Not sure if this has been posted before so apologies if it has.
It doesn't really add too much to what we already know but it is finally, a technical article on Brainchip that appears well written and researched. One of the better ones I have come across. They appear to understand our potential. I even like the fact that they see the licencing deal with Megachips as more of a clear win for Megachips as opposed to us, the newbie minnows.
No hint of snarkiness to be seen, which is also a pleasant change. Hopefully a sign of some long overdue respect starting to come out way. (At least from those that understand what a neuromorphic chip is)
https://www.allaboutcircuits.com/news/brainchip-and-megechips-leverage-neuromorphic-based-processor-ip-for-edge-based-ai-solutions/
Moving away from Cloud-based AI to the network Edge reduces power and bandwidth demands. Akida, the neuromorphic Edge AI processor from BrainChip might help to facilitate that move.
Application-specific integrated circuit (ASIC) Intellectual Property (IP) core development is a complex design, verification, licensing, and deployment workflow. Engineering teams looking to bring a solution to market are often faced with the challenge of deciding how to manage the entire product lifecycle, including design, manufacture, and supply chain.
Last week BrainChip, a developer of a “next-generation” Edge-based AI IP core called Akida, released details on a licensing agreement with ASIC developer MegaChips.
The Akida neural system on chip (NSoC). Image used courtesy ofBrainChip
Deployment of AI services at the network Edge is seen as one of the most pressing industry opportunities in 2021 and beyond. Naturally, before deployment can occur, several hurdles need to be overcome before Edge AI can become as ubiquitous as Cloud AI.
Today, let’s explore some of the challenges relating to Edge AI. Then, look at what the Akida core is said to offer and why MegaChips has signed BrainChip as an IP partner.
On Bringing AI Services to the Network Edge
Traditionally, almost all modern AI services have followed a single deployment model from the user equipment to the cloud for processing and return. This mode is referred to as direct-to-cloud (D2C).
However, there are other potential edge-based modes (or offloading strategies) for AI services, including:
- Device-to-edge
- Device-to-device
- Hybrid architecture (Cloud/Edge), and
- Caching (at the Edge)
Moving part or all of the data collection, algorithms, and computation from cloud-based systems and closer to user equipment could reduce power requirements and overall network traffic.
Below is a current model of D2C, where all modeling is completed in the cloud, and the future where disaggregated workloads can be completed at the Edge of the network.
The concept for Cloud-based AI services and Edge-based AI services. Screenshot used courtesy ofXu et al
Effective offloading is one of four areas of research that may lead to effective solutions for Edge services. Three other areas complicating Edge technology deployment include caching of data, AI training, and inferences.
Four “pillars” of Edge intelligence represent significant challenges to moving away from the cloud. Screenshot used courtesy ofXu et al
Possibly the most pressing challenge related to all four areas is the limited processing capability at the Edge when compared to cloud servers. This issue becomes more pronounced when considering the finite battery capacities of mobile user equipment.
Moving from challenge to "solution": BrainChip's Akida IP core might address one of the core challenges to Edge deployment; the need for AI training at the Edge. That is, Akida is said not to require a retraining process to learn new things.
Overviewing the Akida Event-domain Neural Processor
TheAkida event-domain neural processoris said to be the first commercially available neuromorphic AI chip.
Neuromorphic computing is a field of study based on the brain and aims to model computing cores after biological processes using convolutional neural networks (CNN) or spiking neural networks.
General neural processing set up vs Akida set up. Screenshot used courtesy ofBrainChip
Akida is said to be capable of "instantaneous learning" adopted through a process called time-dependent spike plasticity (STDP) and the cores' inherent event-based operation.
Between STDP and event-based processing, Akida claims to perform incremental learning. This capability would significantly reduce the challenge of AI retraining at the Edge, especially in isolated environments where data input to the system is sporadic.
On Licensing Akida to MegaChips
Presuming the stated advanced capabilities of Akida, MegaChips' procurement of a licensing agreement for the IP is a clear win for the ASIC developer.
As a provider of turnkey and COT (customer-owned tooling) services, MegaChipsprovides access to an extensive ecosystemof IP ranging from in-house development to validated third-party IP partners.
The Akida IP core features a fully synthesizable
level (RTL) and an IP deliverable that includes complete test benches and simulation results, RTL scripts and timing constraints, and customized IP packages.
Highly scalable, the neural mesh can be configured in a parallel format for high-performance applications or optimized for space-constrained applications.
Concept architecture for the Akida neural fabric. Image used courtesy ofBrainChip
The engineering challenges needed to move towards a decentralized compute deployment have not been overcome today, but this neuromorphic event-based processor IP might be a step towards that goal.
The bottom-line benefit of this licensing deal is the strengthened industry position for both companies.
Rob Telson, BrainChip's VP of Worldwide Sales and Marketing believes that this partnership could better further both company’s missions by pushing the boundaries of technology and offering unprecedented products.
All in all, it will be interesting to see where this venture heads next, and what other technology will come out to keep the momentum going for processing and AI at the Edge.
TT
(20min delay)