_105_c
EdgeSites uses existing buildings to deploy 1 MW immersion-cooled AI compute
Why data center power validation is moving to real-time simulation
Why data center power validation is moving to real-time simulation
image-schneider-electric-and-amd-release-first-helios-platfo
AMD Helios reference design targets 246 kW racks, 10.4 MW AI clusters

Share Your Data Center Engineering News

Do you have a new product announcement, webinar, whitepaper, or article topic? 

TDK — tower ()

Get Data Center Engineering News In Your Inbox:

By subscribing, you agree to our Privacy Policy and Terms of Use. You can unsubscribe at any time.

TDK — leaderboard ()