Unlocking FP/BF16 issue rate on RTX 3090 with gpu-unlock: what to measure before deciding

What happened and what to do

The post reports that the gpu-unlock project raised measured PyTorch throughput on an RTX 3090 from 75T to 135T, with a 360 watt limit. The GitHub repository describes unlocking the FP/BF16 instruction issue rate up to the equivalent of professional cards. The case shows that FP/BF16 throughput on consumer Nvidia GPUs can differ from professional cards, which matters for ML model tuning.

For a company, the practical point is to treat any gain of this kind as a hypothesis to be measured in its own environment. A reproducible benchmark pipeline, using real training and inference workloads, pinned driver and CUDA versions, and logging of temperature, power and throughput per batch, allows a fair comparison of configurations. It is also worth defining a firmware and hardware modification policy for the fleet, since unapproved changes can affect warranty, stability and internal compliance. In parallel, capacity planning should consider cost per token or per sample and energy use, not only the raw TFLOPS figure.

How the consultancy can help

Diagnosis of your GPU fleet and ML workloads to determine whether FP/BF16 throughput is a real bottleneck. Then design and implementation of a reproducible benchmark pipeline, with dashboards for performance, power and cost per unit of work, plus a hardware and firmware policy for operation in your own environment.

Next step

Send a short description of your case: models in use, available GPUs, training or inference volume, and warranty or power constraints. You receive a scoped proposal with diagnosis, implementation steps and operation.