Test report DSG-7584 · Rev A · tested November 18, 2026

AI Devices & SystemsDevice under test

NVIDIA's DGX Spark HPC Kit: 512 GB, Four Nodes, No Price

NVIDIA's DGX Spark HPC Developer Kit pools four nodes into 512 GB at SC26, but the price is absent, the network docs conflict, and the HPC SDK playbook is not yet shipped.

Read
5 min
Words
919
Node
7nm
Operator
Priya Raman

Spec summary

  1. NVIDIA announced the DGX Spark HPC Developer Kit on November 16, 2026 at SC26: four nodes, 512 GB memory, 16 TB NVMe, no price stated.
  2. NVIDIA claims up to ~700 billion parameters in NVFP4 format, roughly 350 GB of weights, but gives no speed figures.
  3. NVIDIA's SC26 post describes a switchless ConnectX-7 ring; its October 21 product page says four systems need a 200 GbE switch.
  4. The launch came four days after Jensen Huang resigned; Colette Kress is interim CEO.
  5. NVIDIA says its technologies power 81% of the TOP500 supercomputing list.
DGX Spark HPC Kit Links 4 Nodes, 512 GB Memory
Fig. ADGX Spark HPC Kit Links 4 Nodes, 512 GB Memory — AI-generated

NVIDIA announced a four-node desktop cluster with 512 GB of pooled memory on November 16, 2026, and did not state a price. The DGX Spark HPC Developer Kit, shown at the SC26 supercomputing conference, combines four DGX Spark systems with 16 TB of NVMe storage. It is on sale through the NVIDIA Marketplace and partner OEMs. That is what NVIDIA has committed to. The rest of the story carries caveats worth counting.

What did NVIDIA announce at SC26?

The blog post, dated November 16 and updated November 17, names the kit explicitly. DGX Spark alone still means the single desktop machine. The HPC Developer Kit means the four-node bundle. It adds to the family; it retires nothing.

The stated workloads are multi-node scientific computing, AI models and agents. The audience is researchers, academics and developers. NVIDIA frames the pitch narrowly: debug message-passing code, check scaling behavior and tune memory use at a desk, without waiting in a shared supercomputer queue.

One software gap sits inside the launch. NVIDIA said an HPC SDK installation playbook is "coming soon." No public release date exists. The hardware is orderable today; part of the software story is still a roadmap item.

What are the specifications?

NVIDIA's on-record figures:

  • Nodes: four DGX Spark systems, each built on the GB10 Grace Blackwell Superchip
  • Memory: 512 GB combined, at 128 GB per node
  • Storage: 16 TB NVMe total, up to 4 TB per node
  • Interconnect: ConnectX-7 (NVIDIA's network adapter chip) shown in a direct ring, no switch
  • Model claim: up to roughly 700 billion parameters in NVFP4 (a 4-bit weight format)
  • Price: not stated

Each node carries a 20-core Arm CPU, ten Cortex-X925 and ten Cortex-A725 cores, a Blackwell GPU, and 128 GB of shared LPDDR5x. NVIDIA rates one node at 1 petaFLOP. The chassis measures 150 by 150 by 50.5 mm — roughly the footprint of a hardcover book.

The OEM list runs long: Acer, ASUS, Dell, GIGABYTE, H3C, HP, Lenovo, MSI and xFusion. The presence of H3C and xFusion suggests NVIDIA expects demand from Asian research institutions.

Does the 700-billion-parameter claim hold?

Partly, and only as a capacity statement. At 4 bits per weight, 700 billion parameters need roughly 350 GB for weights alone. That leaves about 160 GB for KV cache, activations, the OS and the stack. The arithmetic fits.

What NVIDIA did not say matters more. The announcement gives no tokens-per-second figure. Token generation depends on memory bandwidth, and LPDDR5x is slower than the memory on a discrete flagship card. Treat 700 billion as what fits, not how fast it runs. NVIDIA cited DeepSeek V4.1 as its example; community ports of that model already run on a single 128 GB GB10 machine.

Why do NVIDIA's own pages disagree on the network?

The SC26 post describes a switchless ring over ConnectX-7, with each node cabled to two neighbors at 200 Gb/s-class speed using RoCE (RDMA over Converged Ethernet). The DGX Spark product page, dated October 21, says a four-system configuration needs a 200 GbE switch, with direct cabling only for two or three systems.

Both statements remain live. Neither is withdrawn. The likely reading is that the kit ships as a validated ring and the older page lags, but that is analysis, not confirmation. Buyers should ask their OEM which topology ships, and whether cables or a switch are in the box.

Where does the kit sit in the scaling ladder?

NVIDIA has built a tiered structure since October:

  • 1 node: 128 GB, on sale; a 64 GB option ships via partners only
  • 2-3 nodes: 256-384 GB, direct QSFP cabling, user-built
  • 4 nodes (this kit): 512 GB, announced November 16
  • Up to 8 nodes: 1 TB unified memory, manual configuration with NVIDIA Sync support, announced October 21

A caution on the headline number: four 64 GB nodes total 256 GB, not 512 GB. The figure holds only with 128 GB per node. A developer blog also cites a 753-billion-parameter GLM 5.3 model across eight systems; treat that as a vendor demonstration until independent testers reproduce it.

Who should buy, and who should wait?

University labs, research software engineers testing MPI-style and NCCL-style code paths, and small AI teams with data that cannot leave the building fit the kit well. Training-focused teams do not; NVIDIA itself pitches single nodes for fine-tuning up to 70 billion parameters. Production inference teams should hold off. The announcement offers no redundancy story and the software playbook has not shipped.

What changed around the launch?

The kit landed four days after Jensen Huang resigned as president, CEO and board member on November 12. Colette Kress, the CFO, became interim CEO while keeping her finance role. No permanent successor had been named. The kit's blog post, written by Allen Bourgoyne, reads as a roadmap item already in motion. The practical risk is small; the machines use shipping GB10 hardware regardless of who holds the top job. The softer risk is software cadence under new leadership.

The strategic logic is software gravity. Code tuned on four Blackwell-based Arm machines lives in the same CUDA ecosystem as NVIDIA's data-center systems. NVIDIA says its technologies power 81% of the TOP500 list, and a desk-side on-ramp protects that share. Competitors sell high-memory mini PCs; none has announced a comparable four-node, vendor-backed research bundle.

via shattered.io (Source)

Filed under

  • nvidia
  • dgx-spark
  • hpc
  • gb10-grace-blackwell
  • ai-infrastructure
Share this article:

More from Priya Raman

Priya Raman

Show full bio

Correspondent covering money, funding and market shifts around hardware layer of artificial intelligence at Die Signal.

661 articles

Same lot · LOT-362E

« Previous articleNext article »