NVIDIA’s cheaper 64GB configuration of the DGX Spark desktop AI system reached buyers on October 23, starting at $4,999, TechRepublic reported. The new model keeps the GB10 Grace Blackwell Superchip and NVIDIA’s full software stack but halves the unified memory of the original 128GB system, giving developers a lower-cost way to run AI models locally. The systems are built by six manufacturing partners: Acer, ASUS, Dell, Gigabyte, HP and MSI.
What the 64GB model keeps, and what it gives up
The biggest confirmed difference is memory. The 64GB DGX Spark carries the same GB10 Grace Blackwell Superchip, the same DGX OS operating system and the same ConnectX-7 networking as the 128GB version. NVIDIA says the smaller model supports AI models with up to 100 billion parameters, against 200 billion for the original. Buyers should confirm the exact configuration, storage, ports and regional pricing with their chosen manufacturer, since NVIDIA has not published one universal specification sheet covering all six partner systems.
The software support is identical: DGX OS and NVIDIA’s AI software ecosystem, with support for tools including Ollama, vLLM, PyTorch with CUDA, Nemotron models and the NVIDIA Agent Toolkit. That gives developers a starting point for local inference, coding agents and model experimentation — though TechRepublic notes the 64GB model has not been independently benchmarked, and NVIDIA’s parameter counts are vendor-stated capabilities, not guarantees.
Clustering: two small systems, or one big one
NVIDIA says two 64GB systems can connect directly through their built-in ConnectX-7 networking ports using a QSFP cable, giving 128GB of aggregate unified memory. Its Sync Cluster Assistant helps detect the connected systems and set up the network. In NVIDIA’s own test with the Qwen 3.8 27B model, two clustered 64GB systems delivered up to 1.7 times the performance of a single unit — a result that applies to NVIDIA’s reported test, not a universal guarantee.
The cost comparison matters: two $4,999 systems cost at least $9,998 before cables and accessories. A separate October 21 GTC Berlin update also extended DGX Spark clustering to manually configured clusters of up to eight systems, for up to 1TB of pooled unified memory across eight 128GB units. A Sync Model Launcher meant to simplify downloading and running models on a single DGX Spark or a cluster is arriving later this month, not yet available to every buyer.
Why It Matters
The 64GB model is NVIDIA’s answer to the most common objection to its desktop AI push: price. At $4,999 it still is not cheap, but it undercuts the 128GB configuration enough to put local frontier-model experimentation within reach of smaller teams — the same developers choosing between NVIDIA’s RTX Spark laptops for local AI and cloud GPUs. The clustering story is the strategic tell: NVIDIA is building a ladder from a single desktop to an eight-unit mini-cluster, keeping every rung on its own silicon and software.
That ladder is how NVIDIA’s chips and CUDA built the AI boom — and the DGX Spark line extends the same lock-in from the data centre to the developer’s desk. The open question is whether 64GB is enough for the workloads teams actually run, or whether buyers will end up clustering two units and paying the $9,998 the single 128GB box was meant to avoid.
Background
NVIDIA announced the 64GB DGX Spark configuration on October 2, positioning it as the entry point to its desktop AI platform, and the systems reached buyers on October 23 through Acer, ASUS, Dell, Gigabyte, HP and MSI. The DGX Spark line is NVIDIA’s bid to own local AI development the way its data-centre GPUs own training: GB10 Grace Blackwell silicon, DGX OS, and the CUDA software ecosystem, sold through PC makers rather than cloud providers. The rollout puts that bet in buyers’ hands.
Source: TechRepublic
0 Comments