The RTX 5080 and RTX 4070 Ti Super 16GB are compared for running large AI models locally with Stratat. Use your own benchmarks and measurements on your own hardware to decide which card fits your needs.
Why choose one over the other
Both cards deliver a large increase in FP16 and INT8 throughput for local inference. The RTX 5080 offers more cores and memory bandwidth, which helps when running larger models such as Llama-3‑70B quantized to 4-bit. The RTX 4070 Ti Super keeps a lower power draw, making it easier to run inside a quiet home workstation.
How inference performance is measured
Run your own inference workloads with the Strata engine on each GPU. Measure peak memory usage, time per token, and total electricity consumption in kWh. Use the same model, same quantization, and the same dataset for a fair comparison. Record the power cost in euros: at 0,25 €/kWh you can estimate monthly operating cost based on your measured kWh.
| GPU | Memory (GB) | Approx. power (W) | Inference tokens/s (sample) |
|---|---|---|---|
| RTX-5080 | 16 | 300 | 12 |
| RTX-4070 Ti Super | 16 | 220 | 9 |
| Typical idle power | - | 50 | - |
Strata provides deterministic logging for each inference run, so you can verify the measurements yourself.
The core term inference is the process of generating output from a trained AI model after training.
Practical considerations
Make sure your workstation PSU can handle the chosen card. If you plan to leave the card running for many hours, calculate the kWh cost: (power in W / 1000) × hours used × 0,25 €.
When deciding, weigh the higher raw throughput of the 5080 against its higher consumption. The 4070 Ti Super saves electricity and noise while remaining viable for medium-sized models.
Recommendation
If you need the absolute maximum local inference speed and can pay for the extra energy, the RTX 5080 wins. If you prefer a quieter, greener setup with acceptable speed, pick the RTX 4070 Ti Super.