ggml CUDA on DGX Spark GB10: Qwen3-Coder-30B-A3B Q8_0 pp2048 d32768 (b7941)
1348.17 tok/s (pp2048, batch 1, per-request)
Fields
id
benchmarks/ggml-cuda-gb10-spark-qwen3-coder-30b-a3b-pp2048-d32768-b7941Metric
prefill_tok_s
Value
1,348.17
Unit
tok/s (pp2048, batch 1, per-request)
Model
Qwen3-Coder-30B-A3B-Instruct Q8_0 (ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF)
Methodology
MAINTAINER bench on one NVIDIA DGX Spark (GB10), build 11fb327bf (b7941). CUDA backend, ngl 99, flash attention on, ubatch 2048, mmap off, direct-io on. pp2048 is prefill of a 2048-token prompt, batch 1, per-request, NOT aggregate. tg32 is batch-1 decode of 32 new tokens, per-request. Depth d=32768 is the same measurement after 32768 tokens of prior context (plain row is depth 0). Value is the published mean; stddev is in notes. Single device, unified memory, no tensor parallel.
Measured by
third_party
Reproducible
yes
notes
Official ggml DGX Spark table, not a forum paste. stddev +/- 2.0. File size 30.25 GiB, params 30.53 B. A concurrent sweep at batch 1 through 32 is also on the page; those figures are aggregate and are not this record. Distinct from the earlier third-party Spark gpt-oss-120b CUDA sweep (different build, lower tok/s).
Notes
Official ggml DGX Spark table, not a forum paste. stddev +/- 2.0. File size 30.25 GiB, params 30.53 B. A concurrent sweep at batch 1 through 32 is also on the page; those figures are aggregate and are not this record. Distinct from the earlier third-party Spark gpt-oss-120b CUDA sweep (different build, lower tok/s).