Changelog

Oct. 19, 2026

Added the Inference Performance Index: modeled long-context inference capacity per accelerator, relative to GB300. Version 1 combines decode and prefill requirements across four model architectures and a log-weighted 50–400 SOL tokens/s/request range. Results include A100 SXM4 40 GB. See Records for the field definition and the methodology for assumptions and code.

Oct. 23, 2024