[Hardware]: μBrain - Ultra-Low Power Neuromorphic Processing

[federicohyo]

Product Name

μBrain

Manufacturer / Organization

IMEC

Short Summary

μBrain is a clockless, event-driven chip that merges memory with computation to achieve sub-100 μW power consumption, making it ideal for always-on edge IoT applications without the battery drain of conventional hardware.

Product/Organization Website

https://www.imec-int.com/en/articles/imecs-snn-chip-combines-low-latency-energy-consumption-high-inference-accuracy

Chip Type

Digital

Development Status

Retired / End of Life

Release Year/Date

2021

Neuron Count

336

Synapse Count

37,366 (4 bits - 18.2kB)

Power Consumption

100 μW

Supported Software/SDK

No response

Detailed Description / Overview

μBrain is a highly efficient neuromorphic integrated circuit (IC) specifically tailored for edge AI IoT devices, operating on less than 100 μW of power. Although built with standard, low-cost digital technology, it breaks away from conventional digital neuromorphic designs by offering three major architectural advantages:

Asynchronous, Event-Driven Operation: It eliminates the global clock entirely. Instead, μBrain utilizes local, on-demand oscillators and an innovative delay-cell mechanism to process data dynamically as events occur.

Extreme Idle Efficiency: When there is no incoming data, the system consumes nothing but minimal leakage power. Even in this dormant state, it securely maintains its full internal state—including synaptic weights, network dynamics, and neuron membrane potentials.

Co-localized Memory and Computing: By merging computation and memory within the same physical space on the chip, μBrain removes the need for separate on-chip or off-chip memory blocks. This completely avoids the data-access bottlenecks and heavy energy overheads associated with traditional Von Neumann architectures.

Related Publications

Stuijt, J., Sifalakis, M., Yousefzadeh, A., & Corradi, F. (2021). μBrain: An event-driven and fully synthesizable architecture for spiking neural networks. Frontiers in neuroscience, 15, 664208.

[Nelias]

Where is the VHDL or Verilog for it? How is it in the "open" category? It seems like just an internal project without a solid proof of its existence or that it could be useful for the broader Neuromorphic community. I know about the μBrain since 2024 and that is the exact reason why I did not add it here but instead I did add SENeCA from the same imec because you at least can get the HDL SENeCA code through email. In general keep in mind that research papers are quite often shared to boast and 2021 μBrain is that kind of paper, while 2023 SENeCA isn't.

Science must be verifiable by others and if it isn't then it is not science but just a closed research done for marketing purposes, which should not belong to the Open Neuromorphic website in my opinion.

[federicohyo]

Hello, you are perfectly right. uBrain It is not an open source project, it is an IMEC proprietary IP, thus it is not part of this open category. I don’t think it will ever be shared (for free). I did not know that SENeCA HDL code can be obtained via email, it seems strange, as that is also an IMEC IP, so by public you mean you can get the code ? Without signing any NDA?

On a separate note, I am not sure I understand your comment of uBrain paper. You can learn a lot from that paper, even if reproducing it will require a lof of work (more than just code, as you will need to design a standard cell for implementing delays), but, in that paper, many concepts and ideas are very well described, as for example, the 3-phase oscillatory cell and the overall architecture. In addition, there are several works always by IMEC (private research) that exploit uBrain IP (e.g. https://pmc.ncbi.nlm.nih.gov/articles/PMC7614138/pdf/EMS163509.pdf) , so I’m not sure what you mean by “that kind of paper”.

In any case, since it is not open, and I only have proof of its existence on my desk at IMEC, I agree it does not belong to the OPEN category.

Thanks for your reply.
Federico

[Nelias]

@federicohyo SENeCA paper proposed for those who wanted to confirm it to contact them by email. Yh there are many such papers from China from USA too and ofc also from Europe. It does move the theoretical science indeed but not the engineering sadly. Maybe there could be some other category for all those proprietary hard to verify papers. It is just in my opinion not on the level of ReckOn or SNE or even Intel Loihi, because those labs that wanted to try out Loihi could do so when it was "easy" even with their NDA or contract. Hmmmm maybe uBrain should not be disqualified just like TSP1 was not but I think it might be easier to buy physical hardware TSP1 than it is to buy uBrain (like you said... papers, contracts, IPs, NDA). Same case with BrainChip Akida, it is not open at all but at least you can get it (they also sell IPs like imec).

If you will show me a link where for example I could easily obtain uBrain without deeper email sales contact and PO and contracts then it might be added here. Otherwise it might be just a free marketing for imec and that was not the purpose of Open Neuromorphic maybe idk. Its founders should look into it and post their opinion here.

Obivously there are "worse" papers that truly boast and you could not even get any IP at all so yh those could be imaginary or purely theoretical and if imec did make the uBrain then it is surely 2 levels higher in value.

[Nelias]

@federicohyo look into other hardware proposals, most of them do have a source code link. If they do not at least there is a path to buy it somehow. In the provided link: https://www.imec-int.com/en/articles/imecs-snn-chip-combines-low-latency-energy-consumption-high-inference-accuracy

I see no options to buy uBrain or obtain the IP so I think the path to obtain it is much harder and we can be honest: SOMEHOW HALF CLOSED NEUROMORPHIC except for that theoretical description in the paper. Could I reproduce that paper at home? I think I would have to send it to some fab, the reproduction would be impossible on FPGA unless maybe I would obtain that IP with an NDA.

[Nelias]

@federicohyo there are arguments for uBrain inclusion: on ONM there is closed IBM TrueNorth with a paper, closed NeuroGrid with a paper, closed Tianjic, very closed Innatera T1. You can open the uBrain pull request but include a link to purchase options or contact options or tell the people from imec to add those in their main page link about uBrain that you already included.

[neural-loop]

We dont limit inclusion of hardware based on openness or closed. What is important is that we will probably be one of the first results / authoritative sources that people find when researching the chip. We can prominently discuss in our articles how we evaluate their openness or lack thereof. If they require an NDA to get basic info, or block a spec sheet behind a 'sign up' form.. that is valid information to call attention to.