We looked at Syntiant earlier this year, and the ogre was stirring, but they are worth another look as they claim 1-bit and up operation for their Syntiant Core 2 tensor processor. They have also had a fair suck of the sauce bottle.
Many of their patents relate to the computerized design of analog or digital circuits, voice and image recognition.
They started with analog NNs until 2017, then progressed through hybrid analog/digital NNs:
US2019034790A1 Systems And Methods For Partial Digital Retraining
https://worldwide.espacenet.com/patent/search/family/065138355/publication/US2019034790A1?q=US2019034790A1
and sort of referred to a full dgital nn by 2018:
WO2020028693A1 SENSOR-PROCESSING SYSTEMS INCLUDING NEUROMORPHIC PROCESSING MODULES AND METHODS THEREOF
https://worldwide.espacenet.com/patent/search/family/069229878/publication/WO2020028693A1?q=WO2020028693A1
but they don't describe the circuit layout for individual neuromorphic processing units.
RZV-SYNTIANT-MVVSYNTIANT Multimodal (Vision+Voice) AI Solution
https://www.renesas.com/us/en/document/prb/syntiant-multimodal-visionvoice-ai-solution-renesas-rz-partner-ecosystem?language=en&r=1523096
Product Brief
https://www.renesas.com/us/en/products/microcontrollers-microprocessors/rz-arm-based-high-end-32-64-bit-mpus/rz-partner-solutions/syntiant-multimodal-ai-solution#document
Advanced voice and image processing capabilities at the edge are provided through a unique
combination of low-power, multi-modal, multi-feature AI inference capabilities
This solution combines Renesas’ RZ/V2M vision AI microprocessor unit (MPU) and Syntiant’s
NDP120 Neural Decision Processor™
############################
Syntiant Introduces Second Generation NDP120 Deep Learning Processor for Audio and Sensor Apps
https://www.syntiant.com/post/syntiant-introduces-second-generation-ndp120-deep-learning-processor-for-audio-and-sensor-apps
“The NDP120 is the first of a family of semiconductors using our next generation Syntiant Core 2 tensor processor platform that brings performance levels previously found in plugged-in devices to a power level suitable to run on batteries,” said Kurt Busch, CEO of Syntiant. “We took years of real world, low-power edge deep learning experience to develop this architecture into a scalable design optimized to bring neural processing to power constrained deployments.”
A
n Easy-to-Use Powerful Neural Network
The Syntiant Core 2 moves larger neural networks into always-on domains with capacity to generate shared embeddings, run ensembles, and other neural architectures concurrently or in cascades. The Syntiant Core 2 is a tensor processing core built from the ground up to support energy efficient inference without compromising ease of programming. This second generation architecture delivers 25x the tensor throughput of the Syntiant Core 1™ found in the Syntiant® NDP100™ and Syntiant® NDP101™ devices that are currently shipping in high volumes.
Graph Native:
The Syntiant Core 2 is built around a highly optimized tensor-based memory and processing system designed to avoid inefficiencies in stored program architectures. Each layer independently controls its parameter, input, and output tensors consistent with graph-based execution, enabling neural designers with full control of multiple concurrent independent networks, or swapping network configurations depending on operating conditions.
Framework Support:
The Syntiant Core 2 training development kit includes tools for running bit-exact simulations directly within high-level modeling languages, such as Tensorflow and Keras. All major frameworks can port to the Syntiant Core 2 runtime, including native support for multiple types of convolutional kernels, kernel striding, kernel dilation, downsampling layers, fully connected layers, pointwise operations, and a variety of activations.
Compression When You Want It:
The evolving practice of neural compression is natively supported by the Syntiant Core 2. Network architectures can mix 1-, 2-, 4-, and 8-bit weights and higher precision bias terms. For the most challenging tasks, the Syntiant Core 2 supports high precision modes, including 16-bit inputs and outputs. Quantize inference when wanted, but not just because the edge processor requires it.
Secret Sauce:
The Syntiant Core 2 supports a variety of task-dependent under the hood optimizations for sparsity and time series that speed up inference and reduce power requirements without extensive post-training network optimization. With design tools supplied by Syntiant, it is possible to explore the complete power, latency, memory, and parameter requirements on the Syntiant Core 2 interactively, and perform large scale hardware-aware hyperparameter searches with the same tools. The Syntiant Core 2 can shorten time-to-product by months or years as compared to more constrained and power-intensive solutions. “Syntiant’s highly configurable deep neural network provides enormous flexibility for low-power audio applications and feature extraction,” said Clark Peng, vice president, head of product management at MSI, a world leader in gaming. “The NDP120’s programmable DSP combined with its highly accurate inference engine are ideal for creating high performance voice command applications that can run across both traditional and machine learning algorithms.”
This article provides an english translation:
https://www.enterpriseai.news/2021/01/08/syntiant-aims-new-ai-chip-at-always-on-edge/ Syntiant Aims New AI Chip at ‘Always-On’ Edge January 8, 2021 by George Leopold
(20min delay)