NVIDIA AI Updates: September 28, 2026

1. DSX MaxLPS Fit 37 Percent More GPUs Into the Same Power Budget in an Nscale Test

NVIDIA. NVIDIA detailed DSX MaxLPS, a dynamic power allocation system that monitors actual consumption and shifts GPU power limits across nodes so a data center can run more GPUs within a fixed provisioned power budget. In a joint evaluation with Nscale serving Kimi K2.5 on GB300 NVL72 systems, the same 264.4 kW budget supported 192 GPUs instead of 140, and aggregate throughput rose 49.2 percent to about 1.62 million tokens per second, or 6.12 tokens per second per provisioned watt versus 4.10. Median and P75 latency stayed within 5 percent of baseline, while P99 time to first token rose 17 percent. Source

2. Nemotron 3 Diarization Tops a Strict Speaker-Separation Benchmark at 100M Parameters

NVIDIA. NVIDIA released Nemotron 3 Diarization, an open-weight model of about 100 million parameters that labels which speaker is talking, distinguishes up to eight speakers, and detects overlapping speech on both recordings and live audio. The audio buffer is configurable from 30.4 seconds down to 0.32 seconds, with shorter buffers trading accuracy for latency, and pairing it with an ASR model such as Parakeet yields transcripts with anonymous speaker labels. It currently ranks first on VoiceArena’s Diarization-Bench at a 14.72 percent error rate versus 19.3 percent for the next system, and cuts error by an average of 41 percent over its predecessor Streaming Sortformer at a 1.04-second buffer. Source