llama.cpp Vulkan on nvidia-gb10: Llama-2-7B Q4_0 FA-on pp512 prefill tok/s (scoreboard #10879)
3321.91 tok/s (pp512, batch 512)
Fields
id
benchmarks/llamacpp-vulkan-gb10-llama27b-q40-fa-pp512Metric
prefill_tok_s
Value
3,321.91
Unit
tok/s (pp512, batch 512)
Model
llama-2-7b.Q4_0.gguf (ggml-org, 3.56 GiB, 6.74B params)
Methodology
COMMUNITY-SCOREBOARD (llama.cpp discussion #10879, Vulkan backend, maintainer-curated "Llama 2 7B, Q4_0, FA enabled" table). llama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 1. build 657e011. pp512 is prompt processing of 512 tokens with no generation, one request, NOT an aggregate. Re-read of the discussion page on 2026-10-17.
Measured by
third_party
Reproducible
yes
notes
Vulkan scoreboard row. pp512 tracks compute more than bandwidth. FA-on row; the no-FA sibling for this chip, where one exists, is a separate record.
Notes
Vulkan scoreboard row. pp512 tracks compute more than bandwidth. FA-on row; the no-FA sibling for this chip, where one exists, is a separate record.