MAPLE: A Massively Parallel Digital Learning Processor
T67
H. Graf, S. Cadambi, I. Durdanovic, V. Jakkula, M. Sankaradass, E. Cosatto, S. Chakradhar; NEC Laboratories America

Architecture scales to thousands of cores · Parallel Vector Processing Elements (VPE) · Parallel data I/O · Parallel memory banks

128 VPE on one FPGA chip; 6 memory banks

Comparison of compute speed
50x 40x 30x

(relative to CPU)

Compare with CPU:
· 19x ­ 30x higher performance

20x 10x

· 7x reduced power dissipation · FPGA: Configurable architecture