Arc Pro B60 24GB Qwen3.5-35B-A3B Q8 token-gen, SYCL (Level1Techs)
8.3 tok/s
Fields
id
benchmarks/comm-arc-pro-b60-qwen35-35b-a3b-q8-sycl-tgMetric
decode_tok_s
Value
8.3
Unit
tok/s
Model
Qwen3.5-35B-A3B (MoE)
Methodology
THIRD-PARTY MEASURED (Level1Techs, llama.cpp b8175). Q8 GGUF, SYCL backend tg128 = 8.3 t/s. MoE - ~35B class model in 24GB is only chat-viable because the active set is small; the same tester's RTX PRO 4500 did 133 t/s on the same model.
Measured by
third_party
Reproducible
yes
notes
The 35B-A3B fit on 24GB is real but slow: a capacity win, not a speed win. For interactive 35B-class MoE the 3090-class cards are ~4x faster per dollar.
Notes
The 35B-A3B fit on 24GB is real but slow: a capacity win, not a speed win. For interactive 35B-class MoE the 3090-class cards are ~4x faster per dollar.