Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

[Benchmark] RTX 4070 Ti 12 GB (sm_89) + 64 GB DDR4, IQ3_S, 128K int8, images on CPU, Windows 10, engine 0.1.50

Open Beginner friendly
  • Is something wrong?
#4,234 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 3 days

Nobody has claimed this yet.

Assessment

Difficulty
1/5
Estimated time
Under an hour
Newbie friendliness
62/100
Issue type
Documentation
Clarity
Clearly specified
Activity status
Active
Tech stack
cpp
Domain
machine-learning

Research direction

The report names tools/strata_bench.py and says it ran suite 1 with three runs per test. Read that script to understand the benchmark format, then compare this report with other community benchmarks. Done means the hardware, software, model configuration, timings, and stated limits are recorded clearly; no correctness check was run.

Written by the indexing model from the issue text.

Description

Community benchmark on NVIDIA GeForce RTX 4070 Ti

Measured on 2026-10-26 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 4070 Ti, 12 GiB VRAM
  • CPU: 13th Gen Intel(R) Core(TM) i5-13600KF (14 cores, 20 threads); installed RAM: 63.8 GiB; storage: not measured
  • OS: Windows 10 (10.0.19045); GPU driver: 621.01; CUDA: not read
  • Strata: project 0.1.51, commit 4b9e05a2; engine 0.1.50
  • Background workloads and power limits: not recorded

Model and configuration

  • Model and quantization: qwen3.8-flash-next-iq3_s
  • Context: 131072 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 --long 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 | 50.2 (44.7-60.3) | 0.77 (0.65-0.82) 4K prompt, fresh | 4019 (4019-4019) | 0 (0-0) | 24 cap | 3 | 1155.7 (856.2-1160.2) | not measured | 3.50 (3.49-4.73) 16K prompt, fresh | 16058 (16058-16058) | 0 (0-0) | 24 cap | 3 | 1441.7 (1383.3-1444.6) | not measured | 11.17 (11.15-11.64) 32K prompt, fresh | 32049 (32049-32049) | 0 (0-0) | 24 cap | 3 | 1443.8 (1443.1-1445.9) | not measured | 22.25 (22.22-22.25) agent session, follow-up turns | 12762 (12744-12787) at the last turn | see below | up to 64 | 3 | see below | 56.2 (46.9-68.3) | 3.06 (2.71-5.12)

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: 223 s.

Additional context (not measured by the tool)

  • RAM: 4 x 16 GB DDR4 (Corsair CMW32GX4M2Z3600C18), XMP running at 3400 MHz CL18, 1.35 V, two DIMMs per channel (3600 MHz was not stable with this CPU's memory controller).
  • Storage: Lexar NM790 2 TB, NVMe PCIe 4.0.
  • Configuration: IQ3_S, every expert in RAM (no low-RAM mode, arena_mib 47962), --kv int8, 131072 context, --kv-streaming off (KV cache in VRAM), images on with the CPU encoder (--vision cpu, --vision-tokens 1024), French draft vocabulary (--draft-vocab fr), one slot.
  • Calibration: --calibrate run on this engine after a reboot: --pcie-frac 0.35 --spec-min-p 0.70, adaptive tier at its default. Pool workers at the engine default (9 on this 6P + 8E CPU); an interleaved A/B of 9 against 13 workers (pool_cores 2,4,6,8,10,12-19, 2 rounds) gave +0.3 %, no difference.
  • GPU package: an RTX 40 card (sm_89) running the sm_86 package, as from 0.2.5.
  • Background: idle desktop, no other GPU work during the run; power limits at stock.
Dominant language
C++
Stars
11.6k
Forks
1k
No merged PRs in 30d

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from Niko1221/Strata

All issues in Niko1221/Strata

Similar issues

More C++ issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.