Achronix eFPGA Serves AI Two Ways

摘要

It's a classic chip-design conundrum. By the time you freeze your specification and then design, verify, tape out, and sample silicon, customer requirements have changed. To make matters worse, in the world of AI, neural-network models are evolving at lightning speed. If your deep-learning accelerator can't efficiently map the newest models, someone else's probably can. A great challenge in accelerator design is therefore to find the sweet spot between flexibility and efficiency. Presently, FPGAs occupy that middle ground between general-purpose processors and purpose-built accelerators.