NVIDIA unveiled its GB10 SoC architecture at Hot Chips 2025, a multi-die solution built on TSMC's 3nm process, integrating a scaled-down Blackwell GPU and a MediaTek-designed 20-core Arm CPU. Powering the DGX Spark workstation, GB10 offers 128GB of unified memory for fine-tuning models up to 70B parameters and delivers 1000 TFLOPS FP4 performance, all within a 140W TDP. This strategic release positions a high-performance, accessible AI development platform as a gateway to NVIDIA's DGX Cloud, broadening the reach of its CUDA ecosystem beyond traditional data centers.
NVIDIA's unveiling of the GB10 System-on-Chip (SoC) at Hot Chips 2025 details a significant strategic move to broaden its AI ecosystem. The GB10, which powers the new DGX Spark workstation, is a technically advanced multi-die solution, integrating a scaled-down Blackwell architecture GPU with a 20-core Arm CPU designed by MediaTek, all fabricated on TSMC's 3nm process. Key performance metrics include 1000 TFLOPS in FP4 precision and 128GB of unified LPDDR5X memory, enabling the fine-tuning of 70-billion parameter models within an efficient 140-watt power envelope suitable for a standard wall outlet. Strategically, NVIDIA is positioning the DGX Spark as a 'gateway device' to funnel developers from local testing into its high-margin DGX Cloud services. The successful execution of this complex chip, achieving functionality on its first A0 silicon spin without costly revisions, underscores NVIDIA's engineering prowess and effective collaboration with partners like MediaTek, who were responsible for the critical CPU chiplet and memory subsystem.
AI-powered research, real-time alerts, and portfolio analytics for institutional investors.
Request a DemoOverall Sentiment
strongly positive
Sentiment Score
0.75
Ticker Sentiment