llama.cpp CUDA on nvidia-rtx-5090-laptop: Llama-2-7B Q4_0 FA-on pp512 prefill tok/s (scoreboard #15013)
7641.89 tok/s (pp512, batch 512)
Fields
id
benchmarks/llamacpp-cuda-rtx-5090-laptop-llama27b-q40-fa-pp512Metric
prefill_tok_s
Value
7,641.89
Unit
tok/s (pp512, batch 512)
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, prompt length 512 tokens, no generation. llama-bench default -r 5 repeats, mean reported. Single GPU, one request, NOT an aggregate across concurrent requests. pp512 tracks compute rather than bandwidth. Re-read of the discussion page on 2026-10-16.
Measured by
third_party
Reproducible
yes
notes
RTX 5090 Laptop, build 29f1482, reporter @rs38. Same bus-width caveat as the tg128 row: the scoreboard says 384-bit, the spec says 256-bit. Prefill cell of the same row as the tg128 record for this card.
Notes
RTX 5090 Laptop, build 29f1482, reporter @rs38. Same bus-width caveat as the tg128 row: the scoreboard says 384-bit, the spec says 256-bit. Prefill cell of the same row as the tg128 record for this card.