NVIDIA's first GB10 Grace Blackwell superchip, leveraging the Arm architecture, recently appeared in Geekbench benchmarks. Anticipated to be officially presented at Computex 2025 in Taipei by the end of May 2025, this launch could mark a significant stride for NVIDIA in the workstation-class Arm processor domain. With NVIDIA's deep expertise in artificial intelligence (AI) and accelerated computing, the GB10 is poised to deliver substantial computational power to AI developers and researchers.

The GB10 Grace Blackwell is a system-on-chip (SoC) that incorporates the high-performance NVIDIA Grace CPU and Blackwell GPU, achieving efficient data transfers through NVLink-C2C interconnect technology. It features 20 Arm architecture cores, including 10 high-performance Cortex-X925 cores and 10 Cortex-A725 cores, with a maximum frequency of 3.9 GHz. Geekbench results indicate that the GB10 rivals high-end Arm and x86 processors in single-core performance, though it slightly lags behind Apple's M4 Max. The chip demonstrates balanced multi-core performance, making it well-suited for AI model training, inference, and data science tasks.

Outfitted with 128GB of unified LPDDR5X memory and up to 1TB/s of memory bandwidth, the GB10 supports up to 4TB of NVMe storage. This configuration enables the chip to handle large language models (LLMs) with up to 200 billion parameters locally, and 405 billion parameters with dual-device connectivity via NVIDIA ConnectX network. This capability positions it as an ideal platform for AI prototyping, model fine-tuning, and inference. NVLink-C2C technology overcomes traditional PCIe bottlenecks, enhancing high-performance computing tasks.

A highlight of the GB10’s development is the collaboration between NVIDIA and MediaTek, bringing together MediaTek's prowess in Arm-based SoC design for outstanding power efficiency and connectivity. Suitable for desktop environments, it operates with just a standard power outlet. Fabricated using TSMC's customized 4NP process, the chip integrates the latest CUDA cores and 5th generation Tensor cores, achieving AI performance up to 100 trillion floating point operations (FP4 precision). This ensures excellence in complex AI workloads while maintaining energy efficiency.

Emerging application scenarios for GB10 include its integration into Project DIGITS AI supercomputer, a desktop unit priced around $2,999, introduced at the 2025 International CES and expected in May. Project DIGITS utilizes the Linux-based NVIDIA DGX operating system and supports the NVIDIA AI software suite, encompassing the NeMo framework, RAPIDS library, and popular tools such as PyTorch and Jupyter Notebook, simplifying AI model development and testing. Additionally, ASUS and Dell plan to integrate the GB10 into forthcoming workstation offerings, expanding its professional market reach.

It's noteworthy that Geekbench tests mistakenly label the GB10 as utilizing the Armv8 architecture, rather than the actual Armv9 architecture, possibly due to a software recognition issue. Testing occurred in a Windows environment, diverging from the expected Linux usage for the GB10, intended primarily for Linux platforms. With Linux optimizations foundational to the NVIDIA DGX ecosystem, Linux remains the primary environment for GB10, especially in AI and HPC contexts.

At Computex 2025, NVIDIA might also unveil the N1 and N1X chips, derived from the GB10 architecture, for desktops and laptops respectively. Featuring Blackwell GPUs and Arm Cortex-X5 cores, LPDDR6 memory support, and production on TSMC's 3nm process, the N1 series aims to deliver 150-200 TOPS of AI performance, challenging current Windows on Arm competitors like Qualcomm's Snapdragon X series. MediaTek CEO Dr. Jonathan Tsai is set to keynote at Computex, potentially divulging further details about the NVIDIA collaboration.

The GB10's introduction underscores NVIDIA's strategic focus on accelerated computing and AI, enhancing data-intensive task performance through a unified CPU-GPU memory architecture and NVLink technology. While its CPU performance might not match AMD Epyc or Intel Xeon, it consolidates computing resources for efficient AI and data processing capabilities. This aligns closely with AI-driven industry trends, including generative AI, data analytics, and scientific computing.

As the Arm architecture rapidly gains ground in high-performance computing, noted for its energy efficiency and flexibility in data center and edge computing, NVIDIA is advancing Arm adoption in workstations and AI development by leveraging Grace CPUs with Blackwell GPUs. TSMC’s advanced manufacturing processes provide a competitive edge for GB10 in performance and energy efficiency.

Looking ahead, the GB10 and its derivatives may shift the landscape of workstations and AI development. Its modular design and mature software ecosystem afford developers seamless transitions from local prototyping to cloud deployment. NVIDIA intends to widen GB10's reach through NVIDIA DGX Cloud and AI Enterprise software platforms, aiding enterprise deployment across vast AI infrastructures.

With impressive performance, low power usage, and versatile applications, the NVIDIA GB10 Grace Blackwell superchip exemplifies the potential of the Arm architecture in high-performance computing. As Computex 2025 approaches, more details about GB10 and its associated products will emerge, promising exciting developments for AI developers, data scientists, and tech enthusiasts.