On the RTX 3090 with the iq3-s quantization, the Strata engine runs entirely inside the 24 GB of VRAM and finishes the task without spilling data to system RAM. The read throughput of 2094 tps makes interactive prompts usable from a local laptop or workstation.

Why iq3-s works on this card

Inference is memory-bound, and the iq3-s scheme trims weights while keeping enough precision for natural-language generation. Benchmarks show that the model uses roughly half the VRAM, leaving room for context and safety buffers. No extra acceleration hardware is required.

How to test on your own RTX 3090

The Strata client exposes a simple readout of read tps, write tps, and instantaneous power draw. Measure power with a plug meter and note VRAM usage via the driver monitor. Compare the observed numbers against the published 24 GB limit and 500 W ceiling.

MetricValue
VRAM used24 GB
Read tps2094
Write tps86.5
Power draw500 W

Definition: "iq3-s" is a quantization format that balances accuracy and size to fit large models on limited GPU memory. It is defined by the trade-off between bit depth and precision for the Strata engine's runtime.

Checking performance yourself

Measure read and write throughput with the client's status pane while generating text. Log power consumption over a hour-long session to confirm the 500 W average. Use the same tools to verify the 24 GB VRAM cap on your own machine.

What this means for deployment

Running a locally quantized model keeps data in your own premises and avoids cloud latency. For a small business in France, 500 W sustained power translates to about 0.25 €/kWh, making daily inference costs modest.

Common questions

Q: Can I run this model on other GPUs? A: The settings shown were tuned for the RTX 3090; other cards may require different quantizations. Q: Is the 24 GB VRAM really fully occupied? A: Measured usage hit 24 GB, which is why the fit is described as good but tight. Q: Does 2094 tps make the model interactive? A: On the local hardware, yes, the read speed is high enough for chat-style prompts without noticeable lag.