A Supercomputer on Your Desk: NVIDIA Unveils the DGX Spark AI Superchip

In 2023, Nvidia unveiled a groundbreaking superchip architecture that unites CPU and GPU through the high-speed NVLink bus. While it delivers far superior bandwidth compared to PCIe, for years its use was confined to data centers and cloud platforms.

Now, the company is bringing this technology into the desktop segment through Project Digits, recently rebranded as DGX Spark. At the Hot Chips conference, lead GB10 architect Andi Skande presented the first detailed look at the new chip.

Fabricated by TSMC using a 3nm process, the GB10 is composed of two dies: a CPU designed by MediaTek and a GPU developed by Nvidia. They are linked via 2.5D packaging and the proprietary NVLink Chip-to-Chip interface, providing up to 600 GB/s of bidirectional bandwidth.

The CPU portion integrates 20 Arm v9.2 cores, split across two clusters — X925 and Cortex A725 — supported by 32 MB of L3 cache and an additional 16 MB of L4 cache to accelerate inter-block communication.

According to Nvidia, the GPU die can deliver up to 1 petaFLOP of FP4 compute (with sparsity) or roughly 31 TFLOPS in FP32. In terms of AI performance, this places it on par with the RTX 5070 (MSRP ~$550), but with far greater efficiency: 140W TDP compared to the 5070’s 250W.

The most striking edge lies in memory capacity — 128 GB of LPDDR5x, compared to the RTX 5070’s 12 GB. While it lacks HBM, the LPDDR5x operates at 9400 MT/s, providing 273–301 GB/s of bandwidth. This makes it especially well-suited for training and fine-tuning models where memory capacity outweighs raw peak speed.

Nvidia claims that DGX Spark can fine-tune models up to 70 billion parameters and run inference on models up to 200 billion parameters. For expanded workloads, two units can be linked via a ConnectX-7 NIC with dual 200GbE ports, effectively doubling compute capacity.

With prices starting at $2,999, the system targets professional developers rather than mainstream users. Yet its greatest strength lies in architectural compatibility with data center-grade solutions: models trained or validated on Spark can be deployed to industrial environments without modification.