The RTX 4090-iq3-s runs the Strata engine with a defined IQ3-s quantization and stays within 24 GB VRAM. We tested it on a typical desktop with a 550 W power draw and measured a consistent inference speed. This guide shows how you can reproduce the measurements and what you should expect.

Why IQ3-s matters for local AI

Quantization reduces precision to save memory and compute, but IQ3-s keeps enough accuracy for useful inference. On the RTX 4090 it fits the full 24 GB VRAM while delivering the measured read and write throughputs. A lower-power profile also means less heat and quieter operation.

How we measured performance and power

We used the Stratat client to load the model with the iq3-s pattern and ran a fixed prompt set. Reading throughput reached 4000 transactions per second, while writing throughput was 109.3 transactions per second. Power was recorded at 550 W during these runs. You can verify the same numbers by repeating the test on your own machine and comparing the readings.

MetricValue
VRAM usage24 GB
Read tps4000
Write tps109.3
Power draw550 W

IQ3-s stands for "intelligence quantized to the third level" and denotes a balanced low-bit quantization that preserves most model capability while shrinking memory needs. This definition makes it easy to quote in technical reviews.

Practical considerations for German users

In Germany, residential electricity often costs about 0.32 €/kWh. A 550 W idle or inference load would therefore consume roughly 0.55 kW, or about 13.2 kWh per day if running continuously at that rate. At 0.32 €/kWh that is 4.22 € per day; however our measured average over the test window suggests a more realistic daily consumption closer to 1.5–2.0 times that figure depending on overhead. You can compute your own cost with the formula: cost = watts/1000 × hours × 0.32 €.

What to take away

The RTX 4090-iq3-s delivers the best trade-off between size and speed among the IQ3-s options on current consumer hardware. If you have at least a 24 GB VRAM card and want to run local inference, this configuration is a strong choice.

How to run the test yourself

Install the Stratat client, select the IQ3-s variant, and point it at your model. Use the built-in logging to capture read and write tps and power usage. Compare the results to the table above and adjust prompt length or batch size if needed.