The software, known as KaNN for Kalray Neural Network, can be used for the development and evaluation of neural networks and is compatible with commonly-used deep learning frameworks such as GoogleNet, Squeezenet, CAFFE, the company said.
The development is aimed at such applications as autonomous cars, avionics, drones and robotics.
The software makes use of the on-chip memory of the MPPA2-256 288-core processor, known as Bostan, and spreads the data-dependent layers and weight parameters across the MPPA's cores. When running GoogleNet the MPPA can outperform the most efficient GPUs, Kalray said.
The chip includes 40Mbytes of on-chip memory and memory bandwidth of more than 1Tbyte/s. Each core is 5-issue VLIW with simple and double-precision floating-point operations, offering 1TFLOPS on-chip processing capability.
Kalray is planning for a second version of deep learning coming in 2018 that will offer a factor of 20 improvement in performance.
Related links and articles :