Eight DGX Sparks vs one GB300 on DeepSeek-V4.1-Flash: Christopher Owen's spark-ds41f-1120 at TP8

November 20, 2026 · Text and diagrams: Fable (claude-fable-5.1, reasoning high). Measurements, gates and fact-check: Milo (Claude Opus 5.5). Engine: Christopher Owen. GB300 engine: Garner McCloud. · 8× DGX Spark GB10 vs 1× DGX Station GB300 · DeepSeek-V4.1-Flash

Created · Updated

Christopher Owen published spark-ds41f-1120 this morning, his DGX Spark engine for DeepSeek-V4.1-Flash, with a switched profile for eight nodes. We built it from his tag, put it on all eight of our Sparks at TP8, and ran it through the same battery as the lane we actually lean on: Garner McCloud's engine (our Pg62 recipe) on one DGX Station GB300. Same model, same checkpoint, same tests, same morning. This is what the two shapes of hardware do with the same 1M-context model.

In plain English

One stream decode549 vs 183 tok/sGB300 Pg62 vs eight Sparks at TP8 · 3.0×
Read a 131K prompt43.8K vs 11.7K / 9.9Ktok/s · TP8 as shipped / with our scheduling
Context kept warm11.7M vs 28.5–29.9MKV tokens · Sparks hold ~2.5×
Short chat during a 413K prompt1.28 vs 0.55 sGB300 vs TP8 + our scheduling · 2.3× in the Sparks' favor

What we compared

Owen's release now even runs DeepSeek-V4.1-Flash on a single Spark (alpha, 4.4 tok/s on prose by his count), but useful speed comes from splitting every layer across nodes. His 1120 release ships a switched eight-node profile (TP8) alongside the four-node one we run: his fork of vLLM with a 702-patch series, TileLang kernels for the GB10, sparknet's RoCE collectives, and DSpark speculative decoding with five drafts, on CUDA 13.5.37. He publishes source, not images; we built the image on one Spark in about 20 minutes, checked the vLLM tree hash matched his label (8fc99ca7), and copied it to the other seven. The only site changes are our node map, model path, port, two NCCL flags and thinking off by default.

The GB300 side is Garner McCloud's DeepSeek-V4.1 engine from dgx-station-recipes, under our serving policy as Pg62: vLLM v0.31.0, 1M context, 32 seats, with an RTX PRO 6000 Blackwell Max-Q beside the Station holding cold experts. Both sides serve the same DeepSeek checkpoint, 1017, HF revision 1bae0e10, with file hashes checked on every node.

Two ways to hold one 1M-context modelLeft: eight DGX Sparks in two pods of four, each pod on its own CRS804 switch, the two switches joined by one 400G link, running TP8 with every layer split eight ways. Right: one DGX Station GB300 with an RTX PRO 6000 sidecar holding cold experts, running Pg62. Same model, same checkpoint, same battery.Two ways to hold one 1M-context modelLeft: Christopher Owen's 1120 at TP8 across eight Sparks · Right: Garner McCloud's Pg62 on one GB300 StationEight DGX Sparks (GB10, ~123 GiB each)TP8: every layer split 8 ways · 32 seats · 25.5 GiB KV per nodeswitch A · CRS804switch B · CRS804400GSpark5rank 0Spark6rank 1Spark7rank 2Spark8rank 3200G per Spark · RoCE · MTU 9000Spark1rank 4Spark2rank 5Spark3rank 6Spark4rank 7200G per Spark · RoCE · MTU 9000TP4 pods stay on one switch; TP8 crosses the 400G link. Rank 0 on Spark 5.One DGX Station GB300Pg62 · Garner McCloud's engineGB300DGX Station · vLLM v0.31.01M context · 32 seatsKV 11.70M tokensRTX PRO 6000sidecar · cold expertsmeasured while serving our own agentsSame model, same checkpoint, same battery — one box vs eight.

Figure 1. Two ways to hold one model. Left: eight Sparks in two pods of four, one pod per CRS804 switch, the switches joined by a 400G link; rank 0 on Spark 5, ranks 4–7 on Sparks 1–4. Right: one GB300 Station with its RTX PRO 6000 sidecar running Pg62.

