Advertising disclosure: we earn commissions when you shop through the links below.
NVIDIA introduced the DGX Spark HPC Developer Kit at SC26 on November 16, 2026. It is a four-node desktop cluster of DGX Spark systems wired directly to their two neighbors in a ring over ConnectX-7 networking, so no network switch is needed. The four nodes give 512GB of combined memory and 20TB of NVMe storage, and NVIDIA says models up to 700B parameters can run locally with NVFP4 quantization. It is available through NVIDIA Marketplace and partners including Acer, ASUS, Dell Technologies, GIGABYTE, HP and Lenovo; pricing was not announced.
Sized for production serving of 70B–200B class models at full or lightly-quantized precision. Overkill for a homelab; right call when the workload pays for itself in token volume.
Generated from this product’s spec sheet. Editor reviews refine it over time.
The NVIDIA DGX Spark HPC Developer Kit is a compact, four-node desktop supercomputer designed to bring distributed AI computing and multi-node high-performance computing (HPC) workflows directly to the engineer's desk. Introduced at SC26, the system addresses a persistent bottleneck for machine learning researchers and AI engineers: waiting in shared data center queues just to debug distributed code, test scaling strategies, or run sensitive enterprise models. By packaging four interconnected nodes running Grace Blackwell architecture into a switchless deskside footprint, it carves out a distinct tier between single-chassis prosumer workstations and centralized enterprise clusters.
At the core of the DGX Spark Developer Kit is the GB10 Grace Blackwell Superchip, delivering a unified memory architecture across an Arm-based CPU and Blackwell GPU on each node. The kit provides an aggregate 512GB of shared memory and 20TB of local NVMe storage. Instead of requiring a costly, loud, and complex InfiniBand or Ethernet leaf switch, the four nodes communicate directly through an integrated NVIDIA ConnectX-7 ring topology running at 200 Gbps per link. This architecture gives developers a practical environment for running distributed training scripts, fine-tuning large models, and orchestrating agent swarms locally.
Distributed across OEM partners including Dell Technologies, HP, Lenovo, ASUS, GIGABYTE, and Acer, as well as the NVIDIA Marketplace, the platform targets labs and teams that need absolute data privacy and hardware-level fidelity with production Blackwell clusters. For teams evaluating the best hardware for running AI models locally, the DGX Spark HPC Developer Kit delivers multi-node execution semantics without enterprise data center infrastructure.
Understanding the NVIDIA DGX Spark HPC Developer Kit for AI requires looking past raw monolithic GPU metrics and evaluating distributed system throughput. Because the kit consists of four discrete GB10 nodes, performance depends on both local compute efficiency and inter-node communication latency.
The memory subsystem is the defining performance characteristic for autoregressive inference. Each GB10 node operates with a unified pool of 128GB LPDDR5X memory running at 273 GB/s. For local token generation, 273 GB/s per node is modest compared to the multi-terabyte-per-second HBM3e found on server-grade B200 accelerators. However, because memory is unified between the 20-core Arm CPU and the Blackwell GPU, zero-copy operations and massive context offloading are straightforward.
In a distributed layout, tensor parallelism splits matrix multiplications across the ConnectX-7 ring at 200 Gbps, while pipeline parallelism splits model layers sequentially across nodes. For large language models, the 200 Gbps interconnect allows distributed all-reduce operations without catastrophic serialization delays, though careful layout of tensor versus pipeline parallelism is required to prevent communication bottlenecks from stalling the Blackwell tensor cores.
The Blackwell architecture introduces native NVFP4 (4-bit floating point) tensor cores. Running at FP4 cuts memory footprint in half compared to FP8 and by a factor of four compared to FP16, dramatically reducing memory bandwidth pressure. This architectural feature is what transforms a 512GB aggregate pool into viable hardware for running 700B parameter models at practical response times.
With 512GB of combined memory across four nodes, the DGX Spark HPC Developer Kit handles model architectures that typically require an entire rack of PCIe cards or an HGX baseboard.
For local LLM execution, token throughput varies by model size and parallelism strategy:
The clear sweet spot for this hardware is FP8 and NVFP4. The Blackwell tensor cores process native FP4 formats with full hardware acceleration, preserving model reasoning capabilities while staying within the combined 273 GB/s per-node bandwidth limits.
The DGX Spark HPC Developer Kit serves specific professional scenarios:
This hardware is not designed for casual hobbyists running single, unquantized 8B chatbots, where a single consumer GPU suffices. It is built explicitly for multi-node engineering.
Evaluating the DGX Spark kit requires comparing it against the two most common alternatives for local AI development: multi-GPU workstations and unified Apple Silicon clusters.
A quad-RTX 6000 Ada setup provides 192GB of aggregate VRAM over PCIe Gen 5.
Many developers link multiple Apple Mac Studios over Thunderbolt or 10GbE to assemble 384GB to 576GB of unified memory pools.
For developers seeking the best AI chip for local deployment who also require standard CUDA toolchains and true distributed execution, the DGX Spark HPC Developer Kit provides a purpose-built desktop supercomputer.
The top models this device can run at 4-bit, ranked by fit and speed.
Specs not available for scoring. This product is missing VRAM or memory bandwidth data.

Also in Our Store
Docks, fast storage, cables, and a UPS sized for a local AI machine.

Rent this class of GPU by the hour before you buy. Live prices from RunPod and Vast.ai.

Find the break-even between buying this hardware and paying for a cloud API.
Mac vs NVIDIA for local inference, if you are still choosing a platform.