NVIDIA DGX Spark and ASUS Ascent GX10: desktop AI systems

The superchip pairs a 20-core Arm CPU — ten Cortex-X925 cores and ten Cortex-A725 cores — with a Blackwell GPU carrying fifth-generation Tensor Cores and fourth-generation RT Cores. NVIDIA rates the combination at up to 1 PFLOP of FP4 tensor performance, and lists a 140 W TDP for the GB10 itself inside a 240 W total system power figure.

Memory is the architectural point. Rather than a CPU pool and a GPU pool connected over PCIe, GB10 presents 128 GB of coherent unified LPDDR5X that both processors address directly. Model weights do not have to be copied across a bus before the GPU can read them, and the working set is bounded by the whole 128 GB rather than by the capacity of a discrete card.

Storage and networking follow the same desk-side logic. NVIDIA lists a 4 TB self-encrypting NVMe drive on DGX Spark; ASUS offers the Ascent GX10 with 1 TB and 2 TB PCIe 4.0 options alongside a 4 TB PCIe 5.0 option. Both carry a ConnectX-7 SmartNIC — NVIDIA specifies it at 200 Gbps — plus 10 Gigabit Ethernet on RJ-45, which is how two units are linked to work on a model larger than one can hold.

A tower workstation with a professional graphics card gives you more raw compute per dollar for training and far more display and expansion capability, but the model has to fit in the card's own video memory. That is the trade: the workstation wins on throughput and on everything that is not model size, and loses the moment the working set exceeds the card.

A rack server wins on both counts and loses on every other one — it needs a rack, three-phase-capable power, real cooling and a data center or at least a properly conditioned server room. A GB10 system draws 240 W and sits on a desk. DGX Spark measures 150 mm by 150 mm by 50.5 mm and weighs 1.2 kg.

The decision therefore turns on where the work happens rather than on benchmark tables. If a developer needs a large model resident locally, iterating without scheduling cluster time, the desk-side machine is the answer. If the job is production training at scale, it is not.

Storage tier is the main configurable, because the 128 GB of unified memory is fixed by the superchip. The choice is between the 1 TB, 2 TB and 4 TB options and the PCIe generation each one uses — checkpoints and datasets are what fill a machine like this, not the operating system.

Then check the software path. These systems run NVIDIA's DGX OS on Arm, so verify that the frameworks, containers and drivers your team depends on are published for that combination before the purchase order rather than after. An x86 assumption buried in a build script is the most common surprise.

Finally, decide whether one unit is the whole plan. The ConnectX-7 interface exists so that units can be linked for a model that exceeds a single machine's memory — NVIDIA documents connecting up to four DGX Spark systems for models of up to 700 billion parameters — and that is a different cabling and switching conversation from a single desk-side box.