llama.cpp CUDA on Tesla V100 32GB SXM2: Llama-2-7B Q4_0 FA-on tg128 decode tok/s (scoreboard #15013)
134.76 tok/s (tg128, per-request, batch 1)
Fields
id
benchmarks/llamacpp-cuda-v100-sxm2-32gb-llama27b-q40-fa-tg128Metric
decode_tok_s
Value
134.76
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. Tesla V100 32GB SXM2 (32 GB HBM2 4096-bit). Build commit 51f5a45, reporter @Hedede. llama-bench default -r 5 repeats, mean +/- std. Single-GPU run per thread rules. Per-request batch-1, NOT aggregate.
Measured by
third_party
Reproducible
yes
notes
Maintainer-verified scoreboard row from ggml-org/llama.cpp#15013. Identical model/test/flags across every GPU, making this the cleanest cross-GPU llama.cpp comparison in the repo. tg128 is bandwidth-bound and tracks memory bandwidth; pp512 tracks compute. Flash-attention enabled.
Notes
Maintainer-verified scoreboard row from ggml-org/llama.cpp#15013. Identical model/test/flags across every GPU, making this the cleanest cross-GPU llama.cpp comparison in the repo. tg128 is bandwidth-bound and tracks memory bandwidth; pp512 tracks compute. Flash-attention enabled.