Nvidia trims its desktop AI box to live with costlier memory

Nvidia cuts the memory in its desktop AI computer to keep the price down as DRAM costs climb.

ChipNews Staff
2 Min Read

Nvidia is adding a smaller configuration to its DGX Spark line, betting that a lower price matters more than maximum memory while DRAM costs keep rising.

The 64GB model goes on sale October 23 through Acer, ASUS, Dell, Gigabyte, HP and MSI at $4,999. The same week, Nvidia lifted the 128GB version to $6,950, roughly 75 percent above where it sat a year earlier, with soaring memory prices behind the move.

Both machines carry the GB10 Grace Blackwell superchip, a 20-core Arm processor co-developed with MediaTek, ConnectX-7 networking and the same CUDA software stack. Bandwidth is unchanged at 273 GB/s, which points to lower-capacity LPDDR5X modules rather than fewer of them.

The smaller box suits inference and private agents rather than heavy fine-tuning. Nvidia says models in the 26 to 35 billion parameter range, such as Qwen 3.8 27B, run comfortably, and the company puts the ceiling near 100 billion parameters.

Two 64GB units can be joined with a cable through the Sync Cluster Assistant, pooling memory to 128GB and lifting model support toward 200 billion parameters. Nvidia measured a 1.7x gain when running Qwen 3.8 27B across a pair.

AMD’s newly launched Gorgon Halo parts compete on memory ceiling, reaching 192GB in top Ryzen AI Max+ 495 systems for less than a 128GB DGX Spark. Nvidia’s answer is to let buyers start small and add a second node later.

Share This Article