Decode · tok/s
40
Prefill
—
tok/100W
16.7
DGX Spark x8 Cluster : nvidia/GLM-5.3-NVFP4. C1 : 40 tok/s. C2 : 41 tok/s. C4 : 52 tok/s. vLLM: Eugr’s spark-vllm-docker.
First-person 8× Spark. Rank is the quoted C1 40. C2 and C4 stay in the note as concurrent. Separate from the 2× GB300 SGLang 81.5.
NVFP4