IBM's New Computer Chip is Pushing the LIMITS! 🔥
Вставка
- Опубліковано 23 чер 2024
- In this video I discuss New IBM's Analog chip for Artificial Intelligence with IBM Staff Researcher Manuel Le Gallo-Bourdeau: research.ibm.com/people/manue...
The Paper: www.nature.com/articles/s4192...
Thumbnail Image Credentials: IBM research.ibm.com/blog/analog-...
B-ROLLS Sources:
IBM • The future of computer...
Nature Electronics www.nature.com/articles/s4192...
Timestamps:
00:00 - The Problem
02:10 - New IBM Chip
03:31 - How In-Memory Computing Works
09:06 - How to run NN on the Analog Chip
14:10 - Will Analog Computers Happen?
14:47 - Training NN on Analog Chips
The Video To Watch Next:
New CPU Technology: • New CPU Technology jus...
👉 Support me at Patreon ➜ / anastasiintech
📩 Sign up for my Deep In Tech Newsletter for free! ➜ anastasiintech.substack.com
Let me know what you think!
I absolutely loved how you anthropomorphized computer memory with your finger gesture
I wonder when Laguerre polynomials / spherical harmonics (roughly speaking, describing atoms' shapes perfectly - whole periodic table is precisely this only because our universe is exactly 3D!) come into play. Physics' 101% aesotherics compared to this kind of computer science.
I think the video is great, but the title is grammatically incorrect. It should be "IBM's New AI Chip Explained". If you had left the 's off of IBM, it would have been ok, because then IBM would have been a descriptor of the type of chip along with New and AI. Throwing an 's on IBM made it a possessive proper noun, and therefore the subject. It would be like writing "new his shorts" instead of "his new shorts". Keep up the good work!
Doesn't a quantum computer store memory in QBits??? And then when you expose it to a gate (program) it collapses the wave function? So you have processing, and memory storage as a component, it you split the two processes up, and view them individually?? Is this correct? The gate being analogue, and the qbits being digital.. Hence the 0, and 1 nature of the qbits?
If this is right then it follows the same pattern of mixing analogue and digital. At the point of video 6:00 mixing digital with analogue.
In the video 2:44 - 2:56; This is the nature of the quantum computer; and the difference between the classic and quantum, is the quantum can executing millions of lines of code in one pass.
If all this is correct, then making a classical computer that is both analogue and digital, makes it more like a quantum computer. And the more this happens, the more it solves the memory/power problem:) I find that interesting:)
The more classical computers have an analogue, and a digital component, like a quantum computer (2:44 - 2:58 min), the more it solves the original problem see: 0:44 min to 1:05 min.
Thanks! @@Starlesslight
I really like you are inviting engineers behind those state of the art chips ! These people deserve recognition and they have lots of interesting details to share as Manuel did.
Thank you
Might even encourage them to practice their public speaking skills, we could use more of them fully engaged with society at large.
Totally agree, we can have a better understanding about how these chips work with the experts working on it.
He just seemed a little reserved and introverted, he explained what he was doing clearly though@@stinkymccheese8010
Melhor e usar quem já tem Problema Como Joe Biden mesmo 😃 o bobó bobó Mental 🤦🤪aí tu vai ver se o Chiaps Funciona mesmo 😃#
People at the bleeding edge of science and technology deserve more recognition than actors, singers and entertainers, and yet we don't hear about them in the media. Thank you for these fascinating videos.
At least, they are paid lavishly (hopefully) 😊
@@gennadiyleyfman6920getting to the bleeding edge isn't easy. Expect to spend 10 years getting a doctorate. I support developers at the bleeding edge, they are mathmaticians working on industrial control algorithms.
College level coaches get paid multiple fold more than professors at the same Colleges. As a civilization our priorities are backwards.
Amen. We live in a shallow ignorant world
But when you figure the average American has 108 you now you know why they can only stick to actors and sports fans because they can’t comprehend how deep this goes and that’s genetics. Can’t do anything about it right now anyways.
I am 65 year old female and know nothing about chips, beside the one in the supermarket🤣Still I watch your videos just to get an idea. You explain things in a fashion that is easy to understand, so thank you for that. I learned a lot from you.
🍟
Let me enlighten you, they are great with fish and mushy peas.
In my professional field, analog is almost considered a pejorative term. At some scales the use of analog active filter elements using op-amps in combination with digital potentiometers, peak detectors and SAR type ADC's can significantly reduce the computational overhead. As channels are added however the scale increases and the footprint with it. Seeing so much effort going into integrated analog/digital systems is very exciting for me, thanks for the great video.
It's low computation but high chip real estate. For example, the cosh function on the old ETANN chip used 4 op-amps, and took up so much space it had to be time multiplexed between rows of the matrix - which gives up a lot of the advantages of analog NN. I've been pondering how to do an acceptable sigmoid func with as low as 2 transistors (with trick gates that resist saturation) so that every matrix can afford to have its own realtime outputs.
