boltbeam.fyi

Qwen3.8-27B speed on every GPU

15 measured runs across 9 GPUs, fastest first.

GPUQuantEnginetok/sRooflineShare
H200W4A16-AWQvLLM100.8299.134%
2× Radeon R9700Quark W4A16vLLM + MTP k=292.3––
H100 80GBW4A16-AWQvLLM87.8211.941%
RTX 5090W4A16-AWQvLLM87.5109.780%
H100 80GBUD-Q4_K_Mllama.cpp83.6205.741%
RTX 5090UD-Q4_K_Mllama.cpp81.9108.775%
H200 NVLUD-Q4_K_Mllama.cpp78.6286.727%
2× RTX 3090W4A16-AWQvLLM68.690.676%
2× Radeon R9700Quark W4A16vLLM56.1––
RTX 4090W4A16-AWQvLLM54.461.788%
RTX 4090UD-Q4_K_Mllama.cpp50.461.182%
RTX 3090W4A16-AWQvLLM46.257.680%
L40SW4A16-AWQvLLM43.648.989%
RTX 3090UD-Q4_K_Mllama.cpp41.557.073%
L40SUD-Q4_K_Mllama.cpp40.048.483%

Download