#3609 [Benchmark] RTX 4070 12 GB (165 W), Core i5-12600K, 64 GB DDR5 on Strata v0.1.45

本文

## Problem and result

No baseline community benchmark report exists for an NVIDIA GeForce RTX 4070 12 GB running on Linux with DDR5 memory, dynamic VRAM expert caching, and Transparent Huge Page (THP) kernel acceleration. The only existing RTX 4070 report is an older Windows 11 run with DDR4-3200 and IQ2_XS (`2026-10-08-community-rtx4070-iq2-xs`).

This pull request contributes a community benchmark report and results dataset for Strata v0.1.45 on an RTX 4070 12 GB (165 W power limit cap) paired with an Intel Core i5-12600K and 64 GB DDR5 on CachyOS Linux, testing ISTA-DASLab Qwen3.8-Flash-Next IQ3_XXS at 131,072-token context (`128k`) with native MTP. It provides empirical measurements for the live DDR5-5200 baseline (59.0 t/s decode), DDR5-5800 scaling (65.8 t/s decode), and full uncompressed Q8_0 PLE SSD streaming (74.8 t/s decode).

## Related discussion

Follows the community benchmark contribution workflow in `docs/COMMUNITY_BENCHMARKS.md` (§ "Submit a report").

## Approach

- Created `bench/results/2026-10-18-community-rtx4070-ddr5-5800/` containing `README.md`, `results.json`, `strata-iq3_xxs.json`, and `engine-log-excerpt.txt`.
- Added the report entry under `## Community reports` in `docs/COMMUNITY_BENCHMARKS.md`.
- Added the summary row to `bench/results/COMMUNITY.md`.
- Results-only contribution; no engine or build code modified.

## Validation

**Passed:**
- `python3 serve/server.py --engine strata --config strata-iq3_xxs.json --port 8088`:
  - Revision/Environment: Strata v0.1.45 source build (commit `v0.1.45`, SM89 Ada, CUDA 12.8 / 13.0, CachyOS Linux 7.2.9-2-cachyos, performance governor).
  - Hardware: RTX 4070 12 GB (AD104, 165 W power cap via `nvidia-smi`), Intel Core i5-12600K (6P + 4E, 16 threads, AVX2), 64 GB DDR5 (4x16 GiB).
  - VRAM Cache: 3,086 hot experts cached in GPU VRAM (5.04 GiB VRAM allocation, 75–78% hit rate).
  - Host Arena: Pinned 43+ GB expert arena mapped with Linux 2 MB Transparent Huge Pages (`MADV_HUGEPAGE`).
  - Observed decode throughput:
    - DDR5-5200 live baseline: **59.00 tok/s** median (probes: 54.6, 56.4, 59.9, 65.1 t/s; 2k pp 924.3 t/s, 8k pp 1,899.3 t/s).
    - DDR5-5800 tuned run: **65.80 tok/s** median (probes: 65.1, 65.4, 65.8, 66.5 t/s, peak burst 70.0 t/s; 2k pp 947.0 t/s, 8k pp 1,885.2 t/s).
    - DDR5-5800 + uncompressed Q8_0 PLE (NVMe streaming): **74.80 tok/s** median (probes: 73.8, 74.8, 75.1, 75.4 t/s).
  - Correctness: 4,096-token needle-in-a-haystack retrieval probe passed (100% accuracy, 1/1).

**Failed:**
- none observed

**Not run:**
- Non-CUDA backends (HIP/ROCm, SYCL, Metal): this node is NVIDIA Ada Lovelace only.
- Context lengths beyond 131,072 tokens: the model was validated at native 128k context with `k8v4` KV cache within 12 GB VRAM; YaRN 262k / 1M sweeps were not tested in this session.

## Limits

- Single desktop node measurement with a 12 GB GPU, where 43–45 non-cached expert layers drain through host RAM per token.
- Platforms with DDR4 memory or narrower PCIe links will experience lower decode throughput during host expert evaluation rounds.