The RTX 5090 paired with the rtx5090-iq3-xxs quantization delivers exceptional token generation speeds while consuming significant power, making it a high-performance choice for users with robust cooling and electricity budgets. With a measured reading speed of 5803 tokens per second and a writing speed of 168.4 tokens per second, this configuration is optimized for rapid inference rather than energy efficiency. The model fits well within the 32GB VRAM limit of the GPU, ensuring stable operation without the need for offloading to system RAM.

Understanding the Hardware Fit

When you run the rtx5090-iq3-xxs model on an RTX 5090, the primary constraint is the 32GB of video RAM. The "Good" fit rating indicates that the model parameters and context window align efficiently with the available memory, preventing out-of-memory errors during standard use cases. You can verify this fit by monitoring your GPU memory usage in the Stratat client or your system's hardware monitor; if memory usage stays consistently below 32GB, the fit is stable. This efficiency allows the GPU to focus its processing power on computation rather than memory management overhead.

Performance Metrics and Power Consumption

The performance of this setup is defined by two distinct speeds: reading tokens from the context window and writing new tokens to generate a response. The measured reading speed of 5803 tokens per second is exceptionally high, suggesting that the RTX 5090's bandwidth and compute units are heavily utilized during context processing. However, the writing speed of 168.4 tokens per second is more typical for high-quality generation, balancing speed with coherence. It is important to note that this performance comes at a cost. The power consumption is measured at 700 watts, which is substantial. To understand the real-world cost, multiply 700 watts by the duration of your session and the local electricity rate of $0.17 per kWh. This will give you a precise figure for your operational expenses, helping you decide if the speed justifies the energy bill.

How to Measure Your Own Results

To validate these numbers on your own machine, you should use the Stratat engine's logging features or a hardware monitoring tool like HWInfo. Check the tokens per second (TPS) output during a stable generation phase to confirm the writing speed. For power, ensure your PSU and motherboard support the 700-watt draw, as this is near the upper limit for high-end consumer GPUs. Always monitor temperatures to ensure the 700-watt draw does not trigger thermal throttling, which would reduce the 5803 TPS reading speed.

MetricValueUnit
VRAM Capacity32GB
Reading Speed5803tokens per second
Writing Speed168.4tokens per second
Power Consumption700watts
Model FitGoodStatus

The core term iq3-xxs refers to a specific quantization method that reduces model precision to extreme levels to maximize speed and minimize memory footprint, allowing larger models to run on consumer hardware.

Cost Considerations

Using the local electricity price of $0.17 per kWh, running this setup for one hour at 700 watts will cost approximately $0.119. While this may seem small for a single session, it adds up quickly during long training or extensive generation tasks. Consider this cost when planning your workflow. If you are running multiple models or longer contexts, the 700-watt draw will be a recurring expense. Always factor in the ambient temperature of your room, as cooling a room heated by a 700-watt GPU in summer can add significant AC costs, further increasing the effective price of each token generated.