How fast a model would answer on the cards, as this cluster would lay it out
CI / test (push) Successful in 9m1s

For each layout modelplan.Options offers here — one card, tensor parallel
across a node's cards, llama.cpp's layer split, a pipeline across nodes on
a fabric — modelplan.GPUTokensPerSecond gives one reply's speed: the bytes
a token reads over each card's bandwidth, by its model name (a table, SXM
and PCIe apart), plus two all-reduces a layer for tensor parallel (8 us on
NVLink, 25 us and up on PCIe), turns between cards for a layer split, and
a hop per node over InfiniBand or RoCE for a pipeline. The explorer and a
repository's files show the fastest, with the layout; mixed classes are
judged class by class, FP8 on 8.9 cards only.

Calibrated on cuda01's Quadro RTX 6000 (672 GB/s): qwen3-vl-8b Q8, 8.7 GB,
62.1 tok/s measured, 59 estimated; Qwen3-Coder-30B-A3B Q4_K_M, 18.6 GB,
144 measured, which says a model of experts reads 1.6 times its active
share (ActiveShare now does, on the CPU too). Tensor parallel and fabrics
are not measured here.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-10-28 00:19:36 +01:00
1 parent e70d6caa2f
commit f2c6f12cfb
13 files changed
+449 -14

No files matched your search

+13
View File
@@ -163,6 +163,19 @@ model's size — for a model of experts named like `30B-A3B`, its active part.
On the test cluster it came within 15% of what llama.cpp measured. Several
replies at once share the reads and each is slower.
**GPU tok/s** is the same on the cards, for the layout this cluster would
give the model — the fewest cards that hold it, of each class, the fastest of
those; its tooltip says which ("2× h100-80g, NVLink, vLLM", "2 nodes × 1×
a100-80g over ib0 (InfiniBand 200 Gb/s), vLLM"). A card's memory bandwidth is
known from its model name (an H100 SXM and an H100 PCIe differ by half); a card
not in the table gets no figure. Tensor parallel across a node's cards divides
what each reads, and adds two all-reduces a layer, short over NVLink and
longer over PCIe; llama.cpp across cards splits by layer, so the cards take
turns and a reply is no faster than on one; a pipeline across nodes takes turns
too, and adds a hop over the fabric for each node. One card's figures were
measured on the test cluster's RTX 6000; tensor parallel and the fabric are
from what NCCL is known to take, not measured here.
1. Click a model to read its card.
2. Pick one of its files, usually a `q4_k_m` for a first try, and click
**Fetch**. It is downloaded into the mirror once, for every node.