Startup launches near-binary neural network accelerator
Efficiera specifically targets inference processing of convoluted neural networks and can be implemented in FPGA or ASIC forms. The company claims its solution is “ultra-low power” due to its use of single- or two-bit quantization.
Typically processing is done on 8- or 16- or 32-bit data types depending the accuracy of result required. However, for neural networks results high resolution processing can be a waste of area and energy for the resolution required. Binary neural networks can be used to reduce the processing burden but at a loss of accuracy.
LeapMind claims that by using 1-bit weight coefficients and 2-bit activation values for intermediate data it is possible to maintain accuracy while achieving a significant reduction in the model area, thereby maximizing speed, power efficiency, and space efficiency.
The company has benchmarked an instantiation of its core as achieving occupying 0.422 square millimetres in a TSMC 12nm manufacturing process and achieving 6.55TOPS at a 800MHz clock frequency. This represents an efficiency of 27.7 TOPS/W.
The company is seeking applications for Efficiera in edge devices that are power- and cost-constrained such as household electrical goods, industrial machinery, construction equipment, surveillance cameras, and broadcasting equipment as well as miniature machinery and robots with limited heat dissipation capabilities.
Alongside the core available for license LeapMind is launching a software development kit to provide a dedicated learning and development environment. There is also the Efficiera Deep Learning Model and Efficiera Professional Services. This combination allows customers to build binary or near-binary deep learning models applicable to their own unique requirements.
LeapMind was founded in December 2012 as AddQuality and since its formation has raised 4.99 billion yen (about US$46.5 million).
The company plans to ship Efficiera later in 2020.
Related links and articles:
Edge startup offers legacy support and ML acceleration
GrAI Matter, Paris research gives rise to AI processor for the edge
Development environment eases machine learning on to microcontrollers
ResNet-50 – a misleading machine learning inference benchmark for megapixel images