llama.cpp CUDA on nvidia-rtx-4090-d: Llama-2-7B Q4_0 FA-on tg128 decode tok/s (scoreboard #15013)
191.57 tok/s (tg128, per-request, batch 1)
Fields
id
benchmarks/llamacpp-cuda-rtx-4090-d-llama27b-q40-fa-tg128Metric
decode_tok_s
Value
191.57
Unit
tok/s (tg128, per-request, batch 1)
Model
llama-2-7b.Q4_0.gguf (TheBloke, 3.56 GiB, 6.74B params)
Methodology
COMMUNITY-SCOREBOARD (llama.cpp discussion #15013, maintainer-curated). llama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 1. llama-bench default -r 5 repeats, mean reported (the +/- std is in the source row, not stored). Single GPU, per-request batch 1, NOT an aggregate across concurrent requests. Prompt length is the bench default for the tg test (the scored quantity is the 128 generated tokens). Re-read of the discussion page on 2026-10-16.
Measured by
third_party
Reproducible
yes
notes
RTX 4090 D, 24 GB / GDDR6X / 384 bit, build 79c1160, reporter @autonomous-ai-lab. China-only AD102-250 cut of the 4090; see nvidia-rtx-4090-d. Maintainer-verified row. tg128 is the bandwidth-bound cell of the pair.
Notes
RTX 4090 D, 24 GB / GDDR6X / 384 bit, build 79c1160, reporter @autonomous-ai-lab. China-only AD102-250 cut of the 4090; see nvidia-rtx-4090-d. Maintainer-verified row. tg128 is the bandwidth-bound cell of the pair.