Types of Deep Learning Hardware - Interview with Bradley Geden of Synopsys
Artificial intelligence based on “deep learning” is rapidly being implemented in a wide variety of edge systems, mostly using commercial chips that are one of several types. Ed Sperling, Editor of Semiconductor Engineering, interviewed Bradley Geden of Synopsys about these different types, and the various applications for each type.
These include systolic arrays, 2D course-grained reconfigurable arrays, and parallel pipelines. In each case, there is a 2D neural network array of artificial neurons, with matrix operations (multiply-accumulates or MACs) on the values in the neurons. These are all digital computational arrays, in contrast to analog arrays of memristors that are under development elsewhere.
Systolic arrays are hard-wired for MAC operations, and are used for basic image recognition neural networks. The reconfigurable arrays are more like FPGAs, with greater flexibility for a wider range of algorithms, but are more complex to program. Parallel pipelines are optimized for high-speed throughput, involving complex calculations on real-time data.
Mr. Geden also emphasized design synthesis approaches for programming these chips. Given the repetitive nature of these structures, a hierarchical approach may often be more flexible and easier to alter than a “flat” approach. Further information from Synopsys on design tools for AI systems is available at the Synopsys website.