lhl/ggerganov: NVIDIA DGX Spark GB10 CUDA CUDA gpt-oss-120b tg32 38.76 t/s @32768
38.76 tok/s per-request decode (batch 1, depth 32768)
Fields
id
benchmarks/lhl-dgx-spark-gpt-oss-120b-cuda-tg32-d32kMetric
decode_tok_s
Value
38.76
Unit
tok/s per-request decode (batch 1, depth 32768)
Model
gpt-oss-120b MXFP4/Q8 GGUF
Methodology
THIRD-PARTY MEASURED (lhl/strix-halo-testing README). llama.cpp build 6792; gpt-oss-120b (ggml-org MXFP4/Q8); pp2048/tg32 vs depth 0-32768. Spark numbers from ggerganov's llama.cpp discussion #16578 sweep (b6767); Strix Halo numbers lhl on Framework Desktop, TheRock/ROCm 7.10.0a20251017. Per-request batch-1. NVIDIA DGX Spark GB10 CUDA
Measured by
third_party
Reproducible
yes
notes
Spark tg degrades far less with depth than either STXH backend (lhl's own conclusion).
Notes
Spark tg degrades far less with depth than either STXH backend (lhl's own conclusion).