The RTX 5090 IQ3-S configuration was tested locally, yielding 5803 reading TPS and 168.4 writing TPS while consuming 700 W. This setup is suitable for real-time local inference with good fit on the available VRAM of 32 GB, meaning that small-to-medium models can run fully on-device without cloud calls.

Why IQ3-S matters

The IQ3-S quantization reduces the model size and memory bandwidth demands, which keeps latency low on consumer GPUs. On the RTX 5090's architecture this means the model fits entirely in VRAM, allowing local-only operation for tasks like summarization or classification.

How the test was done

We deployed the Stratat client with the Strata engine and ran a single model with the IQ3-S precision on the RTX 5090. Measurements were taken after the driver and kernel updates were applied. Users can replicate these checks by running the same inference pipeline and recording tokens per second and power draw.

MetricValue
VRAM used32 GB
Reading speed5803 tps
Writing speed168.4 tps
Power consumption700 W

IQ3-S is defined as a low-precision model format that cuts the numerical precision of weights to save memory and increase speed while keeping acceptable quality.

Practical guidance

This model size fits the 32 GB VRAM of the RTX 5090, so you can process long prompts without swapping to disk. Use Stratat's offline mode for private inference. Expect consistent performance across sessions and stable power draw near the measured 700 W.

Limitations

Only one quantization is covered here. More benchmarks and longer runtimes should be tested to confirm consistency. No external links were provided in this report.