# FasterGPU

> Independent GPU performance engineering for companies that own their own GPUs but have
> no dedicated performance team. We measure first, fix the bottleneck, and deliver a report
> the customer can re-run themselves.

## Measured results (our own hardware, not a customer system)

- 88.7% memory-bandwidth utilization on an already-tuned vLLM baseline
  (Qwen2.5-7B, RTX 4090, FP16) — 1.13x headroom remaining
- 1.44x throughput from FP8 weights alone, no scheduler changes (63.0 to 90.5 tok/s, batch 1)
- 55.5% average prefix agreement between FP8 and FP16 outputs (6 prompts, 5 diverged)

These are published so they can be checked, not as a claim about your workload.

## Pages

- [Home](https://fastergpu.com/en/)
- [About](https://fastergpu.com/en/about/)
- [Contact](https://fastergpu.com/en/contact/)

## Research

- (Articles are being published; the measured numbers on the home page are the honest summary.)

## Contact

erik041223@gmail.com