News · · 9 min read

DGX Spark HPC Developer Kit: a switchless four-node ring, and what NVIDIA's own docs say about MPI

In this post
  1. What the kit changes
  2. What three NVIDIA documents say about the ring
  3. Ring vs switch, in numbers
  4. Our calculations: 512 GB and link speed
  5. Developer forum reactions
  6. What a research lab should check
  7. Limitations
  8. Sources

At SC26 in Chicago on November 16, 2026, NVIDIA announced the DGX Spark HPC Developer Kit: four DGX Spark systems cabled directly to each other in a ring, with no network switch. The pitch is a desk-side cluster where researchers can run multi-node scientific workloads and AI models without waiting in a shared queue. Combined memory is 512 GB.

The short version: for LLM inference, NVIDIA's own measurements put the four-node ring at about 90% of the decode speed of a switched setup running NVIDIA's dedicated communication backend (one model, one request at a time). For the "HPC" in the kit's name, and MPI (the standard for splitting a computation across machines that exchange messages) in particular, NVIDIA's technical blog said a month ago that a switch is required. The HPC setup playbook for the kit has not been published yet. This post covers what the kit changes, how NVIDIA's documents line up, the ring-versus-switch numbers, what developers on NVIDIA's forum have reported, and what a research lab should check before buying.

What the kit changes

There is no new hardware. DGX Spark is a small computer built around one NVIDIA GB10 Grace Blackwell chip with 128 GB of unified memory. The kit packages four of them, linked through the ConnectX-7 ports on the back, as an official configuration. Here is what the NVIDIA announcement states:

ItemAnnouncement
LayoutFour DGX Spark systems, each wired to its two neighbors in a ring
SwitchNone to buy or set up
Memory and storage512 GB combined, 20 TB NVMe
Model sizeUp to 700B parameters with NVFP4 (NVIDIA's 4-bit format)
HPC SDK playbook"Coming soon" (C, C++ and Fortran compilers, libraries, MPI tools)
PurchaseKit on display at SC26 and on sale; DGX Spark units and cables sold via Acer, ASUS, Dell, GIGABYTE, H3C, HP, Lenovo, MSI, xFusion and NVIDIA Marketplace

What changes is the old guidance that four units go through a switch. The FAQ on the DGX Spark product page still says two or three systems can be cabled directly and that "clustering 4 systems requires an external network switch"; its configuration table lists a 200 GbE switch for four units. The kit makes the four-unit, no-switch layout official.

The announcement gives no price for the kit. The only performance figure in it is a customer story: a UC Berkeley lab got results 15x sooner on a DGX Spark by skipping an 11-hour cluster queue. That is a single-unit example, not a measurement of the four-node ring.

What three NVIDIA documents say about the ring

A four-node ring is not new; three NVIDIA documents already discuss it, and they do not quite agree with the announcement. We lined them up (compiled by us):

DocumentGuidance on connecting four units
Product page FAQFour units need an external switch
Multi-Spark playbookA four-unit direct ring works, but only via the manual setup path; NVIDIA Sync supports four units only through a switch
Technical blog, Oct 22A four-node ring gives good inference throughput; point-to-point patterns such as MPI need a switch
SC26 announcementFour-node ring, no switch, HPC SDK playbook coming soon

The playbook's wiring table shows the shape of the ring. Port 0 on each unit goes to port 1 on the next, and a fourth cable runs from Spark 4 back to Spark 1. The playbook then notes: "Spark 1 and Spark 3 are not connected directly. Spark 2 and Spark 4 are not connected directly." Two of the six possible pairs have to go through a third machine.

That extra hop is what matters for MPI. The October 22 blog says: "Anything that depends on point-to-point patterns (for example, MPI-based clustering) requires a switch; a ring would pay a hop penalty." Its summary recommends a ring for "the simplest, lowest-cost four-node inference box" and a switch for eight nodes, split clusters, or communication patterns a ring cannot serve. The new kit carries the HPC label and promises a playbook with MPI tools. No published document yet explains how the kit handles traffic between the two non-adjacent pairs.

Ring vs switch, in numbers

NVIDIA has published its own measurements. Table 2 of the October 22 blog runs the NVFP4 version of Qwen3.8-27B on one to eight nodes in different topologies. The chart below redraws the decode (answer generation) column from that table.

Figure 1. Decode throughput by node count and topology. A four-node ring reaches 90.3% of the switched setup; at eight nodes the ring reaches 76.3%

At four nodes the gap is small. Decode runs at 32.82 tokens/s on the ring and 36.34 on the switch, so the ring gets 90.3% of switch speed. Prefill (reading the input) runs at 4,446 versus 4,556 tokens/s, or 97.6% (ratios calculated by us). NVIDIA explains that at four nodes NCCL (its GPU collective communication library) defaults to ring-based algorithms, so a physical ring matches the traffic pattern.

