Setup matrix

How estimates are made

Every speed on this page carries one of three labels.

• Measured: the recipe's own number, and your setup matches every link field the recipe states (card count, card variant, PCIe generation and width, P2P). Fields a recipe leaves out are listed as not stated. "Measured on N cards" means the recipe runs on N of your cards; no speedup is assumed for the rest.
• Estimated: your setup differs on a point where the wiki has a measured effect. The page starts from the recipe's number, applies each effect below as a range, and lists every step under the recipe.
• Unknown: your setup differs on a point with no measured effect on the wiki. The recipe's own number is shown for reference only.

Effects with a measured basis

• One card, x4 against x16 link width: up to about 3% decode and up to about 4x prefill. Measured once by @exile_1ck (2026-10-07) on one 10 GB card at 40 GB running Strata, an offload engine that streams over the link, so a model held whole in VRAM may gain less. Applied as a range from no change to the full measured effect. Source: PCIe x16 capacitor mod.
• One card, Gen1 against Gen2: up to 9% decode (3 to 9% across answer lengths) and up to 37% prefill. Measured by @eksitus (2026-07-27) on one 10 GB card, Qwen3.6-27B Q8 on ik_llama with MTP, both runs at x4. Applied as a range from no change to the full measured effect. Source: Benchmarks.
• No P2P on a tensor-parallel recipe that ran with P2P, or does not say: decode 0 to 40% lower, prefill 0 to 49% lower. The wiki's A/B results disagree (Contested): no difference on unmodded x4 cards (@deejay9983); with P2P about 8% more decode and 14% more prefill on two cards (@chonsworth); TP2 prefill at 128k 4,900 tok/s with P2P against 2,500 without (@ropuls); MiMo V2.6 Flash TP4 about 40% slower with P2P off (@nathang6866). Recipes marked P2P required are not estimated: they need P2P. Source: P2P.
• More cards than the recipe used: only where the wiki has the same model, quant and engine measured on fewer and on more cards than yours does the page show the range between those two measurements, and it names both recipes.

No measured basis

These differences are shown as Unknown or as a note, never as a number.

• Link width or generation on recipes with two or more cards: The wiki reports that the x16 mod is behind the top multi-card numbers and that each link upgrade raised multi-card tok/s, but gives no multiplier. Source: PCIe x16 capacitor mod.
• PCIe Gen3 and Gen4: No recipe or benchmark on the wiki was measured at Gen3 or Gen4. Source: PCIe Gen3 on the 10 GB card.
• x8 link width: No recipe states x8. Fitting 12 of the 24 capacitors gives x8, and a slot bifurcated x8x8 runs x8 regardless. Source: PCIe x16 capacitor mod.
• SM count: Recipes do not record it. Compute peaks rose about 5.7% from 70 to 74 SMs (the 74/70 ratio), but no tok/s pair exists, at 74 or above. Source: Full SM unlock.
• ECC: A cost of 0 to 10% was stated as an expectation, not measured, and the ECC page lists no measured performance impact. Source: ECC.
• 48 GB on the 10 GB card: No recipe was measured at 48 GB; it changes what fits, not the speed. Source: Memory unlock on the 10 GB card.
• The other card variant: An 8 GB recipe says nothing measured about the 10 GB card, and the other way round (64 GB against 40 GB per card, different HBM). Source: Memory unlock on the 10 GB card.
• Many users (aggregate) decode: None of the measured effects above was measured on aggregate throughput, so aggregate numbers are Measured or Unknown, never Estimated. Source: Benchmarks.

Comparing numbers

• Recipes use different benchmarks. Most use the MiaAI-Lab protocol (400-token answers, temperature 0, thinking off); the lab's bench uses 1,024-token answers; speculative decoding (DFlash2, DSpark, MTP) is on in most rows, and temperature 0 inflates draft acceptance against real use. Each recipe's own note is in its details. Source: Benchmarks.
• Status is recomputed the way the wiki does it: Proven once someone other than the tester adds a confirmation with result Reproduced, otherwise the recipe's own status. Each recipe's details link to the wiki's confirmation form for it, the same form as the "Confirm on your own card" button on the recipe page. Source: Status taxonomy.