The rtx4070ti_super-iq3-xxs configuration delivers a high-throughput reading speed of 1700 tokens per second while maintaining a manageable power draw of 435 watts on an RTX 4070 Ti SUPER. With a "Good" fit within the 16GB VRAM limit, this setup allows you to run the model locally without swapping, though you should monitor the writing speed of 72 tokens per second during active generation. To verify these metrics on your own machine, ensure the Strata engine is running with the specific quantization enabled and observe the client dashboard during a sustained prompt load. The power consumption will fluctuate based on the specific workload, but the baseline remains consistent with the hardware capabilities of the RTX 4070 Ti SUPER. You can check your own electricity costs by multiplying the hours of operation by the local rate, which we estimate at $0.17 per kWh for this calculation. This approach ensures you are not exceeding the thermal or power limits of your desktop setup while achieving the desired performance. The following table breaks down the core metrics derived from our measurements to help you plan your hardware requirements accurately.

Understanding the Metrics

To understand how this quantization performs, you must look at the balance between speed and resource usage. The reading speed indicates how fast the model can process input tokens, which is crucial for long context windows. The writing speed reflects the generation latency, which affects the user experience during conversation. Power consumption is the third critical factor, as it impacts both your electricity bill and the thermal environment of your workspace.

Checking Your Setup

You can verify the fit by launching the model and checking the VRAM usage in the Strata client. If the fit is marked as "Good," it means the model fits comfortably within the 16GB buffer without significant overhead. You can also measure the actual power draw using hardware monitoring tools to ensure it stays near the 435-watt mark under load. This helps you avoid unexpected spikes that might trip your circuit breakers or cause thermal throttling. The quantization level iq3-xxs is designed to optimize this balance, providing a sweet spot for performance and efficiency.

MetricValueUnit
VRAM16GB
Reading Speed1700TPS
Writing Speed72TPS
Power Consumption435Watts
Fit StatusGoodStatus

The iq3-xxs quantization is a specific compression method that reduces the model size while preserving essential performance characteristics for local inference.