NVIDIA’s blog says the DGX Spark with 64GB of unified memory goes on sale Friday, October 23, from Acer, ASUS, Dell, Gigabyte, HP, and MSI, starting at $4,999. This configuration keeps the same GB10 Grace Blackwell chip, DGX OS, and NVIDIA AI software stack as the 128GB model. The post says one unit can run models of up to about 100 billion parameters on the device, and the agent applications built on them.

[1]
Two dark blocks of the same height. The right block is split by a thin line with a small copper mark in the seam.
The left block is whole. The right block is the same height and split in two, with a small copper mark in the seam. The height does not double. That is two units joined without becoming twice as large. An illustration, not a product photograph., AI-generated illustration, not a news photograph

Two 64GB units can be joined with the NVIDIA Sync Cluster Assistant. In NVIDIA’s own Qwen 3.8 27B test, two clustered systems delivered up to 1.7 times the performance of one system. That is the company’s test, not a third-party retest, and the figure is an upper claim, not a result on every task.

Elsewhere in the same post, two units can connect directly with a QSFP cable, pool memory to 128GB, and support models of up to 200 billion parameters, while delivering twice the memory bandwidth and up to 1.7 times the performance. Twice the bandwidth and 1.7 times the performance are different quantities. Twice the bandwidth is not twice the speed. The post does not give the batch size, the context length, or the software version used as the comparison.

[1]

Each machine includes a ConnectX-7. When scaling from one unit to two, every node runs the same software stack, so nothing needs to be reconfigured. At the end of the month, NVIDIA Sync Model Launcher is also coming. Developers will be able to download and launch Qwen3.8 27B on one system or a cluster, with the launcher placing the model across connected devices and setting up OpenCode so coding can start in a browser. That launcher is described as arriving at the end of the month, not as a feature that ships with the October 23 hardware.

The post uses a “personal supercomputer” label and wording about scaling without friction. Those are the vendor’s phrases. This piece keeps the sale date, the price, the memory, the parameter counts, and the multiple from that test.

[1]

要点

  • The 64GB DGX Spark goes on sale Friday, October 23, from six makers, starting at $4,999, with up to about 100 billion parameters on one unit.
  • In NVIDIA’s own Qwen 3.8 27B test, two units reached up to 1.7 times one unit.
  • Pooling to 128GB and up to about 200 billion parameters comes with twice the memory bandwidth. Twice the bandwidth is not twice the speed.
  • The model launcher is described as arriving at the end of the month, not with the October 23 hardware.