These numbers cover a narrow case. NVIDIA measured one model split across all nodes, one request at a time, with 8,192 input and 256 output tokens. The switch rows use RoCEnante, a Spark-specific communication backend that needs every node to reach every other node directly, so it only works with a switch. NVIDIA has not published any ring-versus-switch comparison for MPI-based scientific code, so there are no performance numbers yet for the use the kit is named after. At eight nodes the ring falls to 76.3% of the switch, but the kit stops at four.

The "700B parameters" claim holds up for the weights alone. NVFP4 stores one 8-bit scale per 16 values, so each parameter costs about 4.5 bits. 700 billion × 4.5 bits ÷ 8 = 393.75 GB, leaving 118.25 GB of the 512 GB (our calculation; 1 GB = 10⁹ bytes). The operating system and the KV cache share what is left. Per unit, weights take about 98.4 GB of 128 GB. We used the same method to check which models fit on two DGX Sparks in our Inkling 2 Small post.

Link speed is an order of magnitude below memory speed. Per the spec sheet, one DGX Spark has 273 GB/s of memory bandwidth, while one QSFP port tops out at 200 Gb/s, or 25 GB/s. A single ring hop is about 10.9x slower than local memory (our calculation). The more data a job moves between machines, the less it gains from splitting. The two non-adjacent pairs cross that slow link twice, so code that talks between them often looks likely to feel the ring's penalty most.

Developer forum reactions

One day after the announcement, we found no discussion of the kit itself on Reddit or Hacker News. NVIDIA's developer forum, however, has threads from users who built rings of four or more Sparks on their own before the kit existed. All quotes below are from May to July 2026, before the announcement.

Critics focus on the non-adjacent nodes. In May, user dbsci wrote that with a ring "you'll end up needing to add hops whenever you need communication across the ones not directly connected. It sorta defeats the purpose of trying so hard to go without a switch" (NVIDIA forum). In July, user alpaslan.erdag reported that on a six-node ring, RDMA connections between non-adjacent pairs failed outright. In the same thread, elsaco replied: "If you want to cluster more than three Sparks just get a switch" (NVIDIA forum). A six-node ring has more hops than the four-node kit, so that case does not transfer directly.

Others run rings in practice. In July, user FujitsuPolycom, who said the project "started as a response to 'Your Mikrotik 804 is backordered until September…'", reported "roughly 19–20 tok/s single-stream decode" running GLM-5.2 on a four-node ring (NVIDIA forum). The same post notes that "stock NCCL/RoCE on a four-node ring tries to establish queue pairs between non-adjacent ranks" and "the communicator fails before choosing a collective algorithm." The user worked around this with a custom neighbor-only transport layer, falling back to a patched ring-only NCCL for operations not yet moved to it. NVIDIA's October blog does not say what setup it used for its four-node ring run. This is one user's measurement, so it is not a general benchmark. It is also an inference workload. On the forum, the success stories for rings come from inference, and the problems come from workloads where nodes talk to each other directly.

What a research lab should check

For LLM inference, the kit's ring looks like a reasonable choice. NVIDIA's own numbers put a four-node ring at 90% of a switched setup with the dedicated backend, and you skip buying and configuring a 200 GbE switch. That fits labs that want to run open models in the hundreds of billions of parameters on their own hardware because they cannot send internal documents or unpublished data to an outside API.

For developing and debugging MPI simulation code, it looks wiser to wait. NVIDIA's technical blog recommends a switch for that work, and the kit's HPC SDK playbook is not out. When it ships, check first how it handles traffic between the non-adjacent pairs and whether NVIDIA publishes ring-versus-switch numbers. If you need to buy now, it seems prudent to confirm with the seller that a switch can be added later. The multi-Spark playbook warns against mixing direct links and switch links in one cluster, so when you add a switch later, you cannot move just some of the units over.

Treating the kit as a replacement for a shared supercomputer also looks like a stretch. NVIDIA itself frames it as a way to "develop and debug on their own, saving server hours for production runs." The roughly 10.9x gap between link and memory bandwidth points the same way. Large runs stay on the existing cluster; the kit moves the code-checking step before them onto the desk.

Limitations

ItemDetail
Performance dataThe only ring-versus-switch comparison is NVIDIA's own (Oct 22, one LLM inference model). No public HPC or MPI numbers for the kit, and no third-party measurements found
Hands-on testingWe do not have the hardware and did not run the kit. Ratios and memory and bandwidth figures are calculated from published numbers
StorageThe announcement says 20 TB across four units, but the spec sheet lists "up to 4 TB" per unit, which makes 16 TB. We could not confirm why
Price and availabilityNo bundle price has been published, and we did not confirm which regions or resellers carry the kit
CommunityThe forum posts quoted are all pre-announcement, self-built setups. One day after launch, we found no reactions to the kit itself

Sources

1 / 1