The so-called Superchip super chip GB10 Grace Blackwell can be said to be NVIDIA's most anticipated product this year. It is also regarded as NVIDIA's first step into the desktop CPU market (the notebook mobile version is the N1 series), but it failed to be released as scheduled in July.
Recently, NVIDIA officials finally announced many details of the GB10 chip.Specifically emphasized that this is the result of successful cooperation with MediaTek.
In fact, at the beginning of this year, when NVIDIA first announced the desktop mini AI workstation code-named Project DIGITS and later named DGX Spark, it mentioned the joint participation of MediaTek, which mainly contributed to the design of the CPU and interconnect parts.
In terms of technical specifications, GB10 adoptsTSMC 3nm process manufacturing, 2.5D packaging technology, where the CPU part and memory become S-dielet, and the GPU part becomes G-Dielet.
The CPU part integrates 20 Armv9.2 architecture cores, divided into two groups, each group has 10 cores. Each core has its own independent level 2 cache (the capacity is not mentioned), and each group shares 16MB level 3 cache.
The GPU part is based on the Blackwell architecture. The number of CUDA cores is not stated but it should be 6144, which is equivalent to the RTX 5070. There are also fifth-generation Tensor cores, 24MB L2 cache, support for light tracing, DLSS4, FP32 format computing power of 31 TFLOPS, and NVFP4 format computing power of 1000 TOPS.
Between the CPU and GPU is a high-bandwidth, low-power C2C channel based on the NVLink bus architecture.
There is also a 16MB system-level cache, which can be used as the CPU’s level 4 cache., further improving the efficiency of sharing data between different engines such as CPU and GPU.
memory is256-bit wide LPDDR5X unified memory, maximum frequency 9400MHz, the original bandwidth is about 301GB/s.
The overall thermal design power consumption is up to 140W, and the operating system is a customized DGX Base OS. It is not yet known whether it can run other distributions of Linux or even Windows on Arm.
NVIDIA claims,Paired with 128GB of memory, GB10 can run large AI models with up to 200 billion parameters, or fine-tuned models with up to 70 billion parameters.
Two DGX Sparks can be interconnected through the ConnectX-7 bus, and the large model parameters can reach up to 405 billion.
