Dual GX10 Benchmark

Ran MiniMax-M2.7-AWQ-4bit without issues on a dual Asus GX10 (dgx spark) setup.

• Power consumption per GX10 during inference is around about 100W
• Works without errors in open code and hermes agent
• Time to first token feels long, but the low-noise, low-heat setup is an advantage

Performance based on llama benchy --depth 0 4096 8192 16384 32768 --latency-mode generation is as follows.

• pp2048: 3452.05 t/s, ttfr 626.82 ms
• tg32: 38.84 t/s
• pp2048 @ d4096: 2848.85 t/s, ttfr 2022.61 ms
• pp2048 @ d8192: 2579.85 t/s, ttfr 3523.69 ms
• pp2048 @ d16384: 2411.34 t/s, ttfr 6791.62 ms
• pp2048 @ d32768: 1988.05 t/s, ttfr 15512.61 ms

As context length increases, prompt processing speed drops, but generation speed stayed in the range of about 37-39 t/s.