#3097 Community benchmark: RTX 5070 Ti Laptop 12 GB + Core Ultra 7 255HX, native IQ3_XXS experts pack, Strata 0.1.44 (tuned 12 GB fork)

Descrição

# Community benchmark on NVIDIA GeForce RTX 5070 Ti Laptop GPU

Measured on 2026-10-18 with `tools/strata_bench.py` (suite 1, 3 runs per test, medians with ranges). Speeds are the engine's own timings; first-token times are measured by the client over HTTP. Speed only: no correctness check was run.

## Hardware and software

- GPU and VRAM: NVIDIA GeForce RTX 5070 Ti Laptop GPU, 12 GiB VRAM
- CPU: Intel(R) Core(TM) Ultra 7 255HX (20 cores, 20 threads); installed RAM: 63.4 GiB; storage: not measured
- OS: Windows 11 (10.0.26200); GPU driver: 617.42; CUDA: not read
- Strata: project 0.1.44, commit 01fc2c80; engine 0.1.44
- Background workloads and power limits: not recorded

## Model and configuration

- Model and quantization: qwen38-uncensored
- Context: 262144 tokens; parallel slots: 1
- Settings: the install's own configuration, unchanged; thinking off for the suite; temperature 0

## Method

Run by `python tools/strata_bench.py` against the HTTP API of a Strata server (already running); model loading is not timed. One warm-up chat and a 1K prompt are discarded first. Every prompt starts with a unique tag, so the conversation cache cannot help the fresh-prompt tests.

- short chat: the three questions of `tools/ab_engine.py`, up to 100 new tokens, 3 runs
- 4K / 16K prompts: generated word lists sized by a calibration request, 24 new tokens, 3 fresh prompts each
- agent session: system prompt and four tools, then 5 turns that each add about 1500 words of a project file and continue from the model's own earlier answers (the way agent clients resend a conversation), 64 new tokens per turn, 3 sessions; the cache numbers come from the answer's usage (`cached_tokens`) and timings (`prompt_n`)
- concurrency (when run): N requests sent at the same moment, total tokens over the wall time of the burst, latency over all requests

## Results

| Configuration | Actual prompt tokens | Reused tokens | Generated tokens | Runs | Prompt tok/s median and range | Decode tok/s median and range | TTFT seconds median and range |
| --- | ---: | ---: | ---: | ---: | --- | --- | --- |
| short chat | 46 (43-51) | 0 | 100 (100-100) | 9 | not measured | 108.8 (92.7-122.7) | 0.41 (0.35-0.45) |
| 4K prompt, fresh | 4019 (4019-4019) | 0 (0-0) | 24 cap | 3 | 1571.8 (1569.7-1643.8) | not measured | 3.01 (2.91-3.03) |
| 16K prompt, fresh | 16058 (16058-16058) | 0 (0-0) | 24 cap | 3 | 2180.0 (2162.9-2214.2) | not measured | 16.90 (16.89-16.96) |
| agent session, follow-up turns | 15265 (15265-15266) at the last turn | see below | up to 64 | 3 | see below | 108.7 (90.5-125.7) | 2.15 (1.35-2.63) |

Ranges are the lowest and highest of the runs; `Runs` counts requests that returned a token.

### Agent session (conversation cache)

Cache hit over the follow-up turns: 70.6 % (70.6-70.7); earlier-turn tokens read again per session: 143 (73-143); first turn (nothing cached) first token: 1.35 s (1.35-1.79).

Per-request data: `results.json`.

## Build measured (the part that differs from a stock install)

This machine runs a locally built engine, not a stock download, so the numbers above are not a stock-install expectation. Everything is public:

- Fork: **https://github.com/zsariboga/Strata**, branch `laptop-12gb`; the tuning write-up is in `docs/LAPTOP_12GB_TUNING.md` there.
- Source tree this report was generated from: commit `01fc2c80` (branch `loopbrk-0441`); the running engine is the locally built `v0.1.44-lb2`.
- Patch set on top of upstream: cache-aware routing (`STRATA_CAR_THRESHOLD/BUDGET/SCOPE`), `STRATA_ROUTE_PRIOR` (+ map / keep-p), `STRATA_TYPICAL_MARGIN/ANSWER`, `STRATA_HC_HOST=1` + `STRATA_HC_Q8=1`, `STRATA_LOOP_BREAK=1`, `STRATA_ADAPT_*` (every 1, 160 swaps, decay 0.9), `STRATA_QSA_SPLIT_ATTN=1`, `STRATA_PREFILL_CPU_SHARE=auto`.
- Experts: a **native-experts pack** (`swift-iq3_xxs`, `experts.bin`, built with `tools/iq_pack.py` from the Swift-Qwen3.8-Flash-Next abliterated IQ3_XXS GGUF, iMatrix-calibrated source; mixed per-layer IQ2_XS/IQ2_XXS/IQ2_S/IQ3_S/IQ3_XXS gate/up and Q2_0/IQ4_NL down), so the experts are read from the pack, not from the GGUF.
- Draft: MTP `rt-tr-v2` with the `dv-wide` draft vocabulary, `--mtp-window 4096`, `--mtp-q4 all`.
- KV: `--kv int8`, `--kv-resident 20480`.

## Steady-state decode on longer answers (not part of the suite)

The suite's chat test caps answers at 100 tokens, which is dominated by warm-up (the same machine measured 86.4 tok/s in an earlier run of the same suite, 108.8 here). For a steady-state figure, 30 fixed questions were sent with `max_tokens 6144` and the decode rate was taken from the engine's own `generated in N ms` lines:

- **143.3 tok/s** (6710 tokens over 46.83 s of engine-reported generation), per-request mean 141.3, median 140.1, range 112.9-164.3.
- Correctness on that set: 30/30.

Method and caveats: the questions and the scorer are local to the fork (not part of `tools/strata_bench.py`), so treat this as a same-machine steady-state reference, not as a comparable community row.

## Limitations

- A browser and an editor were open during the run; `strata_bench.py` asks for a quiet machine and this one was not fully quiet. The same suite measured 86.4 tok/s and 108.8 tok/s in two runs 20 minutes apart, which is the size of the noise here.
- Windows power plan was "Balanced" (CPU max 100 % on AC), not a high-performance plan.
- No correctness suite ran with the benchmark; only the local 30-question check above.
- Not tested: `--clients` (concurrency), `--long` (32K prompt), other quantizations, and any comparison against a stock build on this machine.