#2580 bench/results: RTX 5070 Ti 16 GB on Linux, Swift 1.5 IQ3_S at a 1M context (0.1.44.1)

说明

Results only: one new folder, `bench/results/2026-10-14-community-rtx-5070-ti-swift-iq3_s/`, its line in
`docs/COMMUNITY_BENCHMARKS.md` and its row in `bench/results/COMMUNITY.md`. No engine, tool or setup change.

**Machine:** RTX 5070 Ti 16 GB (gen 5 card in a gen 4 x16 slot, 300 W limit), Ryzen 7 5800X3D (AVX2, no AVX-512),
80 GiB DDR4-3200 (76.5 GiB usable), Ubuntu 26.04, kernel 7.0, driver 595.99.02, CUDA 13.1. Strata 0.1.44.1 source
build (`296b59d2`, sm_120), Swift 1.5 IQ3_S, the installed config: 1,048,576-token context (YaRN x4), int8 KV with
32,768 cells in VRAM, MTP 4. The first Linux RTX 5070 Ti report in `bench/results/`.

**Results** (engine timings; medians and ranges of 3):

| | |
| --- | --- |
| `strata_bench.py --long`: chat decode | 61.2 tok/s (54.9-67.4, n=9) |
| prompt at 4K / 16K / 32K | 1,987 / 3,122 / 3,054 tok/s |
| decode, 256 tokens after 4K / 32K / 128K prompts | 72.7 / 70.5 / 66.3 tok/s |
| prompt at 4K / 32K / 128K (same runs) | 1,988 / 3,058 / 2,902 tok/s |
| needles (`needle_bench.py`) | 9 of 9, from 32,778 to 1,034,541 tokens (1M prompt: 1,967 tok/s, 518 s) |
| 262,144 instead of 1,048,576 context, 3 alternating rounds | +6.6% / +5.8% / +7.4% short decode, +5.2% prompt (about 750 more experts in VRAM) |

GPU clocks, power and temperature for every window are in the folder (SM 2,865-2,872 MHz median, 238-263 W median).

**What it leaves out:** the main tables are one warm engine session on a desktop in use (only the context A/B
restarted the engine); one model size; no vision, concurrency, thinking-on or quality test. One agent-suite turn
generated a single token and is listed as a failure. Every limitation is in the README.

**How it was checked:** every number in the README was recomputed from the attached data (results.json,
decode-sweep.json, the needle JSON, the A/B rows and the telemetry CSVs) by a second reader. All files were searched
for user and host names, home paths, addresses and keys. The model shards' SHA-256 match the Hugging Face
repository's.

🤖 Generated with [Claude Code](https://claude.com/claude-code)