#3300 [Benchmark] 2x RTX 3080 20 GB + RTX 2000 Ada 16 GB (mixed sm_86/sm_89), IQ3_S, 3-card layer split: 0.2.0-beta

Description

Reported from `tools/strata_bench.py` (suite 1, 3 runs per test, medians with ranges) on an
existing `serve/server.py` install. Full `results.json` (per-request data) is available on
request; it is not pasted here to keep the issue readable.

A note on this machine's shape: the three cards are **2x sm_86 (RTX 3080) + 1x sm_89
(RTX 2000 Ada)**, so it is a *mixed-architecture* layer split. The 0.2.0 sm_86 speed table does
not apply here (`generate.cpp` requires every visible card to share one compute capability), and
the start-up log does not say so — separate feature request on that.

---


Measured on 2026-10-20 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 3080, 20 GiB VRAM
- CPU: Intel(R) Xeon(R) W-2195 CPU @ 2.30GHz (18 cores, 36 threads); installed RAM: 123.2 GiB; storage: not measured
- OS: Linux 7.0.0-38-generic; GPU driver: 595.99.02; CUDA: 13.2
- Strata: project 0.1.45.3, commit e88ee39d; engine 0.1.45.3
- Background workloads and power limits: not recorded

## Model and configuration

- Model and quantization: qwen3.8-flash-next-iq3_s
- Context: 262144 tokens; parallel slots: 2
- 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 | 65.5 (60.3-92.0) | 0.47 (0.43-0.77) |
| 4K prompt, fresh | 4019 (4019-4019) | 0 (0-0) | 24 cap | 3 | 877.8 (826.0-879.1) | not measured | 4.61 (4.60-4.90) |
| 16K prompt, fresh | 16058 (16058-16058) | 0 (0-0) | 24 cap | 3 | 1497.3 (1474.0-1497.5) | not measured | 10.77 (10.77-10.95) |
| agent session, follow-up turns | 14657 (14651-14671) at the last turn | see below | up to 64 | 3 | see below | 70.0 (52.9-87.7) | 4.21 (2.88-5.18) |

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: 69.5 % (69.5-69.5); earlier-turn tokens read again per session: 3 (3-3) (earlier turns that the cache did not hold; the 13811 (13811-13812) new tokens of the files are read once); first turn (nothing cached) first token: 2.46 s (2.38-2.64).

| Turn (run 1) | Prompt tokens | Reused | Read | Prompt tok/s | First token (s) | Decode tok/s |
| ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| 0 | 609 | 0 | 609 | 259.5 | 2.377 | 71.0 |
| 1 | 4119 | 634 | 3485 | 678.4 | 5.177 | 87.7 |
| 2 | 5652 | 4182 | 1470 | 444.9 | 3.342 | 63.8 |
| 3 | 9085 | 5716 | 3369 | 666.9 | 5.092 | 52.9 |
| 4 | 11812 | 9129 | 2683 | 587.2 | 4.608 | 76.0 |
| 5 | 14651 | 11844 | 2807 | 658.9 | 4.303 | 68.6 |

Per-request data: `results.json`. Engine log lines from the start: `engine-log-excerpt.txt` (when this run started the server).

## Correctness and limitations

Not measured: this is a speed suite. The conversation uses the model's own replies, so token counts differ a little between runs. Machine details come from the server, `nvidia-smi`/`rocm-smi` and the OS; background load and power limits are not recorded. Total time of the suite: 151 s.