llama.cpp Vulkan Intel Arc A750 llama-2-7b Q4_0 pp512 FA
303.37 tok/s (pp512)
Fields
id
benchmarks/llamacpp-vulkan-arc-a750-llama27b-q40-fa-pp512Metric
prefill_tok_s
Value
303.37
Unit
tok/s (pp512)
Model
llama-2-7b.Q4_0.gguf (ggml-org, 3.56 GiB, 6.74B params)
Methodology
COMMUNITY-SCOREBOARD (llama.cpp discussion #10879, maintainer-curated Vulkan backend). llama-bench llama-2-7b.Q4_0.gguf (ggml-org, 3.56 GiB, 6.74B params), pp512 and tg128, batch 1, full GPU offload, Vulkan. Per-request batch-1, not aggregate across concurrent requests. Flash attention on (-fa 1). Build c1b1876. Chip Intel Arc A750. Printed mean 303.37 +/- 1.44.
Measured by
third_party
Reproducible
yes
notes
Flash-attention scoreboard row. Prompt processing collapses versus noFA on this chip (303.37 vs 1075.94), so the FA path is not the one to quote for prefill.
Notes
Flash-attention scoreboard row. Prompt processing collapses versus noFA on this chip (303.37 vs 1075.94), so the FA path is not the one to quote for prefill.