The rt x4070ti_super-coder setup keeps the entire model inside the 16 GB VRAM of a single RTX 4070 Ti SUPER card. This means the model loads entirely on‑device without swapping.
Why this setup works
The coder quantization was chosen because it keeps the parameter count low enough to fit within 16 GB of VRAM. Other quantizations exceeded memory and forced the model to spill to system RAM, which hurt both latency and stability.
How we measured it
On a standard desktop with the RTX 4070 Ti SUPER 16 GB we loaded the Stratat client, ran the coder model and recorded tokens per second for both reading and writing. The read speed hit 1700 tps while the write speed dropped to 72 tps. The card reported a power consumption of about 435 W under load.
| Metric | Value |
|---|---|
| VRAM usage | 16 GB |
| readingTps | 1700 |
| writingTps | 72 |
| powerWatts | 435 |
Core term defined
A quantization is a process that reduces a model's precision to shrink its size so it can run locally on limited hardware like the RTX 4070 Ti SUPER.
Strata lets us run this coder quantization on a laptop or desktop without sending data to a cloud server. The measurement was done by timing a batch of prompts and counting finished tokens per second on the same machine.