PyTorch Adds New Mode to Make AI Math Far More Accurate on NVIDIA GPUs

Artificial intelligence runs on massive math: multiplying huge tables of numbers. These calculations soak up power, so chip makers created shortcuts. One shortcut, called TF32, lets AI run faster by rounding numbers harshly — but that can cause errors. PyTorch, one of the most popular tools for making AI, just added a third option called BF16x9, which uses smarter math to avoid those errors.

How does it work? Instead of rounding numbers to one small piece, BF16x9 splits each input into three smaller pieces and multiplies them nine ways. This catches nearly all the detail that fast mode throws away. In tests, BF16x9 was about 1,000 times more accurate than TF32 and almost as accurate as using full, chunky FP32 numbers. For AI model builders, that means you can keep speed without sacrificing reliability.

Why should you care? When a self-driving car reads a road sign or a medical chatbot reviews lab results, tiny math errors can become real-world mistakes. More accurate multiplication helps AI give answers that are closer to reality. It also means developers don't have to pick between two extremes: maddeningly slow but perfect math, or quick but sloppy math. There is now a middle path — given the right hardware.

The catch? BF16x9 only works on NVIDIA's very newest 'Blackwell' GPUs (SM100 or SM103) with CUDA 12.9 or newer. Most computers today don't have this chip yet, so the practical effect is limited to early adopters and data centers. Still, it's an important step: AI accuracy is not only about clever software, but also about making the lower-level building blocks more trustworthy.