llama.cpp CUDA DGX Spark llama-2-7b Q4_0 pp512 noFA
3062.31 tok/s (pp512)
Fields
id
benchmarks/llamacpp-scoreboard-gb10-nofa-pp512Metric
prefill_tok_s
Value
3,062.31
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 #15013, maintainer-curated CUDA backend). llama-bench llama-2-7b.Q4_0.gguf (ggml-org, 3.56 GiB, 6.74B params), pp512 and tg128, n_gpu_layers 99, batch 1, full GPU offload, CUDA. Per-request batch-1, not an aggregate across concurrent requests. Flash attention off (-fa 0). Build 5acd455, reporter @ggerganov. Printed mean 3062.31. Re-read of the discussion page on 2026-10-17.
Measured by
third_party
Reproducible
yes
notes
DGX Spark, 128 GB / LPDDR5x, build 5acd455, reporter @ggerganov. prompt processing cell.
Notes
DGX Spark, 128 GB / LPDDR5x, build 5acd455, reporter @ggerganov. prompt processing cell.