Lightmatter just started selling a new kind of computer chip that uses light instead of electricity. Their latest processor puts six chips together in one 3D package, with each chip having special light-based cores that speed up AI tasks. Scientists wrote about it in Nature magazine, showing how it combines around 50 billion transistors with a million light parts, all connected through fast light links. Regular computer chips have hit walls because they can't keep getting smaller or faster like they used to. The light-based chip fixes these problems with something called adaptive block floating point format that controls light signals better.
This new processor can do 65.5 trillion math operations every second but only needs 78 watts of electricity and 1.6 watts of light power. It runs normal AI models with almost perfect accuracy without changing them. Users can run big famous AI systems like ResNet for pictures, BERT for language, and DeepMind's game-playing programs straight out of the box. No one needs to retrain these models or tweak them for the light chip. This marks the first time anyone has sold a light-based AI chip that works with regular software right away.
The chip works differently because it uses actual light to do calculations instead of just electricity. This helps solve problems with next-generation graphics cards that cost tons of money and use huge amounts of power. The company made sure their chip works with popular AI software like PyTorch and TensorFlow. They hope companies will start using their chips right away for real work. Light computing might change how we build machines that think and learn in the future. It offers a way past the roadblocks that have slowed down progress in making faster computers.
This new processor can do 65.5 trillion math operations every second but only needs 78 watts of electricity and 1.6 watts of light power. It runs normal AI models with almost perfect accuracy without changing them. Users can run big famous AI systems like ResNet for pictures, BERT for language, and DeepMind's game-playing programs straight out of the box. No one needs to retrain these models or tweak them for the light chip. This marks the first time anyone has sold a light-based AI chip that works with regular software right away.
The chip works differently because it uses actual light to do calculations instead of just electricity. This helps solve problems with next-generation graphics cards that cost tons of money and use huge amounts of power. The company made sure their chip works with popular AI software like PyTorch and TensorFlow. They hope companies will start using their chips right away for real work. Light computing might change how we build machines that think and learn in the future. It offers a way past the roadblocks that have slowed down progress in making faster computers.
