v1.2.0

What's Changed

Features

• JSON output support (#22)
• GPU selection — run only a given list of GPUs (#19)
• Handle GPUs reporting 0 FLOPS — NVML wrongly reports them as healthy (#11)

Bug Fixes

• Compute real standard deviation in output instead of mean (#21)
• Don't hardcode NVML lib path — let the NVML library search for it (#18)

Other

• Upgrade base image to CUDA 13.1.0 (#15)
• Add GB200 benchmark reference (#20)
• Add V100 & A100 benchmark references with new table format (#9)

Thanks

Thanks to all contributors for this release! @anisse @Arno500 @hunoutl @itechdima

v1.1.0

What's Changed

• [feat] Add Blackwell/Gracehopper compatibility + FP8 computations by @Hugoch in #8

Full Changelog: v1.0.0...v1.1.0

v1.0.0

What's Changed

• tflops_tolerance is based from the fastest GPU by @glutamatt in #1
• feat: Add BF16 burn if GPU supports it. by @Hugoch in #5

New Contributors

• @glutamatt made their first contribution in #1

Full Changelog: https://github.com/huggingface/gpu-fryer/commits/v1.0.0