NVIDIA GB10 Grace Blackwell (DGX Spark) — Benchmarks & Specs

Bottom line: how fast is the NVIDIA GB10 Grace Blackwell (DGX Spark)?

For local LLM inference it generates 60.6 tokens/sec running gpt-oss:120b at MXFP4 under llama.cpp (the source did not report the machine's memory configuration; 59 GB peak use), per llama.cpp GitHub discussion #16578. Unified memory is the binding constraint for local inference: the NVIDIA GB10 Grace Blackwell (DGX Spark) ships in configurations of up to 128 GB, per the manufacturer spec sheet, and a 120B model at MXFP4 needs roughly 73 GB for its weights alone (120B parameters at ~4.85 bits per weight, per llama.cpp), so it runs only on configurations with more unified memory than that. In Geekbench 6 multi it scores 18,183 points, per daniel.lawrence.lu.

Every figure above is a row in the tables below, and each row links out to the review or public benchmark database the number was taken from; the memory ceiling and the weight-size estimate are not table rows and link to their sources inline. SpecPicks aggregates published measurements; it does not report first-party benchmark runs.