#6498 [Benchmark] RTX 3060 12GB / Qwen3.8 Flash-Next Q2_0 / standard v1 draft.1 + 260K xhigh

open · @bgofche · 0 Kommentare · Auf GitHub

BenchmarksMulti-GPUNVIDIA / CUDAModels & quantsDocumentation

Beschreibung

# RTX 3060 12 GB — Q2_0 up to 260K real input

**Main result:** Single RTX 3060 12 GB eGPU, Strata 0.1.63, Qwen3.8-Flash-Next Q2_0, **Thinking XHIGH**, **GPU Unlock ON**, drafting OFF. At **260,050 actual input tokens:** **1,160.3 prefill tok/s**, **42.8 decode tok/s**, **224.25 s TTFT**, **2/2 synthetic recall checks**.

## Benchmark overview

| Context configured | Actual long-input tokens | Prefill tok/s | Decode tok/s | TTFT | Recall |
|---|---:|---:|---:|---:|---:|
| 32K | 32,051 | **1,422.1** | 49.3 | 22.57 s | 2/2 |
| 64K | 64,044 | **1,386.9** | 49.6 | 46.23 s | 2/2 |
| 128K | 128,030 | **1,308.1** | 47.8 | 97.95 s | 2/2 |
| **262K** | **260,050** | **1,160.3** | **42.8** | **224.25 s** | **2/2** |

**Validation:** 8/8 runs completed, 84/84 required standard requests, 48/48 official sanity checks, 20/20 long-context requests, 8/8 dedicated synthetic recall checks. Optional official concurrency completed in each configuration. 262K hardware telemetry: **72°C peak**, **168.92 W peak**, **10.7 GiB minimum available RAM**; PCIe Gen4 ×4 during all 562 samples.

**Environment:** Minisforum AI X1-255 / Ryzen 7 255 / ~60 GiB RAM; RTX 3060 12 GiB via DEG1 OCuLink Gen4 ×4; Ubuntu 26.04.1, NVIDIA 595.99.02, CUDA 13.2, Strata 0.1.63; Q2_0; thinking `xhigh`, drafting OFF, GPU Unlock ON, vision OFF, parallel 1.

## Evidence and reproducibility

- **[Main report with full results, telemetry, methodology and limitations](https://github.com/bgofche/Strata/blob/benchmark/rtx3060-q2-0-20261121-112854/docs/bench-results/2026-11-21-community-rtx-3060-strata-0.1.63-260k/README.md)**
- Official standard pilot runs: [32K](https://github.com/bgofche/Strata/blob/benchmark/rtx3060-q2-0-20261121-112854/docs/bench-results/2026-11-21-standard-rtx-3060-q2-0-32k/README.md), [64K](https://github.com/bgofche/Strata/blob/benchmark/rtx3060-q2-0-20261121-112854/docs/bench-results/2026-11-21-standard-rtx-3060-q2-0-64k/README.md), [128K](https://github.com/bgofche/Strata/blob/benchmark/rtx3060-q2-0-20261121-112854/docs/bench-results/2026-11-21-standard-rtx-3060-q2-0-128k/README.md), [262K](https://github.com/bgofche/Strata/blob/benchmark/rtx3060-q2-0-20261121-112854/docs/bench-results/2026-11-21-standard-rtx-3060-q2-0-262k/README.md)
- Individual raw JSON/JSONL/CSV, SHA256 manifests and both long-context v3 reproduction scripts are committed under these five report directories. No model weights or raw private host logs.

## Interpretation and limits

- **Different workloads:** the upstream standard protocol `strata-qwen38-thinking-v1-draft.1` has prompts up to ~16K; only the **separate non-standard synthetic extension** tests ~260K *actual input*. The word `draft.1` is the protocol ID, **not** a Draft PR.
- **Synthetic key recall**, not an evaluation of general reasoning quality. **No Unlock-OFF control** was run, so no Unlock-caused speedup can be inferred.
- Some source/tokenizer revisions, runtime fields and full engine logs were unavailable. Any capped official outputs remain reported as capped, not as successful full answers. These are community results submitted for maintainer review, not an upstream-approved leaderboard entry.

Mehr auf der Site

Links zu Install, Modellen, Releases.