On our RTX 5080 running Stratat with the IQ3‑S quantization, we observed 3712 reading TPS and 80.7 writing TPS while generating prompts locally. The GPU consumed 510 watts during the test, confirming stable power use on a desktop machine.
Why IQ3‑S was tested
The IQ3‑S format reduces model size while keeping enough accuracy for practical local deployment. We chose it because it fits within the 16 GB VRAM of our RTX 5080, enabling smaller footprints without sacrificing too much quality.
How the test was done
Measurements were taken with Stratat on a clean desktop install. Reading TPS is the number of prompts successfully processed per second, and writing TPS is the throughput for completed outputs. Power was logged via the card's sensor.
| Metric | Value |
|---|---|
| VRAM usage | 16 GB |
| Reading TPS | 3712 |
| Writing TPS | 80.7 |
| Power draw | 510 W |
The IQ3‑S quantization maps closely to what the RTX 5080 can deliver in terms of real‑world inference speed on locally hosted models.
IQ3‑S means a “intelligence quantization 3 short” engineered for small‑footprint, high‑efficiency local runs on consumer GPUs.
Practical implications
With 16 GB VRAM and the IQ3‑S format, the model fits comfortably on the RTX 5080, leaving time for other tasks. Users should expect around 3712 reads per second under light load, which is useful for interactive applications.
Recommendations
If your workstation has an RTX 5080 and needs to run a locally hosted AI model, consider the IQ3‑S quantization for a balance of speed and memory use. The 510 W draw is typical for the card's performance class and stays reasonable for a desktop setup.