The RTX 4070 Ti SUPER 16GB with the ud-iq4-xs quantization fits well within the limits of a single local GPU for running the Strata engine. Measured on a typical workstation, it sustains about 1700 reading tps while producing around 72 writing tps, which provides smooth interactive inference. Power draw stays near 435 watts under load, matching the board’s specification.

Why choose the ud-iq4-xs quantization?

The ud-iq4-xs format was designed for small-footprint inference on consumer GPUs with limited VRAM. With only 16 GB of VRAM, the 4070 Ti SUPER can hold the entire model plus the activation buffer, avoiding swaps. The quantization trades a modest amount of accuracy for a sizeable speed gain, making it practical for local workflows without a cloud connection.

How to test on your own RTX 4070 Ti SUPER 16GB

Use the Strata client to load the ud-iq4-xs model profile and run the built-in benchmarks. Check the reported reading and writing tps, note the power reading in watts, and compare against the published values. If the observed reading tps is close to 1700 and power is near 435 W, the setup is healthy and matches the documented performance.

MetricValue
VRAM availability16 GB
Reading tps1700
Writing tps72
Power consumption435 W

The core term "ud-iq4-xs" refers to a uniform-distribution 4-bit integer quantization variant optimized for small models like Strata, used here to shrink the model size while keeping inference fast.

Practical tips for local deployment

Keep the driver updated and use the latest Strata engine release for best compatibility. Run benchmarks before and after changes to confirm that VRAM usage stays within the 16 GB limit. Monitor system power during long inference sessions to ensure the PSU is adequate for sustained loads.