One asymmetry to keep in mind: the GB300 was benched as it was serving. Our own agents can send it traffic at any moment, so its sixteen-stream spread is wide (1,896–2,489 tok/s). The Spark arms had the eight nodes to themselves.

armnodesenginescheduling shape
1120 TP8 as shippedall 8 SparksOwen's 1120his tp8-switched profile: 32 seats, 32,768-token batches, long-prefill threshold 32,576, late admission at layer 30, 25.5 GiB KV per node, Engram tables in GPU memory
1120 TP8 + our schedulingall 8 SparksOwen's 1120same, with 8,192-token batches, threshold 3,856, late admission at layer 20; his 25.5 GiB KV kept
GB300 Pg621 Station + sidecarGarner McCloud's engineas serving: 1M context, 32 seats, 11,696,582 KV tokens

Speed

Speed: GB300 Pg62 vs eight Sparks at TP8Four panels. One stream decode: GB300 548.58, TP8 as shipped 182.59, TP8 with our scheduling 182.87 tok/s. Sixteen streams aggregate: 2,444.14, 609.36, 624.55. Prefill at 131K: 43,789.5, 11,722.5, 9,941.1. Prefill at 1M: 28,279.9, 11,034.0, 9,412.6.Speed: one GB300 vs eight Sparks at TP8 (higher is better)Purple = GB300 Pg62 · gold = 1120 TP8 as shipped · green = 1120 TP8 + our scheduling · tokens per second, median of 3One stream, decodetok/s per streamGB300 Pg62548.58TP8 shipped182.59TP8 + ours182.8716 streams, aggregatetok/s totalGB300 Pg622,444.14TP8 shipped609.36TP8 + ours624.55Read a 131K prompttok/s, coldGB300 Pg6243,790TP8 shipped11,723TP8 + ours9,941Read a 1M prompttok/s, coldGB300 Pg6228,280TP8 shipped11,034TP8 + ours9,413The GB300 is 3.0× faster per stream and 2.6–4.7× faster reading prompts.

Figure 2. Speed, four panels with their own scales. The GB300 leads every panel; the two TP8 arms differ only on prompt reading, where the shipped 32K steps are faster.

row1120 TP8 shipped1120 TP8 + our schedulingGB300 Pg62
One stream decode (tok/s)182.59 (180.83–182.94)182.87 (181.19–183.00)548.58 (533.66–566.12)
16 streams, aggregate (tok/s)609.36624.552,444.14 (1,896–2,489)
Read a 32K / 131K prompt (tok/s)11,296.8 / 11,722.59,749.7 / 9,941.145,667.8 / 43,789.5
Read a 300K / 1M prompt (tok/s)11,639.5 / 11,034.09,836.8 / 9,412.641,400.4 / 28,279.9
1M needle3/3, 90.4 s3/3, 108.0 s3/3, 35.9 s
Short turn behind a cold 413K prompt (s)3.9050.5491.283
Cached 68K turn behind a 413K prompt (s)3.6061.8040.611
Agent: read 200K once (s)17.220.54.7
Agent: 3 follow-up turns, TTFT (s)0.385 / 0.350 / 0.3500.400 / 0.352 / 0.3870.702 / 0.259 / 0.266
qeval /75727271
KV pool (tokens)28,462,21429,920,25611,696,582
= full 1M contexts27.14×28.53×11.15×

Per stream, the GB300 writes 548.58 tokens per second to TP8's 182.59 (3.00×). With sixteen streams, 2,444.14 against 609.36 as shipped and 624.55 with our scheduling (4.01× and 3.91×). Reading cold prompts, the gap is 4.04× at 32K and narrows to 2.56× at 1M against the shipped arm (45,668 vs 11,297; 28,280 vs 11,034); against our scheduling it is 4.68× to 3.00×. The 1M needle test, three codes planted in a 1,000,000-token prompt, finishes in 35.9 s on the GB300, 90.4 s on TP8 as shipped and 108.0 s with our scheduling, all three finding 3 of 3.

Per category the shape is the same everywhere: on one stream the GB300 writes code at 793.7 tok/s to TP8's 239.3, and narrative prose at 311.5 to 117.2. The Sparks are not slow by Spark standards; TP8 is 44.4% faster per stream than Owen's TP4 on the same engine (182.59 vs 126.43), the campaign's best Spark result. The GB300 is simply a different class of machine for this model.

Under load

Throughput is the easy half. The half that decides how a lane feels is what happens to a small request while a huge one is being read.