Yes that's what I was thinking 😂
@@MattOGormanSmith Throughput though?
The usage of electronic networks instead of digital was used at Brunel University, UK by Stonham, Alexander and Wilkie when they made the WISARD system. This was a neural network that worked and was able to identify patterns on problems that they were non-linear which was an accusation by Marvin and Minsky in which paper they have stopped the developpement of NN. The Wisard system was able to bring back the NNs since then but after that they went all digital. It is good to see that IBM has revisited that idea.
20+ years ago Penrose was saying we'll need analog networks to get some real AI, and I agree.
(Well, he was saying quantum, but I think this will do)
@@artdehls9100 Also Brunel was 20y+ when i was a student there and already it was ancient history but its good to know that we do revisit good ideas.
You said it, "simple & elegant"! Fascinating to see analog chips pushing limits. Great video. 💯🤖
This analog technology reminds me of sideband transmission. 1 it uses a lot less power 2 single transmission can have many layers of information. Fascinating presentation Anastasia
Thank you so much for doing such videos and encouraging young chipmakers. Analog chip design was always a big question in my mind as a fresh graduate electronics engineer. Could you make an explanatory video for titles of a variety of chip design engineers and show what the industry is searching for? Should all the future chipmakers go on the digital design path?
Yes, Anastasia please make a video like that 🔬
Here's smth big for you. All engineering, all of it- is exactly like esotherics when opposed to maths, like physics+metaphysics (computer science's algos + physics, physical ontologies, models like Wolfram's, this sort of... well, rules of all. So one would probably want to know this sort of QMy math a bit and think what structure of compitation would fit both larger qorld and the world of atoms and molecules. Binary trees will definitely be there. Can you model real-life everythings with binary trees? This plus bits of structures single steps more complicated here and there
Yes that would be a good video definitely
Doing a comparison between chip maker’s career paths is a good idea
Yezzzz plsss
Honestly, it's scary to see how much money and effort goes into this at the moment, and how many overworked brillant engineers tinker away at their particular project, trying to find the next big thing that injects even more speed into everything AI. Breakthroughs are bound to happen and it's clear that we can expect BIG jumps in performance in terms of hardware, training methods, model architecture, algorithms and software as well as data. This is going to go through the roof. Ready or not, here it comes.
No, they R just taking existing tech like SSD & LCD & whatever & pretending it is something 'new' 2 jack their $cok price = $cam =P Kind of like calling a laundry list of pre-written statements 'A.I.' = no, it's just a 'fast database engine' LOL Like the so-called 'AI generated music' that is really just a compilation of 'royalty free clips' from elsewhere, or 'AI art' that is just combining other bits from other art & applying filters 2 it = can B impressive, but often it's just basic 'warp sharpening' effects =P There's a channel on UA-cam with a bunch of so-called 'AI' created videos that U can totally tell is $hit U can do with basic software like FFDShow running the same 'source' video through it. They just don't tell U where they got the original video they R tweaking out with the 'blob' filters. The face changing stuff 2 = U can do that on a fuking phone it's so E Z. No 'intelligence' 2 it = just a bit of clever SOFTWARE =D
Então e aí que tá o Sucesso de ver quem já tem dificuldade de similar Palavras isso já são coisas Nelrais uma função que está dando um Curtosirquito. Trás o Chips a Memória que não tá mais fazendo a Leitura de Palavras e fica se debatendo em memorizar #
A Lógica e Parar um Psicopatas Mentais e Racista que se Acha o Dono Do Mundo de se Impor na Vida das Pessoas e pior se Nem se quer conhece e sabe quem e . Só se sabe que e um Famoso por aí neste Mundo 🌎 qué com sua capacidade Mental teve esse Sucesso.que o Capacitou chegar há Onde chegou.o Resto e como um Labirinto que você mesmo tem que se achar um lugar de saída na Rede Social neste Método Convencional que temos em Mãos 💻🖥️📱#
Evolution tends to favor the most adaptive. The AI-human symbiosis co-evolution is inevitable. "Scary" is irrelevant . . . some will adapt accordingly, others maybe not, such is the consistent nature of evolution.
Doesn't photonicas have the same benefit of doing multiple calculations at once with low power? At 70% the speed of light no less.
Then further there are quantum computers that use photons as well?
All without some of the issues here?
I am really enjoying these interview and tech breakdown videos. You are truly inspirational, keep it up :)
A divided approach of training with digital chips & inference on analog would be the best way to get some immediate results. Thank you for the video Anastasia, keep it up.
Wow, this is amazing! I learned so much from this video. Thank you for sharing your insights and knowledge with us, Anastasiia. You are such an inspiration for anyone who wants to pursue a career in chip design. I’m looking forward to your next video. Keep up the great work!👍🤓
Thank you for always bringing us such great information.
Very informative content as always, keep it up!
Did you know IBM did 3nm node around 8 years ago?
