Is clustering officially supported, or a hack?

Official: NVIDIA's Sync Cluster Assistant supports two to four DGX Spark devices. Two or three connect directly with approved QSFP cables; four requires a 200G switch. Beyond four is undocumented and has no published measurements — we do not sell it.

Should I buy the cluster up front?

Usually no. One box runs the flagship at 3-bit today. Buy boxes two through four when your target model demands the pool: FP8 fidelity at two, GLM-5.2-class frontier models at four. Every box added later is a cable, a switch port and a configuration visit.

Does GLM-5.2 really run? And Kimi K3?

GLM-5.2 runs on four boxes at 22–42 tok/s single-stream — measured by three independent parties with published recipes, and we build that exact configuration as Cluster 512. Kimi K3 (594GB at its smallest quant) has never been demonstrated on any home hardware by anyone, ours included, and we will not pretend otherwise.

Two to four boxes

Capacity, not velocity

NVIDIA officially supports clustering two to four GB10 machines — a pair over a cable, four over a 200G switch, up to 512GB pooled. Here is what that genuinely buys, measured — and the sentence most vendors will not say: it does not make one stream faster.

What more boxes buy

Two: the flagship at FP8-class fidelity instead of 3-bit, and 397B-class models at real quant quality
Four: GLM-5.2 — a 744B frontier model — at 22–42 tok/s, measured by three independent parties
Aggregate serving: ~104 tok/s across 4 streams while each stream stays usable
Failure domains: one box down, the rest still work

What it never buys

Single-stream speed. Decode is autoregressive — every token pays the fabric at least once, so a model that fits one box runs FASTER on one box. The cluster is for models that cannot exist on one box at the quality you want. That is the whole, honest pitch.

The ceiling is the fabric, and it is 1/11th of local memory

One ratio explains every clustering result ever published for these boxes.

25 vs 273 GB/s

Each 200Gb/s ConnectX port moves 25GB/s; local unified memory moves 273GB/s. Models stride the fabric only where the software makes them.

2.56×

Measured decode swing on identical hardware from simply verifying RDMA is actually ON — many container setups silently fall back to TCP with no error.

4.4 tok/s

The same 469B model class on a 3-node TCP/pipeline setup vs 22–42 on 4-node RDMA tensor-parallel. Topology and transport matter more than node count.

Beyond four nodes there is no published measurement anywhere — 8× and 16× clusters exist as arithmetic and forum plans, not evidence. We sell what has receipts.

Measured on real clusters

Community hands-on runs with published recipes — sources with URLs in docs/research/agents/spark-clusters.md. Not projections.

GLM-5.2 (744B) 4 boxes · Int4-Int8Mix · vLLM TP=4 · RDMA · MTP drafter 32.5 tok/s mean, 36 peak single-stream; 65–75 tok/s at 6 concurrent

GLM-5.2 (744B), 1M context 4 boxes · NVFP4 compact-KV · custom vLLM · dual RoCE 22–42 tok/s single-stream; ~104 tok/s aggregate at 4 streams

Qwen3.5-397B-A17B 4 boxes · int4 AutoRound · vLLM TP=4 over 200G switch 37–41.5 tok/s single user; 87–94 tok/s at 4 concurrent

GLM-4.7-FP8 (355B) 4 boxes · SGLang TP=4 · EAGLE speculative decoding 25.1 tok/s — and 9.8 with RDMA silently off. Same hardware.

Qwen3.5-397B (pair) 2 boxes · UD-IQ4_XS 176.7GB · llama.cpp RPC 13.9 tok/s — a 397B model at 4-bit quality on two desks

DeepSeek V4 Flash (pair) 2 boxes · FP8-class fidelity in the 256GB pool The flagship with no quantisation compromise at all

The fine print that travels with the 4-node numbers: driver 580.142 pinned (590+ leaks memory on these boxes), community patches required (one TP=4 kernel bug produced silent zeros — coherent-looking gibberish with no error), and RDMA must be verified on. This is exactly the class of pain our install removes — it is why the cluster ships as a configured product, not a parts list.

Questions, answered straight

Is clustering officially supported, or a hack?: Official: NVIDIA's Sync Cluster Assistant supports two to four DGX Spark devices. Two or three connect directly with approved QSFP cables; four requires a 200G switch. Beyond four is undocumented and has no published measurements — we do not sell it.
Should I buy the cluster up front?: Usually no. One box runs the flagship at 3-bit today. Buy boxes two through four when your target model demands the pool: FP8 fidelity at two, GLM-5.2-class frontier models at four. Every box added later is a cable, a switch port and a configuration visit.
Does GLM-5.2 really run? And Kimi K3?: GLM-5.2 runs on four boxes at 22–42 tok/s single-stream — measured by three independent parties with published recipes, and we build that exact configuration as Cluster 512. Kimi K3 (594GB at its smallest quant) has never been demonstrated on any home hardware by anyone, ours included, and we will not pretend otherwise.

See Cluster 512 Configure one — or a pair