PEX8749 switch allows 6 MI50s with negligible performance loss

A user tested running six AMD MI50 GPUs on an Asus X99-E-WS motherboard by using a PEX8749 PCIe switch to connect four cards, while the remaining two connected directly. The goal was to determine if this configuration improved or degraded token generation (tg) and prompt processing (pp) speeds compared to all six GPUs running directly.

• Testing the dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B model on an Intel Xeon E5-2680 v4 with 128GB RAM.
• The switch configuration resulted in a -0.74% change for pp512 and a -0.36% change for pp16384+tg128.
• Token generation speeds improved by +2.73% for tg128 and +0.83% for pp65536+tg128.
• The combined pp512+tg128 test showed a +1.90% improvement, while pp4096+tg128 remained nearly identical at +0.01%.

The results indicate that using the PEX8749 switch to expand GPU capacity introduces negligible performance loss, with token generation speeds actually seeing slight improvements across several benchmarks.