» Current | 2020 | 2019 | 2018 | Subscribe

Linley Newsletter

Flex Logix Accelerates DSP

April 28, 2020

Author: Mike Demler

Recently, Flex Logix customers recognized the common mathematical foundation shared by CNNs and DSPs, leading the embedded-FPGA developer to evaluate the performance of its NNMax deep-learning accelerators (DLAs) on finite-impulse-response (FIR) filters. Flex developed NNMax as licensable intellectual property (IP), which it offers as a tile comprising systolic multiply-accumulate (MAC) arrays, memory blocks, programmable logic, and the company’s XFLX configurable interconnect.

MAC operations typically consume 90% or more of the cycles in a CNN, but FIR filters employ the same operations. In a CNN, each artificial neuron multiplies its inputs by a set of weights (also called a filter), adds the result, and feeds the sum forward as an input to the next layer. FIR filters are similar, comprising a series of MAC operations called filter taps. Each stage in the filter multiplies the input signal by a filter coefficient, combines the result with the output of the previous filter tap, and sends the sum to the next tap.

Flex Logix plans to develop a DSP compiler that maps Matlab output to NNMax. The DLA’s ability to serve as both a DLA and a DSP accelerator will be attractive to some customers, and the initial performance looks promising. But for NNMax to be useful beyond FIR filters, the company must develop a more comprehensive function library, which will likely require new architectural features as well.      

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »

Events

Linley Fall Processor Conference 2020
October 20-22 and 27-29, 2020 (All Times Pacific)
Register Now!
More Events »