RX 7900 XTX (GPU0) llama.cpp ROCm 7.1.1: Llama-2-7B Q4_0 FA-off tg128
118.82 tok/s (tg128, per-request, batch 1, one GPU)
Fields
id
benchmarks/rocm711-rx7900xtx-gpu0-llama27b-q40-nofa-tg128Metric
decode_tok_s
Value
118.82
Unit
tok/s (tg128, per-request, batch 1, one GPU)
Model
llama-2-7b Q4_0 (3.56 GiB, 6.74B)
Methodology
llama-bench, build 34ce48d97 (7356), ROCm backend on ROCm 7.1.1, Arch Linux 6.12.61-1-lts. The published table gives ngl 99, fa 0, test tg128, on device 0 of a 2x RX 7900 XTX gfx1100 box, 24 GB. Reported 118.82 +/- 0.05. Per-request batch-1 decode, not aggregate. Device 1 scored 130.67 +/- 0.24 on the identical test.
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
Well below the maintainer scoreboard row for this card: discussion #15021 recorded 167.11 tok/s on the same model, quant, and tg128 test with build 2f0c2db (llamacpp-rocm-7900xtx-llama27b-q40-nofa-tg128). Kept as a second measurement rather than a correction, because the build, the ROCm version, and the clocks all differ, and this repo's own two cards already disagree by about 10 percent.
Notes
Well below the maintainer scoreboard row for this card: discussion #15021 recorded 167.11 tok/s on the same model, quant, and tg128 test with build 2f0c2db (llamacpp-rocm-7900xtx-llama27b-q40-nofa-tg128). Kept as a second measurement rather than a correction, because the build, the ROCm version, and the clocks all differ, and this repo's own two cards already disagree by about 10 percent.