Samuel K. Moore

Research from the College of William & Mary, Jefferson Lab, and Silicon Data reveals significant performance variability among GPUs of the same model when rented from cloud providers. Testing 6,800 benchmark instances across 3,500 GPUs from 11 cloud operators found that H100 PCIe GPUs varied by up to 34.5 percent in computing performance and H200 SXM GPUs by up to 38 percent in memory bandwidth, despite being identical models. The variability stems from manufacturing inconsistencies rather than cooling or configuration differences, creating a real financial risk for customers paying premium prices for GPUs that may underperform older models.