Nvidia said on 2 October that it is adding a 64GB configuration to the DGX Spark line, its desktop machine built on the GB10 Grace Blackwell Superchip. Nvidia gives a starting price of $4,999 and an on-sale date of 23 October, through Acer, ASUS, Dell, Gigabyte, HP and MSI as well as its own channel.
The headline figure is what the halved memory costs you. Nvidia says the 64GB unit runs inference on models up to 100 billion parameters, and that two of them clustered over the built-in ConnectX-7 networking handle up to 200 billion. The original 128GB DGX Spark was specified at up to 200 billion parameters on a single box. In other words, a pair of the new units is roughly where one of the old ones already sat on model size, though Nvidia says clustering also doubles memory bandwidth.
Why it matters: local inference hardware is where small teams decide whether to keep weights in-house or rent them. A cheaper entry point helps, but the parameter ceiling is the number to check against your actual model, not the price.
Source: Nvidia’s announcement.