It has been rumored for some time that Nvidia is preparing a processor based on the ARM architecture. It already has such in the form of Grace and Grace Superchip, while now it is also aiming at lower price levels. It’s called GB10 Grace Superchip and it is a 20-core CPU with ARM architecture, which is complemented by a powerful GPU based on the Blackwell architecture (it also has a 5th generation Tensor Core). This is supposed to achieve 1 PFLOP performance in FP4. It is thus intended primarily for computing with artificial intelligence. The novelty is part of the developer’s mini-PC Project Digitswhich is expected to cost $3,000. It is intended for various data scientists, students, AI researchers. Nvidia collaborated with MediaTek on the processor, and Jensen Huang mentioned that this processor will not only be in the GB10, but MediaTek may use it for its own PC chip (so it’s possible that the Snapdragon X Elite will have competition).
The computer will have a generous 128GB LPDDR5x unified memory, up to 4TB SSD, connectivity via Wi-Fi, Bluetooth and USB, and support for ConnectX (NCCL, RDMA, GPUDirect). All this should be enough to run LLM with 200 billion parameters. If you connect two of these small PCs together via the Nvidia ConnectX interface, they can even handle LLM, which has 405 billion parameters. The operating system here is the Linux-based Nvidia DGX OS, which also allows easy publication of research results on the Nvidia DGX Cloud.
Source: www.svethardware.cz
