NVIDIA DGX Spark

What it is

The NVIDIA DGX Spark is a small computer built by NVIDIA for running AI models. It belongs to the DGX family, the line of machines NVIDIA designs for AI work, but it sits on a desk rather than in a data-centre rack. Its core component is the GB10 Grace Blackwell Superchip, which combines a 20-core Arm CPU with a Blackwell GPU ^[1].

How it is built

On the CPU side, NVIDIA lists 20 Arm cores: 10 Cortex-X925 and 10 Cortex-A725 ^[1]. Arm designs these cores and licenses them to chipmakers, and the two kinds are tuned for different jobs — the X925 for heavier single-threaded work, the A725 for lighter background tasks — so mixing them lets the system spread work between fast and frugal cores. Putting the CPU and the GPU into one package also lets them share memory instead of copying data back and forth across a separate connection.

NVIDIA rates the system at up to one petaflop of FP4 AI performance ^[1]. A petaflop is a thousand trillion operations per second. FP4 is a very compact way of storing numbers, using four bits each; AI chips use such low-precision formats because shorter numbers let the hardware push more values through the same silicon, trading some numerical accuracy for speed.

Memory and the models it can run

NVIDIA offers the Spark with 64 GB or 128 GB of LPDDR5X coherent unified memory, running at 273 GB/s of bandwidth ^[1]. "Coherent unified" means the CPU and the GPU address one shared pool rather than each keeping its own copy, so a model loaded into memory is visible to both processors without duplication.

Because everything sits in that single pool, the size of model a machine can hold depends mainly on how much memory it has. Parameters are the learned weights of a model, and running inference means putting a trained model to work producing answers rather than training it further. NVIDIA says a single 128 GB DGX Spark can run inference on models of up to 200 billion parameters, and that four connected systems can reach 512 GB and 700 billion parameters ^[1].

Size and power

NVIDIA lists 140 W of thermal design power for the GB10 chip, paired with a 240 W power supply ^[1]. Thermal design power is the amount of heat a chip is built to dissipate continuously, and it sets how much cooling and how large a power supply the machine needs; the gap between the two figures leaves headroom for the rest of the system. The chassis measures 150 mm by 150 mm by 50.5 mm and weighs 1.2 kg ^[1].

Where it fits

NVIDIA's DGX systems have generally been large machines for data centres; the Spark is a desktop-sized system with the same brand on it. That combination — a Blackwell GPU, a coherent memory pool of up to 128 GB, and the ability to link several units together — defines what the machine is for: a single local box in place of a shared server for work that fits within its memory ceiling, and a way to scale up by connecting systems when a model does not fit.