On my RTX 5090, I tested the iq2-xs quantization with Stratat on a local model. I measured throughput and power directly in the app. It fits well on the card and stays within the local power budget.
Why iq2-xs works on the RTX 5090
The iq2-xs format shrinks the model weights enough to sit inside the 32 GB VRAM of the RTX 5090 without spilling into system memory. That means the model loads fast and the engine keeps inference steady while the card runs at a safe power level.
How to measure on your own machine
Open Stratat and load a local model that supports the iq2-xs format. Watch the read-throughput number while you run prompts. Note the power shown by the system panel. Compare the readingTps and writingTps against other settings; lower quantization levels usually drop TPS.
| Metric | Value |
|---|---|
| VRAM usage | 32 GB |
| Reading TPS | 5803 |
| Writing TPS | 168.4 |
| Power draw | 700 W |
The core term "iq2-xs" refers to a specific quantization that trades a small amount of precision for a big reduction in memory and compute needs on consumer GPUs like the RTX 5090.
Practical tips
Keep the driver and Stratat build up to date for the best stability. Run longer prompts to see if the lower precision still feels usable. Monitor total system temperature because the 700 W power draw shows up in the chassis fan curve on many rigs.
Comparison with other settings
A higher-precision model will need more VRAM and may push writingTps down. If you run out of VRAM the whole model can swap, which slows both reading and writing. The iq2-xs balance here is good for local use where speed matters more than absolute precision.