4
DGX Sparks
The fleet
Checking the gateway
Four Sparks. One ring.
Four NVIDIA DGX Sparks, cabled to each other and to nothing else, run one model sharded across all four. DeepSeek V4.1 Flash, 1M tokens of context, vision on, served by the head on bs4 behind one OpenAI-compatible endpoint.
1M
Context, tokens
TP4
Tensor parallel
CX7
Ring, no switch
The hardware
Four boxes, four cables.
No switch in the middle. Each Spark has two neighbours and talks only to them.
Every Spark is a GB10 superchip with 128 GB of unified memory and a ConnectX-7 NIC on each end of the ring. The four are cabled into a loop — head to first, first to second, second to third, third back to head — and the model is sharded four ways across them over RoCEv2. The head is bs4, and it is the only node serving the API.
4×
DGX Spark
GB10 superchip, 128 GB of unified memory each, two ConnectX-7 ports.
1,048,576
Context tokens
The model's native window, not a sliding one. Per-key caps decide what you actually get.
4
Ranks
One model, weight-sharded four ways. The head serves; the other three hold shards.
455 MB/s
Fabric, measured
A file copy between two nodes over the CX7 cable. There is no switch to fan out through.
The stack
What is actually running.
One model on the ring, one gateway in front, one dashboard behind a login.
| Role | What it does |
|---|---|
| Head · bs4 | Serves the model to the whole ring and answers on the fabric at 10.10.0.10:8888. |
| Workers · bs1, bs2, bs3 | Hold the remaining shards. The ring passes traffic around the loop, so no two nodes need a direct cable. |
| Gateway · bs1 | api.snail3d.com — keys, aliases, usage, and the search proxy. |
| Model | DeepSeek V4.1 Flash on SGLang, tensor-parallel across the four ranks, vision on. |
The dashboard is private. It can power nodes off and run benchmarks, so it sits behind
Cloudflare Access. Telemetry and per-node detail live there, not here. What this page shows is
the gateway answering, which is the same thing the model does when you call it.
Why it exists
Own the metal, own the latency.
No shared cloud, no per-token meter on my own work, no queue behind another tenant.
The ring exists to run the things this workshop builds, and to find out what four desktop-class superchips can do when they are wired together in a room. It also runs paid work: a dedicated slot is $2 an hour, one request at a time.