Nvidia used the Hot Chips 2025 stage to pull the curtain back on GB10, a new system-on-chip that shrinks its Blackwell GPU architecture into a compact, power-conscious form. The pitch is straightforward and ambitious: bring Blackwell-class AI capabilities to devices and deployments where power, size, and thermals matter far more than in the data center.
Built in collaboration with MediaTek, GB10 combines Nvidia’s graphics and AI expertise with a CPU complex of 20 Arm v9.2 cores. That pairing signals a design focused on balanced performance across AI inference, general computing, and multimedia workloads—ideal for everything from edge servers and smart cameras to robotics, automotive systems, and next-gen consumer devices.
Positioning GB10 as a miniaturized Blackwell makes strategic sense. Developers get familiar tools and model compatibility while hardware makers gain a tighter power envelope. In practical terms, that could mean faster on-device AI, lower latency for perception and language models, and reduced reliance on the cloud for heavy lifting.
The Arm v9.2 foundation is noteworthy in its own right. The latest Arm instruction set generation is built for modern performance and security demands, which aligns well with AI at the edge, mixed-criticality automotive applications, and privacy-sensitive workloads that benefit from on-device processing.
While Nvidia hasn’t shared a full spec sheet, the trajectory is clear. GB10 looks aimed at unifying GPU-accelerated AI with a robust CPU cluster in a single SoC, leveraging the Blackwell software ecosystem. Expect strong ties to established Nvidia toolchains and frameworks, helping teams move models across product tiers without a wholesale rewrite.
For OEMs, the appeal is the promise of Blackwell-derived performance without data center thermals or costs. For developers, it’s the potential to run more capable models locally, scale deployment footprints, and shorten iteration cycles. For end users, it could translate into smarter, faster, and more responsive devices that work even when the network doesn’t.
Key questions that will shape adoption include power targets, thermal design ranges, memory bandwidth, and real-world inference throughput. Availability timelines, partner designs, and standardized benchmarks will also be critical to watch as the platform moves from conference highlight to shipping silicon.
If Nvidia delivers on the vision, GB10 could become a cornerstone for edge AI systems, bridging the gap between Blackwell’s headline-grabbing performance and the everyday products that need efficient, reliable intelligence on the go.
