# RTX 3090 Ti Shows 80% BF16 Gain in One Dense Matrix Test

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/40740
Published: 2026-10-26T01:40:49.000Z
Updated: 2026-10-26T01:40:49.000Z
Section: Technology

> The benchmark also lists higher TF32 and FP16 throughput. Community reports show smaller, workload-dependent gains in large language model prefill, and Windows is not formally supported.

The open-source Zagitta/gpu-unlock project uses a Linux kernel module to change floating-point math throughput on NVIDIA Ampere consumer GPUs. In one dense general matrix multiplication (GEMM) test at a 450-watt power limit, the RTX 3090 Ti delivered 157.41 BF16 teraflops, up from 87.50. The result measures matrix performance in that test; it does not establish an equivalent gain across applications.

The same test lists TF32 throughput at 82.61 teraflops, compared with 42.58, and FP16 at 118.39, compared with 87.46. The BF16 increase is about 79.9%. Across cards, BF16 GEMM gains range from roughly 25% to 80%, and power limits can have a large effect. Cards constrained by lower power limits often show smaller gains, so results vary across cards and power conditions.

The module writes speed selectors before NVIDIA’s driver binds, changing FP16 and BF16 fused multiply-add (FMA) issue rates. It does not require flashing a video BIOS or patching the driver again after updates. Linux is the tool’s main platform, and Windows is not formally supported. The project remains experimental; Secure Boot settings may need to change, and stability and warranty effects vary by card.

Large language model (LLM) prefill gains range from 14% to 36%. Prefill is the stage when a model processes its input prompt before generating responses. Matrix-heavy fine-tuning and FlashAttention-style workloads can benefit more, while decode, INT8 and memory-bandwidth-limited tasks often show little benefit. The tool changes operation issue rates, so gains depend on whether computation or data movement limits a workload.

The tool is compatible with several RTX 30-series cards, including the RTX 3090 Ti, RTX 3090, RTX 3080, RTX 3070 and 12GB RTX 3060. Results are not uniform: BF16 throughput remains at half speed on a few original equipment manufacturer cards even after the process runs. The reported compatibility and performance results vary by card and workload.