A GB10 Grace Blackwell Superchip in charge
Like every year, NVIDIA multiplies the announcements during CES in Las Vegas. We have already detailed the new GeForce RTX 50 which promise double performance compared to the equivalent RTX 40. These graphics cards use the new Blackwell architecture, like the NVIDIA Digits project in question today.
Digits takes the form of a small box in which we find an NVIDIA GB10 Grace Blackwell Superchip SoC. Grace is the name of the CPU part of the “Superchip”, while Blackwell corresponds to the GPU part. We find exactly the same segmentation in the GH200 and GB200 chips, for example, with H200 for the Hopper generation of GPUs and B200 for Blackwell.
Grace CPU part: 20 ARM cores
This GB10 is, as its name easily suggests, lighter compared to the GH200 and GB200. On the CPU part, NVIDIA announces only 20 ARM cores “ optimized for energy efficiency ”, without further details.
On GH200 and GB200, we find 72 ARM Neoverse v2 cores in the Grace CPU part. NVIDIA specifies that its chip is developed in partnership with MediaTek.
Partie GPU Blackwell : 1 PetaFLOP and FP4
For the GPU part, NVIDIA announces a computing power of one PetaFLOP with FP4 precision, or twenty times less than the Blackwell B200 GPU at 20 PetaFLOPs (or 20,000 TeraFLOPs).
We had as a detailed reminder of NVIDIA’s calculation: B200 integrates two GPUs on the same die and goes for the first time to FP4, which allows raw performance to be doubled compared to FP8, itself twice as fast as FP16, etc. GB10 in FP8 is thus at 500 TeraFLOP.
LPDDR5X for memory, NVMe storage
NVIDIA does not give many details on the composition of its GB10 and simply specifies that it is “ featuring an NVIDIA Blackwell GPU with latest generation CUDA Cores and fifth generation Tensor Cores [bref, une architecture Blackwell, ndlr]interconnected via NVLink-C2C to the NVIDIA Grace processor ».
Each DIGITS has 128 GB of unified LPDDR5X memory and a maximum of 4 TB of NVMe (PCIe) storage. The GPU will therefore not have HBM memory (faster than DDR5X) as is the case on high-end versions of Blackwell GPUs for AI.
Local LLMs with up to 200 billion parameters
NVIDIA claims that with its desktop supercomputer, “ developers can run large models with up to 200 billion parameters », provided obviously that you have prepared them for FP4 precision. The manufacturer adds that, thanks to ConnectX, two DIGITS supercomputers “ can be linked to run models with up to 405 billion parameters ».
For NVIDIA, the goal of this GB10 Grace Blackwell chip is to allow “ businesses and researchers to prototype, refine, and test models locally on local systems with the Linux-based NVIDIA DGX OS, then seamlessly deploy them on NVIDIA DGX Cloud “. The hardware architecture and the software part remain the same between Digits and DGX (Blackwell) in the data centers.
Digits will be available in May, starting at $3,000. The details of the configurations are not indicated.
Related News :