Characterizing 2nd Law Efficiency of AI Datacenters

Datacenter workload fluctuations challenge the design and operation of our critical grid infrastructure. Significant gaps exist in assessment of these fluctuations as to their lifetime and short-term impact on the infrastructure and environment. Some mitigation approaches focus on throttling the datacenter workload, thus impacting the performance. Other approaches include use of alternate generation sources or energy storage systems. While disparate, these approaches are reactive and lack foresight and the intelligence to plan and strategize for optimal power management. We introduce an approach to quantify the transitional entropy generated during datacenter power fluctuations as a metric to evaluate datacenter performance using power demand measurements. A comparative assessment is provided between a BESS optimized datacenter and a regular datacenter to demonstrate the reduction of irreversibilities due to power fluctuations. Workloads are used to characterize the datacenter power demand at the point of interaction with utility. This approach can be scaled from datacenters to servers to chips.
Speaker(s): Ratnesh K Sharma,
Agenda:
No-host social at 5:30pm
Presentation at 6:00pm
Dinner at 7:00pm
Presentation continues at 7:45pm
Adjourn by 8:30pm
Zio Fraedo's, 611 Gregory Lane, Pleasant Hill, California, United States, 94523