#2634 bench: RTX 5070 Ti + RTX 3060 vs single 5070 Ti, UD-Q4_K_XL (0.1.44.1)

Descrição

## Title

Related: #2393. Context: #2385; this report measures pinned main, **not** the unmerged planner PR.

## Summary

Same-day community measurements of **Strata 0.1.44.1 (`296b59d2`)** with pinned Unsloth Flash Next **UD-Q4_K_XL**, comparing **RTX 5070 Ti 16 GB + RTX 3060 12 GB** against the **5070 Ti alone**. Intel i9-9900K at **4.7 GHz** (AVX2), **128 GiB DDR4-2800 (2800 MT/s)**, Linux, PCIe 3.0 x8 on both cards, NVMe. CPU/RAM operating settings were confirmed by the owner after measurement on 2026-10-15; they were not independently sampled by the original telemetry. Both cards remain physically installed; the 3060 drives the desktop in every arm.

Two reversed blocks contain **eight fresh services and 360 measured requests**: baseline and saved calibration settings for both GPU selections, 64K context, INT8 KV, 32K resident KV and a fixed 72 GiB resident expert budget. Each service runs the official three-repeat suite including fresh 4K/16K/32K prompts, then nine public calibration-prompt probes.

- Baseline dual improves median short-chat / simulated-agent follow-up decode by **13.57% / 15.32%**.
- Single-GPU baseline improves fresh **16K / 32K prefill by 13.34% / 5.40%**; 4K differs by less than 1%.
- Saved dual calibration does not give a general improvement: chat/probe medians fall; agent median improves 3.07%, but token-weighted decode improves only 0.61% with different generated lengths. Transferred single-GPU tuning regresses the tested decode workloads. No profile/default is promoted.

## What changed

One results-only folder under `bench/results/2026-10-14-community-rtx-5070ti-3060-ud-q4-k-xl/` contains the finished report, all individual request timings/tokens/TTFT/reuse/draft counts, engine logs, losslessly compressed numeric telemetry, exact build/model/asset provenance, calibration source, portable configuration templates, reproduction helpers and an offline verifier. `DATA.md` documents fields and units. Original upstream host GPU inventory is retained and its single-GPU labeling caveat (#2062) is explained.

No engine, server, installer, operational default, model weight or generated pack binary is changed or included. The complete submission is approximately 2 MB; telemetry is gzip-compressed to avoid a large generated-text diff.

## Validation

```text
python3 bench/results/2026-10-14-community-rtx-5070ti-3060-ud-q4-k-xl/verify.py
PASS: 360 requests / 8 services; pooled and weighted rates; comparisons; token/reuse counts; probe answer hashes; safety/OC receipts; relative links; privacy patterns; package checksums.
git diff --cached --check
```

Portable helper compilation/CLI checks, rendering all four configurations with spaced absolute paths and overwrite refusal, and mocked smoke/suite/probe ordering passed offline. No additional GPU run is implied. Backend builds are not applicable to this results-only change.

## Extra Notes

Existing OC is held constant: power limits 330 W / 187 W and unchanged LACT/device-limit receipts. No stock-versus-OC gain is claimed. All services had zero service swap/OOM, but **MemoryHigh reclaim occurred in all runs and global swap peaked at 313 MiB**; these are disclosed alongside temperatures and storage reads.

Two independent service restarts per arm, three within-service repeats, normal graphical background activity and answer-dependent agent continuations limit generalization. No coding correctness, both explicit GPU orders, manual splits, independently calibrated single-GPU optimum, 128K/native context, concurrency or uninterrupted one-hour soak was measured. This is partial unequal-GPU evidence for TEST_REQUESTS.md, not a claim that those outstanding requests are fully satisfied.