Benchmark shows 6x P40s outperform 6x MI50s running MiniMax M2.7 REAP 139B

A user benchmarked the MiniMax M2.7 REAP 139B Q3_K_L model on two multi-GPU setups: six AMD MI50 cards with 96GB total VRAM and six NVIDIA Tesla P40 cards with 144GB total VRAM.

• The P40 setup (CUDA) significantly outperformed the MI50 setup (ROCm) in prefill tests, ranging from ~1.13x to ~3.18x faster depending on context length.
• The MI50 setup was faster in token generation (tg128), achieving 24.87 t/s compared to the P40's 20.49 t/s.
• Both configurations ran on Asus X99-E-WS motherboards with Intel Xeon E5-2680 v4 CPUs and 128GB RAM.

The results indicate that despite the MI50's superior individual specs, the P40 cluster maintained competitive performance for this specific model configuration.