Has this been shared? I haven't come across these guys before but we get a small mention
Giant.AI, Inc.
www.giant.ai
US 11478927 B1 (not BrainChips patent)
Hybrid Computing Architectures With Specialized Processors To Encode/decode Latent Representations For Controlling Dynamic Mechanical Systems
(109) Based on the matching, a hardware machine-learning accelerator may be configured to execute operations of a machine learning model upon inputs received from one or more sensors or encoders. For example, some embodiments of robots and other controlled dynamic mechanical systems described herein may include a plurality of sensors of a modular system hardware design such that each sensor (or a grouping of sensors) is coupled (directly, in some examples) with special-purpose chipsets for performing a space (e.g., like a sub-space or latent-space) or other encoding of sensor data prior to downstream digestion by a higher-level component or model of the system. Moreover, one or more intermediate or downstream models, like various models for encoding inputs, may operate on those encoded outputs to combine sub-spaces into broader representations (which is not to suggest that the broader representation need be of higher dimensionality or size, but rather that it accounts for more properties in aggregate that are reported by sensors of the sensor layer). One or more of the upstream, intermediate (or downstream) encoders may be implemented within one or more hardware ML Accelerators like, but not limited to, Movidius chips, tensorflow edge compute devices, Nvidia Drive PX and Jetson TX1/TX2 Module, Intel Nervana processors, Mobileye EyeQ processors, Habana processors, Qualcomm's Cloud AI100 processors and SoC AI engines, IBM's TrueNorth processors, NXP's S32V234 and S32 chips, AWS Inferentia chips, Microsoft Brainwaive chips, Apple's Neural Engine, ARM's Project Trillium based processors, Cerebras's processors, Graphcore processors, PEZY Computing processors, Tenstorrent processors, Blaize processors, Adapteva processors, Mythic processors, Kalray's Massively Parallel Processor Array,
BrainChip's spiking neural network processors, Almotiv's neural network acceleration core, Hailo-8 processors, and various neural network processing units from other vendors. Different ones of these ML Accelerators may be used to implement different ones of the aforementioned models upon sensor data (or upstream encoder output data), such as based on matching of model performance on a given accelerator for given sensor output.