NeuralBrew

Identify novel AI algorithms with potential for high energy efficiency and performance with logarithmic processing.

Engage with top developers to produce Python floating point scaled-down models to train using CPUs and GPUs.

Convert models to work with logarithmic number system. Synthesize code to run on FPGAs. Produce preliminary estimates for energy and performance.

Scale up models and emulate full size chips. Training targeting specific applications. Produce application targeted weights and parameters. Full chip energy estimation.

Optimize inference chip designs on emulation platform. Prune networks, integrate hard-coded weights and parameters. Add energy management features. Perform design verification.

License IP cores. Produce application specific inference chips for data centers and On-Device AI chips for embedded applications. License software for training.