#2000 bench: community report, RTX 3060 + RTX 4060 Ti (Windows, layer split), IQ3_XXS on 0.1.43

本文

Community benchmark, suite 1 (`--long`), 3 runs per test, medians of the runs, produced on the install's own configuration.

**Results.** Fresh prompts: 4K **508.3**, 16K **910.7**, 32K **1087.2** tok/s prefill; first token 7.95 / 17.69 / 29.56 s. Short chat **46.8** tok/s decode. Agent session, 6 turns with four tools: **56.8** tok/s decode, **68.8%** cache hit.

**Why this pair is worth a row.** It is a deliberately mismatched two-card split on Windows:

- RTX 3060 12 GB in an **x4 Gen3** slot (link max width 4 against the card's 16) — the second slot of a B550 board — against an **x8 Gen4** RTX 4060 Ti. The engine's own probe reads **2.9 GB/s** for the 3060 against **13.0 GB/s** for the 4060 Ti, and drops the 3060's expert share to `pcie_frac 0.08` (default 0.55).
- AM4 Ryzen 7 3700X (Zen 2, no AVX-512), 64 GiB DDR4-3000 dual-channel.
- The layer split (0-20 on the 3060, 21-47 and the head on the 4060 Ti) is **written by hand**, not chosen by the placement search: `layer_split auto` picks an extreme on this pair — a two-layer first stage on 0.1.40.2 and 0.1.41, which collapsed decode to about 25 tok/s.

**Leaves alone:** no engine change, no docs, no tooling.

**Two notes on `tools/strata_bench.py`**, both seen while producing this and neither fixed here:

1. The results folder is named from the first GPU only, so this 3060 + 4060 Ti pair was written as `2026-10-11-community-2x-rtx-3060-...`. The README title inside was already correct; the folder here is renamed to `2026-10-11-community-rtx-3060-4060ti-swift-iq3-xxs`.
2. The Settings block prints the config's `args` array only, so it omitted `--layer-split 21`. `layer_split` and `gpu` are siblings of `args` in the config JSON, and the server is what appends the flag to the engine command line — so anyone reproducing from that block would build a single-GPU setup and get different numbers. The README here carries the corrected line, with a note saying why it is there.

Both are reported here rather than as a separate issue because neither is the subject of an open thread; #1985 is about the opposite case, a report listing GPUs that are not in use.

**Comparability note.** This run carries 0.1.42's new `STRATA_ROUTE_TAIL_SKIP=7` default, which changes the answers slightly (KL 0.0058 on an RTX 3060 with IQ3_XXS, top-1 agreement 97.7%), so it is not directly comparable with a 0.1.41 report of the same hardware with the switch off.

🤖 Generated with [Claude Code](https://claude.com/claude-code)