One GPU, More Work.

We evaluate Nodus Compute’s resident, software-defined GPU partitioning against native NVIDIA Multi-Process Service (MPS) on the same NVIDIA A100-SXM4-40GB. Using Qwen3-1.7B-FP8 and an identical deterministic two-tenant trace, Nodus delivered 15.0% and 15.1% more completed output throughput at 12 and 16 requests per second, while reducing p99 time to first token by 15.0% and 16.6%. The report includes the matched methodology, results, system interpretation, and reproducibility statement. Machine-readable evidence is available at https://research.nodus-compute.ai/gpu-partitioning.