How much electricity does RTX 4090 IQ2-XS use?

The setup draws 550 watts, which costs approximately $0.09 per hour at a rate of $0.17 per kWh in the US.

Is 24GB VRAM enough for IQ2-XS?

Yes, the fit is rated as Good, meaning the model fits entirely within the 24GB VRAM without needing system RAM.

What is the token speed on RTX 4090?

Reading speed reaches 4000 tokens per second, while writing speed is measured at 109.3 tokens per second.

RTX 4090 IQ2-XS Local Inference Benchmarks

Analyze RTX 4090 performance with IQ2-XS quantization. Review VRAM usage, throughput, power draw, and running costs for local AI models.

The RTX 4090 paired with the IQ2-XS quantization offers a balanced compromise between model capacity and inference speed, fitting comfortably within the 24GB VRAM limit while delivering robust text generation rates. This configuration allows users to run substantial language models locally without hitting memory bottlenecks, ensuring that the hardware remains responsive during extended sessions.

Performance Analysis

When deploying the Strata engine locally, the primary constraint is often memory rather than compute. The IQ2-XS quantization reduces the model size significantly, allowing it to fit on the RTX 4090's 24GB of VRAM with room to spare. This "Good" fit status means you are not swapping data to system RAM, which would drastically slow down processing. The reading speed of 4000 tokens per second indicates that the GPU can load context windows rapidly, making the initial response feel instant.

Cost and Efficiency

Running this setup requires managing power consumption effectively. The RTX 4090 draws approximately 550 watts under this load. In the United States, where the average electricity cost is roughly $0.17 per kWh, this translates to a tangible operational expense. You can calculate your own hourly cost by multiplying the wattage by the rate and dividing by 1000. For example, running the model for an hour consumes 0.55 kWh, costing about $0.09. While the inference speed is high, the power draw is substantial, so consider this when planning long-running tasks or batch processing.

Verification Steps

To verify these metrics on your own machine, monitor the GPU utilization and power draw using tools like HWInfo or the Strata client dashboard. Ensure your drivers are up to date to maintain the measured throughput. The writing speed of 109.3 tokens per second is the practical output rate you should expect for single-turn responses.

Metric	Value	Impact
VRAM Usage	24 GB	Fits model with headroom
Reading Speed	4000 TPS	Fast context loading
Writing Speed	109.3 TPS	Standard generation rate
Power Draw	550 W	High energy consumption

The IQ2-XS quantization is a specific method of reducing model precision to save memory, defined as a technique that compresses weights to minimize resource usage while preserving core functionality.