Artificial intelligence / news brief
Epoch AI proposes an inference index beyond peak FLOPS
- Event
- Published

Epoch AI published the Inference Performance Index, a model of concurrent requests per chip that combines compute, memory and interconnect limits across four agentic workloads. It estimates a GB300 has about 25 times an H100's per-chip inference capacity despite 2.5 times the 8-bit FLOPS. The result is analytical, workload-dependent and only partly checked against measured throughput; version 1 had already appeared in Epoch's data methodology before today's report.
Why it matters
The index shifts chip comparisons from peak arithmetic toward users served at target speeds, but its rankings depend on modeled workloads and quality assumptions.
Sources
- Inference Performance Index: changing the way we measure inference capacity (external site)Epoch AI · primary source ·
- Epoch AI's new inference index shows GB300 NVL72 far outpaces the H100 (external site)Crypto Briefing · coverage ·