Nvidia hasn’t yet revealed the specs of its upcoming GB10 desktop AI supercomputer chip, but the details have surprisingly leaked out in the Computex 2025 awards section. The DGX Spark compact supercomputer is revealed to have the same number of CUDA cores as a GeForce RTX 5070 graphics card.
Image source: NVIDIA
The international exhibition Computex 2025 traditionally awards the most unusual, breakthrough and simply unusual products. Previously, this award was often criticized for subjectivity and awarding prizes to overly expensive or impractical products, notes VideoCardz, but this time it became a source of important information – technical data about a mini PC from Nvidia was found on the award page.
Image source: videocardz.com
It turns out that the DGX Spark is equipped with a GPU with 6,144 CUDA cores – exactly the same as the desktop RTX 5070 graphics card. The description of the device says that the DGX Spark, formerly known as Project Digits, uses the Nvidia GB10 Grace Blackwell superchip, which combines the latest Blackwell GPU and a 20-core Grace Arm CPU via NVLink-C2C in a compact package. In addition to 6,144 CUDA cores, the graphics chip received fifth-generation tensor cores, fourth-generation RT cores for ray tracing, as well as new video encoding blocks. The system has 128 GB of shared LPDDR5x memory available to both the CPU and GPU, which allows you to work with AI models with up to 200 billion parameters. The system claims performance of up to 1000 TOPS in FP4 calculations for AI.
Image source: Weibo
Interestingly, the DGX Spark will not be exclusively an Nvidia solution — other companies will offer their systems on the same platform. In particular, Asus presented its version at CES under the name GX10, Lenovo released the ThinkStation PGX, and this week MSI showed the EdgeExpert MS-C931 (pictured above). At the same time, all three systems use the same amount of LPDDR5x memory, which may indicate standardization of the platform.