Optimised speech inference for the new Intel FPGA PAC D5005
The figures were calculated using Myrtle’s AI solution running on the Intel’s new D5005 FPGA Programmable Acceleration Card (Intel FPGA PAC). The results come from the collaboration between Intel and Myrtle to optimise a recurrent neural network (RNN) for speech inference on the Intel FPGA PAC D5005. The results include running over four thousand voice channels concurrently on one FPGA, which brings a six-fold improvement in performance per watt over with general-purpose GPUs with a latency of one-thirtieth that of a GPU.
Myrtle owns the MLPerf speech transcription workload and has open-sourced its code to help the industry benchmark new edge and data centre hardware more consistently.
More information
Related news
Intel adds new FPGA programmable acceleration card
SK Telecom chooses Xilinx FPGAs for AI acceleration
Intel’s hardware+software platform to ease deployment of FPGA acceleration
Development Board accelerates machine learning designs