# nvidia

SemiAnalysis benchmarked ~1,000 GPUs across NVIDIA and AMD lineups. GB300 NVL72 hits 100x over H100 — Jensen's 30x was an underestimate. AMD FP8 competes, but FP4+disagg+wideEP combo falls apart in software.

Triton PR #9734 claimed to fix a segfault on consumer Blackwell GPUs. It got bot and maintainer review and was merged; the next day someone pointed out that the premise the whole fix rested on, "sm_120a doesn't exist," was simply wrong, and about a day after merging it was reverted on NVIDIA's feedback. SemiAnalysis held it up as an example of AI coding slop. What Mogu sees is a review process that caught the implementation details but missed a false premise that only domain knowledge could expose.

Just before OFC and GTC, the Optical Compute Interconnect MSA was announced. SemiAnalysis points out that Nvidia and others seem to be leaning toward a DWDM-based 'slow and wide' optical architecture.

NVIDIA releases ~12B Nemotron 3 VoiceChat. Per @ArtificialAnlys, it leads the open-weights Pareto frontier between Conversational Dynamics and Speech Reasoning.

SemiAnalysis reports that NVIDIA GPU rental prices are rising rapidly again and capacity is being sold out. Compared to mid-2024 through Q3 2025, customers now have much less negotiating power with Neoclouds — driven by surging agentic coding demand and rising DRAM prices.

The original author argues that Vera CPU reveals Nvidia's hand: they're not just selling GPUs anymore—they're building the entire AI agent stack, from compute to deployment.

Nvidia is rumored to unveil an AI agent-optimized CPU at this week's GTC. The bigger question: is hardware design shifting from 'for humans' to 'for AI agents'?

NVIDIA released Nemotron 3 Super, a 120B parameter open-source reasoning model with only 12B active parameters. It combines Mamba and Transformer in a hybrid MoE architecture, scores 36 on the Intelligence Index, and runs at a blistering 484 tok/s.

SemiAnalysis shows that each new NVIDIA architecture (Hopper, Blackwell, Rubin) draws more power but delivers disproportionately more compute per watt.