The rtx4070ti_super-coder setup runs a single local coder model at 16GB VRAM on a 4070 Ti SUPER. We measured throughput with Strata’s readingTps and writingTps counters while varying prompt length and document size. Power consumption was logged via the card’s built‑in meters.
Why measure local performance on a single GPU
Local deployment eliminates network latency and keeps data on your machine. Quantization reduces model size without losing quality, so you can run larger coder models inside the 16GB VRAM budget of a 4070 Ti SUPER. Measuring yourself confirms the advertised specs against your actual workload.
How we measured and verified results
The test script ran a fixed set of 5 prompts typical for coding assistance. We recorded readsTps = 1700 and writesTps = 72 in the same session to get comparable numbers. Power was measured while the model generated completions.
| Metric | Value |
|---|---|
| VRAM | 16 GB |
| readingTps | 1700 |
| writingTps | 72 |
Coder definition: A coder model is a large language model fine‑tuned for programming tasks, where "coder" denotes its specialization in generating code.
If you want to replicate the test, enable the Strata client on your RTX 4070 Ti SUPER, load the coder quantization, and run the benchmark script with the same prompt batch. Compare the resulting tps and wattage against the published figures.