#3491 [Benchmark] 0.1.46.1 GeForce RTX 5060 Ti 16 GB qwen-nvfp4-gptq / RAM 96 GB / PCIe 3.0 / Ryzen 5600 1.5-3.4.-4.4 GHz

Description

# Community benchmark on NVIDIA GeForce RTX 5060 Ti

Measured on 2026-10-20 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

- Mobo : B450
- GPU and VRAM: ("second gpu") NVIDIA GeForce RTX 5060 Ti 16 GiB VRAM GPU Eco Mhz 2150 power 83% /  Stock
- CPU: AMD Ryzen 5 5600 6-Core Processor (6 cores, 12 threads);  1.5GHz, 3.4GHz, 4.4GHz
- RAM (DDR4 3000MHz) : 95.9 GiB
- OS: Windows 10; GPU driver: 616.92; CUDA:  13.13
- Strata: project 0.1.46.1, commit 9ae93f7b; engine 0.1.46
- Background workloads and power limits: "Drivers", tiny wall, Msi monitor.

## Model and configuration

- Model and quantization: qwen-nvfp4-gptq  
NVFP4 GPTQ (qwen-nvfp4-gptq "96GB" [Maximilian228/Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata](https://huggingface.co/Maximilian228/Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata))
- Context: 524288 tokens; parallel slots: 1
- Settings: the install's own configuration, unchanged; thinking off for the suite; temperature 0

(Google FL 8 free, suggestions)
- Orca tokenizer reuse 
-  reduced --expert-cache 2048 (JAN +filesystem, Understory, Exa)
- "some uncensored": true ( [sergqwer/strata-nvfp4](https://github.com/sergqwer/strata-nvfp4))
-  Turn on swap ssd "for stability"

## Method

Run by `python tools/strata_bench.py` 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

### Table 1: Standard Benchmarks (Short Chat, 4K, 16K, Agent Overview)

| Benchmark Workload | Metric | GPU UV + CPU 1.5 GHz | GPU UV + CPU 3.4 GHz | GPU Stock + CPU 3.4 GHz | Delta vs v0.1.45 | GPU Stock + CPU 4.4 GHz |
| :--- | :--- | :---: | :---: | :---: | :---: | :---: |
| Short Chat<br>`~46 prompt -> 100 gen` | Decode tok/s<br>TTFT (s) | 14.9 (13.5 – 16.5)<br>2.45 s (2.17 – 2.69) | 28.0 (24.4 – 30.5)<br>1.25 s (1.10 – 1.42) | 27.2 (25.2 – 31.1)<br>1.25 s (1.08 – 1.37) | +1.1 tok/s (+4.2%)<br>-0.02 s (-1.6%) | 27.2 (24.8 – 31.4)<br>1.16 s (1.01 – 1.27) |
| 4K Prompt (Fresh)<br>`4019 tokens (cap 24)` | Prompt tok/s<br>TTFT (s) | 340.5 (339.3 – 341.6)<br>11.89 s (11.85 – 11.94) | 351.2 (350.8 – 351.3)<br>11.52 s (11.49 – 11.52) | 362.0 (360.1 – 362.1)<br>11.16 s (11.15 – 11.22) | -1.2 tok/s (-0.3%)<br>+0.03 s (+0.3%) | 361.1 (359.5 – 361.4)<br>11.19 s (11.18 – 11.23) |
| 16K Prompt (Fresh)<br>`16058 tokens (cap 24)` | Prompt tok/s<br>TTFT (s) | 571.2 (570.5 – 572.0)<br>28.20 s (28.18 – 28.24) | 582.7 (580.8 – 584.1)<br>27.64 s (27.55 – 27.70) | 624.4 (624.2 – 625.1)<br>25.77 s (25.75 – 25.78) | -2.7 tok/s (-0.4%)<br>+0.11 s (+0.4%) | 626.2 (625.4 – 626.3)<br>25.71 s (25.70 – 25.75) |
| Agent Session (Overview)<br>`~14.7k context @ turn 5` | Decode tok/s<br>TTFT (s) | 15.2 (13.1 – 19.4)<br>10.60 s (10.35 – 10.82) | 26.8 (22.6 – 34.4)<br>10.25 s (10.02 – 10.40) | 27.4 (19.6 – 33.3)<br>10.08 s (9.94 – 10.19) | +1.0 tok/s (+3.8%)<br>-0.07 s (-0.7%) | 28.7 (25.9 – 42.2)<br>10.09 s (9.94 – 10.20) |
| Cache Hit Rate | Follow-up turns | 69.3% (69.3 – 69.5) | 69.4% (69.3 – 69.4) | 69.4% (69.1 – 69.4) | +0.6 p.p. | 69.3% (69.1 – 69.4) |

---

### Table 2: Agent Multi-Turn Session (Conversation Cache Breakdown)

*Context cache hit rate over follow-up turns: ~69.3–69.4% (v0.2.1) vs. ~68.8% (v0.1.45).*