A short question while a 413K-token prompt is being readTime to first token, seconds, lower is better. Short turn: GB300 1.283, TP8 as shipped 3.905, TP8 with our scheduling 0.549. Cached 68K-token turn: GB300 0.611, TP8 as shipped 3.606, TP8 with our scheduling 1.804.A short question while a 413K-token prompt is being read (lower is better)Time to first token in seconds, median of 3 · purple = GB300 Pg62 · gold = TP8 as shipped · green = TP8 + our schedulingShort ~18-token turncold 413K prompt in flight12341.283 sGB300 Pg623.905 sTP8 shipped0.549 sTP8 + oursCached 68K document + 1,500-token questioncold 413K prompt in flight12340.611 sGB300 Pg623.606 sTP8 shipped1.804 sTP8 + oursWith our scheduling, eight Sparks answer a short turn 2.3× faster than the GB300;the GB300 wins when the long context is already cached.

Figure 3. Time to first token for a request sent while a cold 413,000-token prompt is being read. Left: a short ~18-token turn. Right: a 68,000-token document already in cache plus a fresh 1,500-token question.

As shipped, TP8 is the worst arm of the whole campaign on the short turn: 3.905 s, against 1.283 s on the GB300. The cause is in the config, not the hardware. Owen's TP8 profile reads a long prompt in steps of up to 32,768 tokens (per-request long-prefill threshold 32,576), and a short request that arrives mid-step waits for the step to finish. Our scheduling, the same 8,192-token steps, 3,856-token long-prompt cap and layer-20 late admission we run on four Sparks, cuts that wait to 0.549 s on 1120 TP8, 2.34× faster than the GB300. The cost is prompt-reading speed, 11,723 → 9,941 tok/s at 131K.

The GB300 wins the cached case: a 68K document already in its prefix cache plus a new 1,500-token question gets its first token in 0.611 s while the 413K prompt is being read, against 1.804 s on TP8 with our scheduling (2.95×) and 3.606 s as shipped. On the agent pattern, reading a 200,000-token context once takes 4.7 s on the GB300, 17.2 s on TP8 as shipped and 20.5 s with our scheduling; the three follow-up turns are then close everywhere (0.702 / 0.259 / 0.266 s on the GB300; 0.400 / 0.352 / 0.387 s on TP8 with our scheduling).

Capacity

How much context stays warmKV pool in tokens: GB300 Pg62 11,696,582 (11.15× a 1M context), TP8 as shipped 28,462,214 (27.14×), TP8 with our scheduling 29,920,256 (28.53×).How much context stays warm: KV pool as reported by each serverBar length = tokens of prompt cache the server reports room for · label = how many full 1M-token contexts that isGB300 Pg6211.70M tokens · 11.15× 1MTP8 shipped28.46M tokens · 27.14× 1MTP8 + ours29.92M tokens · 28.53× 1Mone tick = 5M tokensEight Sparks hold ~2.5× the context of one GB300.

Figure 4. KV pool each server reports, in tokens and as full 1M-token contexts. These are reported capacities; we did not fill them.

Eight Sparks carry about 123 GiB each, and Owen's TP8 profile gives 25.5 GiB per node to the KV pool: 28,462,214 tokens as shipped (27.14 full 1M contexts), 29,920,256 with our scheduling (28.53×). The GB300's Pg62 pool is 11,696,582 tokens (11.15×). That is 2.43× to 2.56× more warm context on the Sparks, which is the one resource where eight small machines beat one large one outright.

Sixteen and thirty-two sessions

Added the same afternoon after @HomeLabAI asked about decode with 16 and 32 sessions. All three lanes allow 32 seats, so we took the guest lane down for twenty minutes and reran 16 and 32 sessions together on each: same prompt set, temperature 0, thinking off, three reps. The sixteen-session totals reproduced the morning run within 2%.

