NVIDIA broke speed records for artificial intelligence processing with their Blackwell computer chips. The company reached 1000 tokens per second using one DGX B200 system containing eight Blackwell processors. Engineers tested the hardware on Meta's massive Llama 4 Maverick model, which has 400 billion parameters. A single Blackwell server can process up to 72000 tokens per second at maximum capacity. Companies measure AI progress through token generation speeds, according to NVIDIA executives.
The chip maker used special software tricks to achieve these breakthrough performance numbers. TensorRT-LLM optimization tools helped boost processing speeds four times faster than before. Speculative decoding became the key technique that made the biggest difference for large language models. This method uses a small, fast computer program to guess upcoming words ahead of time. The main large model checks these predictions at the same time, rather than one after another.
NVIDIA built its system using EAGLE3 software designed specifically for language model acceleration. The company claims this achievement proves its leadership position in artificial intelligence computing. Blackwell processors can handle the largest language models, like Llama 4 Maverick, efficiently. These improvements make AI conversations faster and more responsive for users. Computer interactions will become smoother as processing speeds continue increasing.
The chip maker used special software tricks to achieve these breakthrough performance numbers. TensorRT-LLM optimization tools helped boost processing speeds four times faster than before. Speculative decoding became the key technique that made the biggest difference for large language models. This method uses a small, fast computer program to guess upcoming words ahead of time. The main large model checks these predictions at the same time, rather than one after another.
NVIDIA built its system using EAGLE3 software designed specifically for language model acceleration. The company claims this achievement proves its leadership position in artificial intelligence computing. Blackwell processors can handle the largest language models, like Llama 4 Maverick, efficiently. These improvements make AI conversations faster and more responsive for users. Computer interactions will become smoother as processing speeds continue increasing.
