NVIDIA’s first Arm-based GB10 Grace Blackwell super chip recently appeared in Geekbench benchmark testing. This chip is also scheduled to officially debut at Computex 2025 in Taipei at the end of May 2025, potentially marking a significant step for NVIDIA in the workstation-grade Arm processor market. Combining NVIDIA’s technical expertise in artificial intelligence (AI) and accelerated computing, the GB10 can provide powerful computing power for AI developers and researchers.
The GB10 Grace Blackwell is a system-on-a-chip (SoC) that utilizes the NVIDIA Grace Blackwell architecture, integrating high-performance NVIDIA Grace CPUs and Blackwell GPUs. It achieves efficient data transfer through NVLink-C2C chip-to-chip interconnect technology. The chip includes 20 Arm architecture cores, with 10 high-performance Cortex-X925 cores and another 10 Cortex-A725 cores, with a maximum frequency of up to 3.9 GHz. Geekbench test results show that the GB10’s single-core performance can compete with high-end Arm and x86 processors, demonstrating excellent performance, but it still lags slightly behind Apple’s M4 Max in comparison. The test data also indicates that the GB10 exhibits balanced multi-core performance, suitable for handling AI model training, inference, and data science tasks.
The GB10 is equipped with 128GB of unified and coherent LPDDR5X memory with a memory bandwidth of up to 273GB/s, and it also supports up to 4TB of NVMe storage. This design allows the GB10 to locally run large language models (LLMs) with up to 200 billion parameters. By connecting two devices via NVIDIA ConnectX network, it can even support models with 405 billion parameters. This capability makes it an ideal platform for AI prototyping, model fine-tuning, and inference. Compared to the traditional PCIe interface, NVLink-C2C technology can reduce data transfer bottlenecks, providing a smoother experience for high-performance computing tasks.
NVIDIA’s collaboration with MediaTek is a major highlight in the development of the GB10. MediaTek’s experience in Arm-based SoC design brings excellent power efficiency and connectivity performance to the GB10, allowing it to run on a standard power outlet, making it suitable for desktop environments. The chip is manufactured using TSMC’s customized 4NP process and integrates the latest generation CUDA cores and fifth-generation Tensor cores, providing up to 1 petaFLOPS (FP4 precision) of AI performance. This performance allows it to excel in running complex AI workloads while maintaining low power consumption.
The application scenarios for the GB10 are beginning to emerge. NVIDIA is integrating it into the Project DIGITS AI supercomputer, a desktop device priced at approximately $2999 that debuted at CES 2025 and is scheduled to be available in May. Project DIGITS runs the Linux-based NVIDIA DGX operating system and supports the NVIDIA AI software stack, including the NeMo framework, RAPIDS libraries, and common tools such as PyTorch and Jupyter Notebook, making it convenient for developers to perform AI model development and testing. In addition, ASUS and Dell have announced that they will be using the GB10 in their upcoming workstation products, further expanding its application in the professional market.
It is worth noting that in the Geekbench test, the GB10 was identified as having an Armv8 architecture rather than its actual Armv9 architecture, possibly reflecting an issue with the testing software’s recognition. The test was conducted in a Windows environment rather than Linux, which deviates from the expectation that GB10 is primarily aimed at the Linux platform. Considering the Linux optimization of the NVIDIA DGX ecosystem, Linux remains the primary operating environment for GB10, especially in AI and high-performance computing scenarios.
At Computex 2025, NVIDIA may also showcase the N1 and N1X chips based on the GB10 architecture, targeting the desktop and laptop markets, respectively. These two chips are expected to combine Blackwell GPUs and Arm Cortex-X5 cores, support LPDDR6 memory, and be manufactured using TSMC’s 3nm process. It is reported that the N1 series will offer 150-200 TOPS of AI performance, aiming to challenge existing players in the Windows on Arm market, such as Qualcomm’s Snapdragon X series. MediaTek CEO Dr. Rick Tsai will deliver a keynote speech at Computex, which may further reveal details of the collaboration with NVIDIA.
The launch of the GB10 reflects NVIDIA’s strategic layout in accelerated computing and AI. Compared to traditional x86 processors, the GB10 optimizes the performance of data-intensive tasks through a unified CPU-GPU memory architecture and NVLink technology. Although its CPU performance may not match that of AMD Epyc or Intel Xeon, its design philosophy lies in integrating computing resources to provide efficient AI and data processing capabilities. This strategy is highly aligned with current AI-driven industry trends, such as generative AI, data analysis, and scientific computing.
In the current industry context, the release of the GB10 comes at a time when the Arm architecture is rapidly rising in the high-performance computing field. Arm’s high energy efficiency and flexibility make it a popular choice for data centers and edge computing. NVIDIA, through the combination of Grace CPU and Blackwell GPU, further promotes the application of Arm in workstations and AI development. At the same time, TSMC’s advanced process technology provides manufacturing support for the GB10, ensuring its competitiveness in performance and energy efficiency.
In the future, the GB10 and its derivatives may reshape the workstation and AI development market landscape. Its modular design and powerful software ecosystem provide developers with a seamless experience from local prototyping to cloud deployment. NVIDIA plans to further expand the application of GB10 through the NVIDIA DGX Cloud and AI Enterprise software platforms, supporting enterprise users in deploying solutions with the same architecture in larger-scale AI infrastructure.
The NVIDIA GB10 Grace Blackwell super chip, with its high performance, low power consumption, and flexible application scenarios, demonstrates the potential of the Arm architecture in the high-performance computing field. As Computex 2025 approaches, more details about the GB10 and its related products will be gradually revealed, bringing new expectations to AI developers, data scientists, and technology enthusiasts.