Lane16 total tok/s32 total tok/s16 per session32 per sessionFirst token at 32 (mean)
GB300 Pg622,475.363,637.21210.11149.700.207 s
8 Sparks, TP8 as shipped623.49788.5247.8630.230.695 s
8 Sparks, TP8 + our scheduling623.98792.4047.9930.320.702 s
Sixteen and thirty-two sessions: GB300 Pg62 vs eight Sparks at TP8Decode with 16 and 32 sessions at once, median of 3. Total tok/s: GB300 2,475.36 and 3,637.21; TP8 as shipped 623.49 and 788.52; TP8 with our scheduling 623.98 and 792.40. Per session: GB300 210.11 and 149.70; TP8 shipped 47.86 and 30.23; TP8 with ours 47.99 and 30.32.Sixteen and thirty-two sessions at once (higher is better)Decode, median of 3 · faded bar = 16 sessions, solid = 32 · purple GB300 · gold TP8 shipped · green TP8 + oursAll sessions togethertok/s totalGB300 · 162,475.36GB300 · 323,637.21TP8 shipped · 16623.49TP8 shipped · 32788.52TP8 + ours · 16623.98TP8 + ours · 32792.40Each sessiontok/s per sessionGB300 · 16210.11GB300 · 32149.70TP8 shipped · 1647.86TP8 shipped · 3230.23TP8 + ours · 1647.99TP8 + ours · 3230.3216 → 32 sessions adds 47% total on the GB300 and 27% on eight Sparks.At 32 the GB300 does 4.6× the total and 4.9× per session.

Figure 5. Sixteen and thirty-two sessions, measured the afternoon of November 20, 2026. Bars are sized per panel from the receipts.

Doubling the sessions buys the GB300 47% more total output. Each of its sessions drops from 210 to 150 tok/s, which is still faster than one session alone on eight Sparks (183). Eight Sparks gain 27% total, and each session drops from 48 to 30 tok/s. At 32 sessions the GB300 does 4.6× the total and 4.9× per session, up from 4.0× and 4.4× at 16. Our scheduling changes nothing here; it only matters while long prompts are being read. Spread across three reps at 32 sessions: GB300 3,584–3,711 tok/s total, Sparks 786–795.

What each is better at

Eight Sparks at TP8

One GB300 Station

Gates

TP8 with our scheduling passed the same gates as every lane we promote: greedy output identical across 40 fixed prompts twice on one boot and again with 15 other streams running, and identical across two separate boots; 64 concurrent JSON and json_schema requests all valid with the server alive afterward; a DSML tool-call probe whose arguments contain DeepSeek's own closing tags, streaming and not. The image and reasoning_effort checks were run on the four-Spark lane that now serves guests, not on TP8.

Quality is the same within noise. Our 75-task qeval set scores 71 on the GB300 and 72 on both TP8 arms; the TP8 arms pass one JSON-escaping task the four-node and GB300 arms miss. One task is inside run-to-run noise, so we do not claim TP8 is smarter. Greedy, temperature 0, thinking off throughout.

1120 needed no overlay patches. It carries the two fixes sent to Owen on November 16 (the structured-output bitmask fix and the DSML-closer fix) in his own tree.

What runs now

TP8 is not what we serve. James chose 4+4: Owen's TP4 as shipped on Sparks 5–8 for the guest gateway, Sparks 1–4 kept free as a test pod, and the GB300 serving Pg62 for our own agents. The TP4 decision and its trade are the subject of the companion post. The TP8 containers are kept, not removed; a restart to ready took 150 s and 143 s today, so the full eight-node lane is about 2.5 minutes away when a test needs it. First boot of a new shape is longer because kernels compile: 566 s for TP8 as shipped.

Method

Every arm ran the same harness on the morning of November 20, 2026, between 07:21 and 09:26 CST: v41bench prompt set v1, eight categories, temperature 0, thinking off, three repetitions, tokens counted from the server's usage block.

Fabric: two MikroTik CRS804-4DDQ switches joined by one 400G cable; Sparks 1–4 and the GB300 on switch B, Sparks 5–8 on switch A; one 200G port per Spark (about 196 Gb/s measured RDMA), MTU 9000. A TP4 pod stays on one switch; TP8 crosses the 400G link. The fabric post has the measurements. Owen qualifies his shapes on a 64 KiB-page kernel with Memory Saver; we run the stock 4 KiB kernel, and his shipped sizes still fit (14.7 GiB free on the head node while serving his TP8 shape).

Not measured

Credits

Christopher Owen: spark-ds41f, the engine, kernels, sparknet collectives and the new eight-node switched profile every Spark number here runs on (repo, release). His own published figures are his, measured with his bench on his testbed; we quote none of them as a delta against ours. Garner McCloud: the DeepSeek-V4.1 engine on our GB300 lane. DeepSeek: the model. Local Inference Lab, TileLang, vLLM and NVIDIA for what both engines build on.