Memory Subsystem IP – DDR Controller, NoC, Artificial Intelligence, Deep Learning
AI Accelerator IP – ENLIGHT™
OPENEDGES™ Artificial Intelligence Compute Engine ENLIGHT™ is a deep learning accelerator IP technology delivers unrivaled compute density and energy efficiency. ENLIGHT™ NPU IP accelerates inferencing computation with excellent efficiency in terms of power/performance/area. It targets +10x compute density (performance/area) and +10x power efficiency (performance / power) vs comparable GPU.ENLIGHT™ optimize precision for common case and then deal with outliers somehow. The key observations are that most of the weights (& activations) can be nicely quantized in 4-bits and higher-precision outliers account for less than 3% of overall weights count.
ENLIGHT™ implementation an array of 4-bit precision MAC unit and it enables 4x compute density & low-power consumption.
Also our ORBIT™ high performance memory subsystem IP can perfectly meet high DRAM bandwidth requirement.
: Deliver unrivaled compute density and energy efficiency.
ENLIGHT™ Neural Network Compiler
: Optimize your neural network for perfect fit to ENLIGHT™ NPU.
ORBIT™ Memory Subsystem
: Pump exceptional memory performance into the ENLIGHT™ NPU and the rest of the system.