NVIDIA unveiled the DGX Spark HPC Developer Kit at SC26 in Chicago on November 16 — a four-node ring cluster of Grace Blackwell systems that fits on a desk, pools 512GB of unified memory, and runs models up to 700 billion parameters locally. The pitch is simple: stop waiting in the HPC queue and start running frontier AI at your workstation.
Four Nodes, No Switch Required
The kit links four DGX Spark systems in a ring using their built-in ConnectX-7 networking — no separate network switch needed. Each node carries a GB10 Grace Blackwell Superchip: a 20-core Arm CPU paired with a Blackwell GPU featuring 6,144 CUDA cores and up to 1 PFLOP of FP4 compute, all backed by 128GB of coherent unified memory in a 150×150mm chassis that draws about 140W. Stack four together and you get 512GB combined memory, 20TB of NVMe storage, and enough headroom for NVFP4-quantized models at the 700B parameter scale.
That ring topology matters. Benchmark data shows ring beats a network switch by roughly 10% for single-user inference throughput, and you skip the cost and cable clutter of a separate switch entirely.
The Numbers Worth Knowing
Community benchmarks on Qwen3.6-27B using tensor parallelism over the 200GbE ring put a single DGX Spark at around 12.6 tokens per second. Scaling to four nodes gets you to 33.1 tokens per second — a 2.6x speedup. For pre-training workloads, four-node scaling efficiency reaches 96.2%, yielding a 3.85x speedup. That is respectable distributed scaling for a desk-side device with no data center infrastructure.
| Configuration | Memory | Max Model Size | Throughput (Qwen3.6-27B) |
|---|---|---|---|
| Single DGX Spark | 128GB | ~200B (FP4) | ~12.6 tok/s |
| 4-Node Dev Kit | 512GB | ~700B (NVFP4) | ~33.1 tok/s |
One caveat worth flagging: NVIDIA’s headline claim of running 700B-parameter models locally depends on NVFP4 quantization. DeepSeek V4.1 is the cited example, but DeepSeek V4.1 Flash actually requires around 614GB of GPU memory — the kit has 512GB. You can run large open-weight models at the 700B class with NVFP4, but verify memory requirements for your specific model before assuming the ceiling is universal.
The Queue Time Argument
Cloud GPU costs and data center access are only part of what slows down AI and HPC researchers. The bigger drag is queue time. At most universities and national labs, shared cluster jobs sit in queues for days or weeks. The DGX Spark HPC Developer Kit is not positioned to replace cloud or production data center clusters — it is explicitly a local development and debugging environment. Test your multi-node distributed workload at your desk, find the bugs, then submit the production run to the shared cluster.
NVIDIA has demonstrated this with real scientific workloads: LAMMPS with Kokkos for materials simulation, NAMD and GROMACS for molecular dynamics, OpenFOAM for fluid dynamics, WRF for weather modeling, and NVIDIA NIM microservices including OpenFold3 and Boltz-2 for biomolecular structure prediction. These are not toy demos — they are the workflows that researchers spend weeks queuing for.
Who This Is For and What It Costs
Individual DGX Spark nodes are priced at $4,999 for the 64GB model and $6,950 for the 128GB unit. NVIDIA has not disclosed a bundled price for the four-node kit — the product page shows a “Contact Us” button. At four nodes of the 128GB configuration, you are looking at roughly $28,000 before cables and add-ons. Research lab budget territory, not personal workstation pricing.
The kit is available now through NVIDIA Marketplace and OEM partners including Acer, ASUS, Dell, Gigabyte, HP, Lenovo, and MSI. NVIDIA is showing it live at SC26, which runs through November 20 in Chicago. An HPC SDK playbook with GPU-accelerated C, C++, and Fortran compilers plus MPI tools is listed as upcoming.
Bottom Line
The DGX Spark HPC Developer Kit does not make a case for replacing cloud GPU infrastructure. What it does offer is a credible answer to a real problem: the gap between a single-node AI workstation and the first available shared cluster slot. If your research requires iterating on multi-node distributed workloads, the queue time cost is real, and 512GB of unified memory at your desk is a genuinely different capability than any consumer GPU build can offer. At roughly $28,000, it is not casual spending — but it is in range for departments and labs that currently bill cloud time by the month.

