This invention describes a specialized computer chip designed for artificial intelligence calculations. It uses a grid of tiny electronic components, where each component stores a single numerical value, like a "weight" in a neural network. The key idea is that each component has multiple ways to read that single stored weight at the same time, allowing it to perform several calculations simultaneously using different inputs. The claims narrow this to a component that multiplies an input by its stored weight and then adds another incoming value to that product, sending the final result out.
Why it matters: Filed before the widespread adoption of large language models, this hardware design could offer energy-efficient compute-in-memory for AI. The demand for specialized accelerators for matrix multiplication and accumulation has significantly increased since 2018, making novel architectures like this more relevant for scaling AI inference and training.
AI gives you a few directions you could take this. Pick one, and we check whether your version is different enough to patent, then write the filing.
Reinvent this with AI