Arc Pro B60 24GB Qwen3.5-9B Q4_K_XL token-gen, Vulkan backend (Level1Techs)
17.8 tok/s
Fields
id
benchmarks/comm-arc-pro-b60-qwen35-9b-q4-vulkan-tgMetric
decode_tok_s
Value
17.8
Unit
tok/s
Model
Qwen3.5-9B
Methodology
THIRD-PARTY MEASURED (Level1Techs, llama.cpp b8175). Q4_K_XL GGUF, VULKAN backend tg128 = 17.8 t/s, pp2048 = 559.6. Compare SYCL backend tg128 = 23.0 / pp2048 = 328.7 on the same card+model (comm-arc-pro-b60-qwen35-9b-q4-sycl-tg): SYCL wins TG, Vulkan wins PP.
Measured by
third_party
Reproducible
yes
notes
24GB GDDR6, ~456 GB/s - VRAM cheap (~$27/GB at $650 new) but ~1/3 the decode of a 3090 on the same model. Same tester's 35B-A3B Q8 SYCL run: 8.3 t/s (vs 133 t/s on an RTX PRO 4500).
Notes
24GB GDDR6, ~456 GB/s - VRAM cheap (~$27/GB at $650 new) but ~1/3 the decode of a 3090 on the same model. Same tester's 35B-A3B Q8 SYCL run: 8.3 t/s (vs 133 t/s on an RTX PRO 4500).