MaxLinear’s augmented neuron technology mathematically augments DNNs reducing the amount of computation required to achieve a target accuracy. This leads to reduced memory usage, lower latency, and increased throughput. We provide optimized integration with TensorFlow and Intel’s OpenVino and OneDNN libraries, enabling enhanced AI using industry-standard workflows.
MaxLinear strives to improve the world’s communication networks for everyone by connecting people through our highly integrated radio-frequency (RF), analog, digital, and mixed-signal semiconductor solutions and licensable machine learning IP for access and connectivity, wired and wireless infrastructure, and industrial and multi-market applications.
MaxLinear’s Augmented Neuron technology enables small DNN models to achieve high accuracy, leading to AI solutions with lower computational and memory cost. Augmented Neurons are provided via a new set of convolutional and dense layers that are fully-compatible, drop-in replacements for standard DNN layers.
Augmented Neurons are delivered via TensorFlow and OpenVino plugins, easing integration to customer’s existing workflows. These plugins include implementations optimized for Intel’s CPUs supporting SSE, AVX-2, and AVX-512 instruction sets.
The complexity of neural networks (NNs) is increasing rapidly. With slowdown in Moore’s law, chipset capacity is unable to keep up with compute requirements to run the best NNs. While the world is focusing on optimizations at NN structural level, enhancements at the core mathematical level provides a new avenue to improve accuracy and reduce compute requirements of NNs. This white paper provides details of MaxLinear’s Augmented Neuron technology developed by mathematicians and how it cuts down compute needs by up to 10X while maintaining the same accuracy
In this post we describe how MaxLinear’s Augmented Neuron technology achieves 2x or greater computational and memory cost reduction with respect to the standard ResNet50 model. These benchmark results highlight gains from MaxLinear’s technology alone. Pruning, quantization, and other optimizations can be combined for further performance improvements.