Decode · tok/s
60
Prefill
—
tok/100W
6
@coldniko Strata 0.1.43.1 on 4x V100 32GB just hit 204 tok/s aggregate with 8-way concurrent (Qwen3.8-Flash-Next Q4_K_XL-DN4). Single 60 tok/s. Beating my RTX 5090 in total throughput. Volta still has headroom - any tips to optimize further? Full bench: https://github.com/Niko1221/Strata/issues/2260
First-person 4× V100 32GB. Rank is the quoted single 60 tok/s. The 204 is concurrent aggregate and stays a note. Separate from the single V100 57 pair.
