NVIDIA unveils PC with 72-core CPU and Blackwell GPU

NVIDIA has unveiled the DGX Spark and DGX Station systems, bringing data center-level power to the desks of everyday users. Powered by the latest Grace Blackwell chips, these devices are set to revolutionize AI training and deployment, making high-performance computing accessible beyond cloud servers.

The DGX Spark is positioned as the world's smallest AI supercomputer. At its core is the GB10 chip, which combines a 72-core Grace processor with a Blackwell GPU featuring fifth-generation tensor cores. The system supports FP4 computing, delivering up to 1 quintillion operations per second — critical for fine-tuning neural networks.

A key innovation is the NVLink-C2C technology, which is five times faster than PCIe 5.0. This eliminates bottlenecks in data exchange between the CPU and GPU, which is particularly crucial for memory-intensive tasks. The system includes 128 GB of shared memory and 4 TB of NVMe storage, enabling local work on large models with seamless cloud integration.

For more demanding workloads, the DGX Station features the GB300 chip, which combines a Grace processor with a Blackwell Ultra GPU, delivering a record-breaking 20 petaflops of FP4 performance. The system offers 784 GB of unified memory (288 GB HBM3e + 496 GB LPDDR5X), enabling training of models that previously required clusters of dozens of servers. Special attention has been given to networking: the ConnectX-8 SuperNIC adapter, with 800 Gbps bandwidth, allows multiple stations to be linked into a single computing cluster.

The systems will be produced by leading vendors — ASUS, Dell, HP, and Lenovo — ensuring global availability. Experts predict that the launch of DGX systems will accelerate the rise of "personalized AI." With this level of desktop power, even startups could compete with industry giants in the AI innovation race. Initial shipments are expected in Q4 2025, but NVIDIA has not yet disclosed pricing details.