This article is written by Alberto Romero, an analyst at Cambrian-AI, and Karl Freund, founder of Cambrian-AI.
We recently tweeted about the startup GrAI Matter Labs (GML) and won a lot of questions about the company’s products and strategy. As one of the first startups to launch a neuromorphic AI platform for next-generation AI, the company deserves a little more attention, so let’s take a closer look.
GML is an AI hardware startup that focuses on Edge AI with near real-time computing. Founded in 2016 through an organization of experts in silicon design and neuromorphic computing, GML believes it will revolutionize inference at the terminal device level, focusing first and foremost on audio. and video processing with very low latencies. By processing knowledge closer to its source, AI algorithms can provide near-instantaneous insights and transformation without incurring the highest latencies and prices typical of cloud servers. GML’s “out-of-the-box” AI provides responses that had such a low load and power. After a demo, we were amazed at the quality and instant latencies they were able to produce.
GML is recently focusing on commercial robotics and drones for a near-sensor understanding, representing a market of around 50 million devices by 2025. Now, the company’s plans will expand its success to come with transforming high-fidelity data into cellular and customer gadgets, a market that, according to the company’s estimates, is 20 times larger with more than a billion gadgets by 2025.
IoT devices are proliferating in smart security cameras on the streets, robotic arms in factories, voice assistants in our homes, and smartphones in our pockets. All of those devices have sensors that capture knowledge. Most corporations that apply AI at the edge of the network Focus on understanding or categorizing this information to enable predictions. GML is literally transforming the AV user experience on the go. To achieve this, they combine 4 technological pillars: high-precision processing (16-bit floating point) to deliver quality content, dynamic knowledge flow to exploit knowledge-dependent parsimony, neuromorphic design for efficiency, and in-memory computing to decrease force input and latency. In short: 1/10 of the reaction time to 1/10 of the force.
GML’s pricing proposition, therefore, is based on those pillars that, combined, create a unique and differentiated solution: low-latency, energy-efficient AI-powered terminal computing for raw knowledge in real-time high-fidelity consumable content, enabling instant applicability. situations
Power constraints at the edge of the grid force terminal AI devices to low power consumption. GML’s state-of-the-art solution produces high-fidelity content through the exploitation of parsimony, the fact that audio and video content is not replaced everywhere, or at the same time, with wonderful precision.
A prototypical example to illustrate the benefits of this technique is a smart security camera. The recorded background remains largely consistent with the day, so it doesn’t provide any new information. By processing and analyzing only people, cars and other moving objects, energy savings and latency discounts can be up to 95%.
GML’s next hardware concept, GML VIP (not yet available for production) is a SoC (System on Chip) that integrates a neural engine, GrAICore, with the required features for low power consumption, ultra-low latency, and high final accuracy. -point inference process.
GrAICore uses neuronFlow technology inspired by the brain. In addition to parsimonious processing, NeuronFlow is based on the paradigm of knowledge architecture, which allows for effective and detailed parallelization. With in-memory computing, which reduces bottlenecks in functionality caused by the transfer of knowledge between memory and the processor, those functions speed up calculations in various orders. of magnitude
The full VIP stack is complemented through the GrAIFlow SDK, compatible with the same old ML frameworks, TensorFlow and PyTorch, to bring traditional templates into effect. It also provides a library of ready-to-deploy templates. Custom and pre-trained models can be optimized and compiled with the ML toolkit to be deployed for peripheral device-level inference with the newest component, GrAIFlow Run-Time Ready.
GML targets the fast-growing market of more than $1 billion (more than 20% consistent with the year) for AI terminals with an exclusive technique subsidized through state-of-the-art technology. They outperform the terminal competition by focusing on floating 16-bit high-fidelity signaling real-time “content transformation” rather than simply “understanding” (categorizing) that uses 8-bit computing.
According to the company, the 4 pillars combine to surpass NVIDIA’s next-generation platform, Jetson Nano, 10 times, with a decrease of > 10 times for Resnet50. However, we note that Jetson Nano is a state of the art. platform, while the GML platform focuses very well on a few tasks.
GML has the potential to revolutionize commercial and customer AV reporting with everyday high-fidelity devices while meeting the stringent force and load requirements to handle device content. -Come take advantage.
Disclosures: This article expresses the authors’ reviews and should not be considered as a recommendation to buy or invest in the corporations discussed. D-Matrix, Esperanto, Graphcore, GML, IBM, Intel, Mythic, NVIDIA, Qualcomm Technologies, Si-Five, Synopsys and Tentorrent. We have no investment position in any of the corporations discussed in this article and do not plan to start in the near future. For more information, visit our online page at https://cambrian-AI. com.