Brain-inspired chips promise ultra-efficient AI, so why aren’t they everywhere?
- Reference: 1662978188
- News link: https://www.theregister.co.uk/2022/09/12/brain_chips_interview/
- Source link:
Intuitively, the idea makes a lot of sense. Our brains are pretty good at making sense of the world, so why wouldn't a chip designed to work like them be good at it too?
Yet, after years of development and the backing of massive tech companies like IBM and Intel, these brain-like chips are still years away from making their way into consumer products.
[1]
That hasn't stopped the tech from grabbing headlines over the years. Neuromorphic chips up to 16 times [2]more efficient ; brain-like chips potentially [3]powering future supercomputers; Samsung wanting to [4]reverse engineer the brain; IBM [5]recreating a frog brain in silicon. You get the idea.
[6]
[7]
While the chips show promise, the reality is the field of neuromorphics is still in a very experimental stage, and faces many challenges that must be resolved before they are ready for prime time, explains Karl Freund, principal analyst at Cambrian AI Research, in an interview with The Register .
This may be one of the reasons many of the more promising neuromorphic processors have seemingly stalled.
[8]
IBM, for example, hasn't given an update on its True North neuromorphic chips, which are capable of simulating more than a million neurons, in more than four years. SpiNNaker, another promising spiking neural networking processor, received an €8 million (c $8.15 million) grant in 2019 to develop a second-gen chip based on the design. However, the company behind the chip, Dresden, Germany-based SpiNNcloud, is only now getting off the ground.
Intel's Loihi and Loihi 2 processors have come the closest to a commercial launch in so far as Intel has made development boards available to outside researchers alongside its Lava software development kit.
The Department of Energy's Sandia National Laboratories, for example, is [9]exploring how these chips could be used to accelerate supercomputers. In a paper published in the journal Nature Electronics, researchers at Sandia demonstrated how Intel's Loihi chips "can solve more complex problems than those posed by artificial intelligence and may even earn a place in high-performance computing."
[10]
Yet, at least as of April, Intel has [11]no plans to productize its Loihi chips anytime soon.
What's the holdup?
So what gives? Why is it these chips, which show such promise in the lab, haven't matured faster given the insatiable demand for AI/ML?
According to Freund, one of the biggest problems comes down to funding.
"I tried to connect some venture capitalists in both neuromorphic and analog [computing] and a fairly consistent response even before the current capital crunch was 'we don't invest in research'," he says. "Their take is pretty much the same as mine, which is most of the technologies, perhaps all, are still in the research phase."
As a result, progress in productizing neuromorphic computing has been limited to large companies with deep R&D budgets, he said.
But it's not just funding that's getting in the way. Freund argues the scope of the problem for neuromorphics has only gotten larger as the tech has grown more mature.
[12]Neuromorphic chips 'up to 16 times more energy efficient' for deep learning
[13]Intel's neurochips could one day end up in PCs or a cloud service
[14]Brain-like neurochips good for supercomputers, not just AI, says Sandia
[15]Intel offers Loihi 2 to boffins: A 7nm chip with more than 1m programmable neurons
With the first neuromorphic test chips, scientists were primarily focused on getting to a point where they could do useful work, he explains.
However, productizing such a chip means solving other problems, like how you get data in and out of the chip effectively.
This isn't a problem unique to neuromorphics by any means. It's one associated with quantum computing and even traditional accelerators, which have accumulated bottlenecks in recent generations due to the speed at which the data can be pre-and post-processed and/or ingested and egressed from the chip, Freund explained.
Finally, there's the issue of developing software that can take advantage of these accelerators.
"It's really going to take a whole community of researchers to solve the programmability problem of neuromorphic computing," Freund says.
Traditional accelerators are good enough
Perhaps the biggest reason that neuromorphic computers haven't taken over is that traditional accelerators are simply getting more powerful and more efficient quickly enough.
"What they're finding is that platforms, like Nvidia Jetson Orin or some new novel platforms from startups, are solving the problem really quickly. So the need to do something super exotic is continuing to lessen as the state of the art in existing technologies evolves," Freund says. "If you look at what Qualcomm has done with their AI engine, you're talking milliwatts… and what it does when you take a photograph is astounding."
