New Trick Makes AI Training 31% Faster Without Losing Quality

Training large AI models usually requires hundreds of specialized computer chips called GPUs working together. This is especially true when the AI learns from very long texts, like entire books or months of customer support logs. The problem: those chips need to constantly share information, and that communication is slow and expensive.

Current methods split the long text evenly across all GPUs, as if one class’s assignment was distributed to every student in the school. But when you send many texts at once, this forces every GPU to talk to every other GPU — like a 100-person conference call where each person must speak to all the others. A team of researchers introduced BASP, which instead creates small groups of GPUs and gives each group its own text to handle. This means GPUs mostly chat inside their own team, cutting the overall chatter dramatically.

Tests on professional-grade Nvidia A100 systems using popular Llama and Qwen models showed that BASP makes training 17% to 31% faster than standard methods. Importantly, the model’s accuracy and memory use stayed exactly the same. In practical terms, if a training run took 10 days, this could shave off up to 3 days, which saves huge amounts of electricity and money.

Why does that matter to you? Cheaper training makes it possible for more organizations to build powerful AI that can understand long documents, have longer conversations, or analyze entire research papers at once. The catch: this is an early optimization, and actual speedups depend on the hardware and setup. But it proves that smart logistics—not just smarter algorithms—can make AI more accessible and affordable for everyone.