ggml benches/dgx-spark table, build 11fb327bf (7941)

Fields

id

sources/ggml-dgx-spark-bench-build-7941

URL

https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md

Publisher

ggml-org (ggerganov, discussion 16578)

Kind

benchmark

Published

2026-02-05

Accessed

2026-10-16

Archived URL

https://raw.githubusercontent.com/ggml-org/llama.cpp/master/benches/dgx-spark/dgx-spark.md

notes

Maintainer llama-bench table for one NVIDIA DGX Spark GB10. aarch64, driver 580.95.05, CUDA 13.0, nvidia-smi dated 2026-02-05. Command from discussion 16578: llama-bench -m model -fa 1 -d 0,4096,8192,16384,32768 -p 2048 -n 32 -ub 2048 --no-mmap. Build footer 11fb327bf (7941). pp2048 is prefill of 2048 tokens at batch 1. tg32 is batch-1 decode of 32 tokens. Depth rows are the same test after N tokens of prior context. Batched-bench tables on the same page are aggregate and are not filed. Initial post was 2025-10-14 commit 7ea15bb; this file is the 2026-02-05 refresh. Distinct from the earlier lhl Spark gpt-oss-120b CUDA numbers, which are a different build.

Notes

Maintainer llama-bench table for one NVIDIA DGX Spark GB10. aarch64, driver 580.95.05, CUDA 13.0, nvidia-smi dated 2026-02-05. Command from discussion 16578: llama-bench -m model -fa 1 -d 0,4096,8192,16384,32768 -p 2048 -n 32 -ub 2048 --no-mmap. Build footer 11fb327bf (7941). pp2048 is prefill of 2048 tokens at batch 1. tg32 is batch-1 decode of 32 tokens. Depth rows are the same test after N tokens of prior context. Batched-bench tables on the same page are aggregate and are not filed. Initial post was 2025-10-14 commit 7ea15bb; this file is the 2026-02-05 refresh. Distinct from the earlier lhl Spark gpt-oss-120b CUDA numbers, which are a different build.

Cross-references

This record references no other records.