The high computational requirements of ML applications pose a challenge to their deployment, especially when targeting resource-constrained devices.
To address it, devised a new hardware solution supporting the high degree of parallelism offered by ML algorithms and leveraging their robustness towards low-range data representation. The pipeline is optimized for parallel, small-bitwidth arithmetics, using flexible Single Instruction Multiple Data (SIMD) formats. Its small area footprint allows its integration at the periphery of memory resources.
Related Publications
An Energy Efficient Soft SIMD Microarchitecture and Its Application on Quantized CNNs | |||
Yu, Pengbo; Ponzina, Flavio; Levisse, Alexandre Sébastien Julien; Gupta Mohit ; Biswas Dwaipayan ; Ansaloni, Giovanni; Atienza Alonso, David; Catthoor Francky | |||
2024-03-05 | IEEE Transactions on Very Large Scale Integration Systems |