Skip to content

README.md

Latest commit

 

History

History
19 lines (15 loc) · 1.06 KB

README.md

File metadata and controls

19 lines (15 loc) · 1.06 KB
Raw

eval — moved to PerfDigest-MCP-Bench

Benchmark workloads and evaluation results live in a dedicated companion repo: https://github.com/onlyxItachi/PerfDigest-MCP-Bench — keeping large report artifacts and result documents out of this code repo. Unit tests stay here (they gate CI); only benchmark/eval artifacts moved.

Headline results measured so far:

  • Token efficiency: the digest costs ~14–130x fewer tokens per turn than raw ncu output for the same correct diagnosis — results/RESULTS.md.
  • Cross-backend, real hardware: NVIDIA (nsight) and CPU (linux_perf) digests validated on a real RTX 4060 + Ryzen host through one MCP call shape — results/CROSS_BACKEND_2026-06-15.md.

To reproduce, see workloads/README.md in the bench repo (capture commands for both workloads).