As a result, meaningful problems can be resolved in the power envelope required by existing digital technologies.
While neuromorphics may not be ready to replace traditional accelerators anytime soon, Freund believes the technology will eventually reach the mainstream.
"These things do take time to mature," he says, citing the rise of Arm processors in the datacenter as something that took more than 10 years to achieve. "And that was for CPUs; CPUs are easy compared to things like quantum and neuromorphic computing." ®
Get our [16]Tech Resources
[1] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Yx9Xls39oenbQ5QKIjsq@QAAABY&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[2] https://www.theregister.com/2022/05/24/neuromorphic_chips_up_to_16/
[3] https://www.theregister.com/2022/03/10/neuromorphic_chips_sandia/
[4] https://www.theregister.com/2021/09/27/samsung_planning_a_silicon_brain/
[5] https://www.theregister.com/2014/08/07/ibm_synapse_chip/
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Yx9Xls39oenbQ5QKIjsq@QAAABY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Yx9Xls39oenbQ5QKIjsq@QAAABY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Yx9Xls39oenbQ5QKIjsq@QAAABY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[9] https://www.theregister.com/2022/03/10/neuromorphic_chips_sandia/
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Yx9Xls39oenbQ5QKIjsq@QAAABY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://www.theregister.com/2022/04/15/intels_neurochips_could_one_day/
[12] https://www.theregister.com/2022/05/24/neuromorphic_chips_up_to_16/
[13] https://www.theregister.com/2022/04/15/intels_neurochips_could_one_day/
[14] https://www.theregister.com/2022/03/10/neuromorphic_chips_sandia/
[15] https://www.theregister.com/2021/10/01/artificial_brain_intel/
[16] https://whitepapers.theregister.com/
Re: Thanks for the update
Seconded! I always find neuromorphic chips very interesting, and I often wonder why we don't see them in actual use.
Re: Thanks for the update
One thing puzzles me - should I understand "neuromorphic chip" as in spiking neurons, as opposed to, say, a chip which implements e.g., a sigmoid transfer function close to the metal (as well as large-scale connectivity)? And if so, why? it may be Nature's Way, but that does not necessarily imply it will be the "best" way to implement ML/AI technology on silicon - especially seeing as (i) we do not have a deep understanding of how biological neurons achieve functionality through spiking, and (ii) most large-scale neural models are not based on a spiking design and implementing learning/training with spiking neurons is harder (and slower!)
The mid-end is interesting too
The situation at the low- to mid-end of the spectrum looks more dynamic. Looking at companies like Brainchip, GrAI, Innatera, ... one might get the feeling that neuromorphic could got the RISC-V way i.e. start small (e.g. embedded computer vision) and make its way towards more high-performance applications.
Some point to automotive being one of the higher potential markets (likely due to the increase in cameras on L2/L3 vehicles) but that remains to be seen (see the point made in the article about current architectures developing fast enough in terms of performance and performance/watt).
IoT/the industrial sector might be a lower hanging fruit. In this sense one would bypass the problem of programmability, as neuromorphic chips hit the market as part of a final product. As an example one can look at Prophesee.
I, for one, have given up on puns
Please tell me you lot have thought through the implications. You scared Stephen Hawking and he wasn't scared by black holes.
If you are going to train them then start them on the Iain M Banks Culture novels. Oppenheimer regretted too late.
Re: I, for one, have given up on puns
The only "Post scarcity" fiction I've seen that has a sensible explanation of how scarcity isn't deliberately maintained by the rich and powerful.
If they are inspired by the brain, how are you "programming them"
Because you don't program brains, you "teach" them.
Until they reach a critical mass and are able to learn for themselves.
Mines the one with the copy of Carver Meads "Analog VLSI Implementation of Neural Systems" in the pocket.
Re: If they are inspired by the brain, how are you "programming them"
Don't rush for your coat - you hit the nail on the head!
Real brains are "programmed" by tens of billions of years of evolution - programmed to learn very, very well indeed. We currently have no idea of the design principles behind that (beyond variations on backprop on multilayer, usually feed-forward networks).
Thanks for the update
I look at the state of spiking neural network hardware development every few months, you saved me an hour or two.