Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/discussion-our-chips-can-never-achieve-the-true-computing-power-of-a-brain-by-having-same-specs-like-it.13959/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Discussion: Our chips can never achieve the true computing power of a brain by having same specs like it.

S

Saicharan1919

Guest
I always have the concept of building an exact brain-like hardware, in my mind. By which I mean, computer architectures inspired by Brain. The other thing which inspired me and pushed me to think in this direction is that the brain has large computation power but consumes small amounts of energy. Hence, I started understanding how exactly the brain works, especially the neurons inside the brain tissues.
In the process, I found that our brain is actually very sensitive in terms of packaging and processing info. In other words, when we compare it with the chips and its purpose, maintaining conditions like in a brain, we can't build a chip. Bare with me, you will understand this clearly in further sentences.
As we know, neuron is the one which gets more attention for all that computational power inside Brain, like a fundamental unit. But what actually happens in neuron is, that the electric impulses (or info) travel because of a potential difference developed between inner and outer body of the axion. This potential is mainly controlled by few voltage driven gates (though they are other factors of control too). The cut-off potential for these is in mV. This less power consumption (low potentials) is what makes brain more efficient. One thing to notice is that all those high computations are handled in the above discussed environment. Though we can't prove that how computation is happening in such environment, we can understand that the effect of this environment (like a supporting feature) is there on that high computational power of the brain.
In real-life chips/hardwares, maintaining such low potential environments are not possible I guess because it will be too sensitive to the external environment. We can also understand it in vice-versa. Our brain is sensitive and misbehave when provided any external direct electric impulses which is done by many neurologists to treat patients. Whereas, these are somewhat common in the environments where chips work.
Now, my opening point is - since we are appreciating the computation power of the brain by not considering the environment in it, hence we can't achieve the same level of computation power for our hardware.
I tried my best to put all my understanding in words. Please mention the places which lacks clarity, if any.
I am expecting a good understanding on this hence ready to learn new things and correct myself as well.
 
Thanks, Fred Chen. I was actually looking for the same thing.. But don't you think that achieving the brain-like computation with the same low power, is not possible until we dig deep into the material properties... Like we use the same silicon to build these spiking NN based hardware right?
 
Thanks, Fred Chen. I was actually looking for the same thing.. But don't you think that achieving the brain-like computation with the same low power, is not possible until we dig deep into the material properties... Like we use the same silicon to build these spiking NN based hardware right?
That's likely true. There were some CMOS demos but those were definitely very space-inefficient and would not be power-efficient if scaled to brain capacity.
 
@Saicharan1919 This is a great discussion and I have been delving into this for the last few days, so I'll share my thoughts. Neuromorphic computing (NC) is the field where people are trying to replicate biological intelligence with computers. Intel R&D has come up with a NC chip called Loihi, which is much more efficient than a CPU for running spiking neural networks. However, to reach the human brain scale Loihi is not enough because it is still based on CMOS logic which, as Fred mentioned, is very space inefficient.

The most promising direction for efficient implementation of NC is magnetic RAM (MRAM), specifically STT-MRAM. See this for example. The current problem is processing and economics of handling non-standard materials in a commercial fab. Tech demos exist, but so far, having invested a lot of money in NC, even Intel hasn't gone down the route of designing/making non CMOS implementations of NC chips. This is the critical step which, once executed, will pave the way for achieving the concept you described.
 
@Saicharan1919 This is a great discussion and I have been delving into this for the last few days, so I'll share my thoughts. Neuromorphic computing (NC) is the field where people are trying to replicate biological intelligence with computers. Intel R&D has come up with a NC chip called Loihi, which is much more efficient than a CPU for running spiking neural networks. However, to reach the human brain scale Loihi is not enough because it is still based on CMOS logic which, as Fred mentioned, is very space inefficient.

The most promising direction for efficient implementation of NC is magnetic RAM (MRAM), specifically STT-MRAM. See this for example. The current problem is processing and economics of handling non-standard materials in a commercial fab. Tech demos exist, but so far, having invested a lot of money in NC, even Intel hasn't gone down the route of designing/making non CMOS implementations of NC chips. This is the critical step which, once executed, will pave the way for achieving the concept you described.
As SRAM replacement, STT-MRAM has been frequently recommended. It will give better space efficiency, but not sure it can provide power efficiency yet (depends on the thermal stability you need). Actually memory-based computing has been proposed for a while. I guess too many designers still bound by von Neumann architecture.
 
@Saicharan1919 This is a great discussion and I have been delving into this for the last few days, so I'll share my thoughts. Neuromorphic computing (NC) is the field where people are trying to replicate biological intelligence with computers. Intel R&D has come up with a NC chip called Loihi, which is much more efficient than a CPU for running spiking neural networks. However, to reach the human brain scale Loihi is not enough because it is still based on CMOS logic which, as Fred mentioned, is very space inefficient.

