llama.cpp Vulkan on amd-radeon-rx-6650-xt: Llama-2-7B Q4_0 FA-off tg128 decode tok/s (scoreboard #10879)
62.14 tok/s (tg128, per-request, batch 1)
Fields
id
benchmarks/llamacpp-vulkan-rx-6650-xt-llama27b-q40-nofa-tg128Metric
decode_tok_s
Value
62.14
Unit
tok/s (tg128, per-request, batch 1)
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 table). llama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0. build dbb852b. tg128 is generation of 128 tokens, per-request batch 1, NOT an aggregate across concurrent requests. Re-read of the discussion page on 2026-10-16.
Measured by
third_party
Reproducible
yes
notes
Vulkan scoreboard row. no-FA commit dbb852b tg128 is the bandwidth-bound cell of the pair.
Notes
Vulkan scoreboard row. no-FA commit dbb852b tg128 is the bandwidth-bound cell of the pair.