| Turn | Context (Tokens) | Split (Reused / Read) | GPU UV + CPU 1.5 GHz<br>*(Prefill \| TTFT \| Decode)* | GPU UV + CPU 3.4 GHz<br>*(Prefill \| TTFT \| Decode)* | GPU Stock + CPU 3.4 GHz<br>*(Prefill \| TTFT \| Decode)* | Delta vs v0.1.45<br>*(Prefill \| TTFT \| Decode)* | GPU Stock + CPU 4.4 GHz<br>*(Prefill \| TTFT \| Decode)* |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| 0 | ~609 | 0 / 609 (cold) | 68.8 t/s \| 8.94 s \| 11.1 t/s | 71.3 t/s \| 8.58 s \| 27.9 t/s | 72.9 t/s \| 8.42 s \| 18.2 t/s | +0.1 t/s \| 0.00 s \| -3.0 t/s | 72.9 t/s \| 8.41 s \| 18.3 t/s |
| 1 | ~4,083 | ~620 / 3,463 | 324.2 t/s \| 10.78 s \| 13.8 t/s | 340.3 t/s \| 10.24 s \| 25.7 t/s | 344.8 t/s \| 10.10 s \| 26.2 t/s | +35.2 t/s \| -0.05 s \| +4.4 t/s | 344.6 t/s \| 10.09 s \| 25.9 t/s |
| 2 | ~5,616 | ~4,146 / 1,470 | 143.0 t/s \| 10.35 s \| 13.9 t/s | 154.6 t/s \| 10.02 s \| 27.3 t/s | 148.7 t/s \| 9.94 s \| 25.1 t/s | +0.4 t/s \| -0.02 s \| -1.0 t/s | 148.6 t/s \| 9.94 s \| 26.8 t/s |
| 3 | ~9,050 | ~5,680 / 3,370 | 321.1 t/s \| 10.82 s \| 15.2 t/s | 326.5 t/s \| 10.38 s \| 23.7 t/s | 332.9 t/s \| 10.17 s \| 22.7 t/s | +2.2 t/s \| -0.07 s \| -8.2 t/s | 333.4 t/s \| 10.16 s \| 26.5 t/s |
| 4 | ~11,800 | ~9,080 / 2,708 | 256.5 t/s \| 10.66 s \| 19.2 t/s | 267.5 t/s \| 10.20 s \| 30.3 t/s | 270.7 t/s \| 10.05 s \| 32.5 t/s | -13.8 t/s \| -0.06 s \| +3.6 t/s | 269.8 t/s \| 10.09 s \| 28.7 t/s |
| 5 | ~14,640 | ~11,815 / 2,819 | 267.9 t/s \| 10.64 s \| 14.2 t/s | 276.7 t/s \| 10.25 s \| 26.7 t/s | 281.5 t/s \| 10.08 s \| 24.8 t/s | -5.6 t/s \| -0.07 s \| -1.4 t/s | 281.4 t/s \| 10.10 s \| 26.6 t/s |

-----

### Technical Summary: Hardware Bottleneck Analysis  (with Google FL 8 free)

### Hardware Bottleneck Analysis (Developer Brief)

- CPU Scaling & Memory Wall:
  - 1.5 -> 3.4 GHz: +88% decode throughput (14.9 -> 28.0 tok/s). Resolves host scheduler and AVX2 worker starvation.
  - 3.4 -> 4.4 GHz: Marginal +0 to +1.3 tok/s gain, -90 ms TTFT. Execution hits a physical wall: non-resident experts (--expert-cache 2048, 42–69% hit) stream over PCIe 3.0 x16 DMA (7.1 GB/s ceiling) from DDR4-3000 (FCLK 1500 MHz, ~42 GB/s). CPU OC past 3.4 GHz yields memory stall cycles.

- Quantization Efficiency (16GB VRAM + 96GB RAM):
  - GGUF Q4 / Unsloth UD-Q5: Exceeds 16GB VRAM. UD-Q5 requires 108–118+ GB RAM, collapsing to 3–6 tok/s due to NVMe paging.
  - NVFP4 GPTQ: 63.28 GiB arena fits 100% in 96GB RAM (20GB OS/WDDM headroom). Yields 92.1% Q8_0 parity (UD-Q5: 94.6%) at 28 tok/s decode and 626 tok/s prefill via Blackwell sm_120 Tensor Cores.

- Load Telemetry (63.28 GiB @ 2.00 GiB/s in 54s):
  - Exceeds raw PCIe 3.0 x4 NVMe link (~1.91 GB/s). Attributed to Windows Standby List page reuse mapped directly into the host arena via STRATA_PIN_BY_NEED=1.

- Engine [v0.1.45](https://github.com/Niko1221/Strata/issues/2956) vs. v0.1.46 Parity:
  - Synthetic Benchmarks (Table 1): 100% identical token sequence. Decode gained +3.8% to +4.2%; prefill flat (bus saturated).
  - Multi-Turn Session (Table 2): +35.2 tok/s prefill gain in Turn 1 via improved --adapt-async 1 PCIe DMA pipelining.


[README eco gpu cpu 1.5 .md](https://github.com/user-attachments/files/33731728/README.eco.gpu.cpu.1.5.md)
[results eco gpu cpu 1.5.json](https://github.com/user-attachments/files/33731727/results.eco.gpu.cpu.1.5.json)

[README eco gpu cpu 3.4.md](https://github.com/user-attachments/files/33731751/README.eco.gpu.cpu.3.4.md)
[results eco gpu cpu 3.4.json](https://github.com/user-attachments/files/33731750/results.eco.gpu.cpu.3.4.json)

[README cpu 3.4  gpu def .md](https://github.com/user-attachments/files/33731801/README.cpu.3.4.gpu.def.md)
[results cpu 3.4  gpu def.json](https://github.com/user-attachments/files/33731800/results.cpu.3.4.gpu.def.json)

[cmd log gpu def   4.4 cpu.txt](https://github.com/user-attachments/files/33731809/cmd.log.gpu.def.4.4.cpu.txt)
[README cpu 4.4 gpu def.md](https://github.com/user-attachments/files/33731807/README.cpu.4.4.gpu.def.md)
[results cpu 4.4 gpu def.json](https://github.com/user-attachments/files/33731808/results.cpu.4.4.gpu.def.json)