The most promising direction for efficient implementation of NC is magnetic RAM (MRAM), specifically STT-MRAM. See this for example. The current problem is processing and economics of handling non-standard materials in a commercial fab. Tech demos exist, but so far, having invested a lot of money in NC, even Intel hasn't gone down the route of designing/making non CMOS implementations of NC chips. This is the critical step which, once executed, will pave the way for achieving the concept you described.
Thank you for showing interest @jaiyam. When I was researching, I also came across memristors.. But again as you told me, the processing is the bottleneck with our commercial fabs. Also, as specified by @Fred Chen, in-memory computation and near-memory computation are getting under the spotlight w.r.t the NC. Can I know both of your thoughts on memristors? This is stated as a fundamental element. Though, the signal is not constant as we increase levels in the design... we are losing the strength of the signal and hence we need to use amplifiers.
 
Humans are idiots (read the news if you doubt). Something that does a particular task much better than a human in a very narrow niche is interesting, and probably more profitable.

I worked on Automata at Micron. The wide parallel nature of memory for comparing states is very interesting. As a memory technology I thought it was very cool. It was implemented in real hardware.

Spiking NN is interesting. The issue I see is the connectivity between neurons. In the brain a single neuron is connected to hundreds of other neurons. Even if you can simulate the single neuron, the connections to adjacent neurons becomes a scaling issue for a physical implementation. Chips are stacks of 2D structures, if we simplify. The brain is a 3D interconnected mass.
You have picked a difficult problem to work on so it should be very interesting.
 
Thank you for showing interest @jaiyam. When I was researching, I also came across memristors.. But again as you told me, the processing is the bottleneck with our commercial fabs. Also, as specified by @Fred Chen, in-memory computation and near-memory computation are getting under the spotlight w.r.t the NC. Can I know both of your thoughts on memristors? This is stated as a fundamental element. Though, the signal is not constant as we increase levels in the design... we are losing the strength of the signal and hence we need to use amplifiers.
Relying on analog memory in combination with digital processing seems to be the direction taken by this approach. It may not achieve the brain-like scenario but is more targeted as an alternative to von Neumann architecture.
 
Not to mention mimicking a brain would have all the same issues that a brain has in the first place...like forgetfulness, loss of focus and a whole slew of other issues. Like the above poster said, humans are idiots! Mostly semantics at that point, but if you just want to make a chip like the human brain, it seems like a pointless endeavor.
 
I wonder why you think that mimicking brain is the correct way? In 1982 I had my first computer, a Comodore Vic-20; only 8 bits and only ~3500 bites of 8 bits. Compare that with today’s Apple a14 and iPhone 12pro max. I think that AI will be the next evolution of thinking and robots will be the next step in evolution that at a certain time later, it will create a much more better biological brain. Just my idea.
 
Our brains have a lot of ancient structures that under pressure of higher levels of brain, mainly the cortex. My only fear is that our engineers have never been able to create a bug free OS or application even at the earliest days of computing. A failure of computing error in the AI could end up as much much worst than the nightmare that I can ever imagine.
 
So I’m very happy that I am at 60 years of age and I hope that I will never ever will see that these will happen. But I’m very very seriously sorry for my children. What a pity for them.
 
@f4rewell1 Yes, AI will create the best brain... But to run that best brain, we need the best hardware resources. And that is what I am talking about. What we are discussing is the point that we can never achieve brain-like hardware with high computation power and low power consumption. If you have good hardware, you can run good software on it. We all know that the gap between the growth of AI and the growth of the respective processing hardware is way too long. We are trying to reduce the gap between them, hence the emerging brain-like hardware gets more attention and interest in the industry.
 
Fundamentally, the brain is not a computer in any conventional sense. Artificial neural networks are one thing, but trying to emulate the actual workings of the human brain on silicon seems misguided to me.
The brain is electrochemical, not purely electrical in nature, and the chemistry is extremely important. Think about how a person can become more productive with a little caffeine (or Adderall in a more extreme example) , how food releases dopamine that makes you happy, ect.
 
True...The brain is not a computer, but we are just wondering and taking steps towards studying how it is able to perform using low energy. That's the reason we are now focusing on other materials which allow us to reach such computational power by consuming low energy.
 
By the way human brain is very different than silicon chips. In the human brain there are only positive signals, no negative signals. For example when the brain does nothing it sands just a few signals; not at least e few megahertz when it is idle. And once again in the brain there are only one way positive signals. The signals goes only one way on the axons; and never goes a signal backwards on the same axon. Because sinapses between axons only passes signals in one way. This saves very much energy.
 
Back
Top