deepsilicon

We’re making hardware to run neural nets with orders of magnitude less latency, energy, and 5x less RAM. Right now, we’re proving it with software.

Deepsilicon announces a full-stack system to run transformer-based models on a single chip, including existing hardware, aiming to reduce latency, energy use, and RAM with a custom ASIC. They offer near-infinite chiplet customization, 5x memory reduction on Nvidia hardware, and a 50% YC-discount for training and deployment.

Step through deepsilicon’s homepage year by year — the pitch, the product and the design, as the web remembers them.

Each year opens the Wayback Machine capture closest to that date.