IBM has some stuff that is not shared globally.
Just wow. So much quality info! Thank you 🎉
given the rush to AI if this tech can increase speeds at lower power consumption it will definitely be adopted the company that cracks the technical issues will own the AI space thanks for the video entertaining and informative as always.
Really interesting, thanks for sharing!
It's so nice to have these complex topics covered and explained in layman's terms. I have learned more about emerging tech and AI from this channel than any other. To be honest, I would watch Anastasi describing paint drying as long as she giggled occasionally! Surely, I'm not the only one?
First video watch impression, very good unpacking of information for those of us who are not chip engs, but still working in IT. I can see why this is exciting to talk about, a lot of potential and groundbreaking technology, will be interesting to watch where this goes over the next few years. Long time fan of IBM, good tech as long as you can afford it. Subscribed.
The "bottleneck" in current computers is the use of X86 derived architectures that while x86-64 somewhat improved some things, it's still not as effective as a properly designed new architecture like the CPUs that were coming out in the mid to late 1990s and early 2000s.
For example the Intel Itanium architecture would have eliminated a lot of these bottlenecks if it had been more widely used but you'd also need a properly engineered Operating System that isn't held back by companies with monopolistic practices and antitrust activities.
I knew this was the future even more than twenty years ago when i learned about digital computing. I love how tech goes parallel with my predictions.
Excellent show. Very informative and entertaining.
Thank you for your show I don't have time to keep up with the latest technology but your show keeps me connected thank you
Thank you !
When you mentioned a substance being heated and cooled, to go from solid, to (liquid?) to Crystalline... how does this happen on such a fast scale? Doesn' heating and forming crystals take SO much longer than just passing currents? That implies a physical change process, not just an electrical state change. If it works, it works, not denying that, but it seems odd that the two would be on a competitive time frame. This stuff is SO fascinating! Thank you Anastasi and all who put these posts together.
If you use heat exchange formulas with the scale size of the material it is incredibly fast. Heat, or energy, transfers almost instantaneously at the electron size. If we have material in nanometer sizes, it would seem instant.
I was wondering why this was not done earlier, but nice to see this tech finally implemented.
This reminds me of an "April fools" article from the old Magazine "Electronics Now" in 1999. They claimed an analog chip that was pin-compatible with Pentium II's and thousands of times more powerful. I was a very young child and I believed the article and I started to get very excited about it. Great video, thank you!
I recently met with an MIT researcher to discuss very similar analog tech for ML. I learned a lot from this vid. And I like your Cartier Santos watch as well!
The power required discussion was very interesting.
Thanks
That was very interesting and informative, thank you.
Very cool, i heard about the potential of analog computers being able to speed up AI training and remember a youtuber demonstrating a calculation, but thats about it, far off from training a network and here's IBM already on top of it, very nice. The in processing ram too which was an unexpected combination. Theres a channel on youtube thats been giving lectures about the data bottleneck and the solution- computing on the memory, its cool to see these 2 bleeding edge concepts working together. Now only to program a game that brings this tech no its knees, crysis 4?
Hi Anastasi. I like your patience at explaining things. Thanks.
I needed this video in may.
Thank you good information
back in the late '80s the audio greeting cards became a thing. They were analog storage. Instead of storing a digital value and having a DA converter, they allocated one bit per sample and used the pulse/measure/pulse method to program the bit to match the analog value being recorded and then at playback, they just toggled through the bits and fed them to an amp. The dirt cheap chips that were put in greeting cards were getting about 8 bit data out of a single bit that way.
programming with multiple pulses like that is much slower than normal storage, but is pretty reliable and depending on the recording cell type can last a long time.
Also, the wear is on erase cycles, not programming pulses (especially partial programming pulses)
The problem with this result (just like Mythic's analog edge AI processor) is the power efficiency is not particularly impressive. The reported 12.4 TOPS/W is no better than efficient digital neutral networks. Oppo's 2 year old phone AP with a 6nm process node has an 11.6 TOPS/W CNN core, but that falls far short of the current leaders. Gyrfalcon claims 24 TOPS/W, but Perceive's Ergo chip (or maybe its Ergo 2) claims 55 TOPS/W.
This is an area I've done some research in. My original Ph.D. topic in 1990 was digital and analog neutral network implementations, and then I worked on the topic again a few years later and then again for a DARPA project around 2009. As you discussed, the fact is a lot of benefits of the analog computation evaporate when you include the DACs (for the input voltage) and ADCs and then if you need to extend the equivalent arithmetic precision beyond the SNR limitations of a basic analog multiplier.
It’s easy to say that reducing the power of ADCs and DACs could resolve that, but it’s much harder to actually do that. Reducing power of both these circuits has been an area of massive amounts of research for other chips. For example, the highly competitive $10B CMOS image sensor market demands ever lower power ADCs, and yet still those ADCs are very power hungry.
There was a big push in the direction of analog computation after Carver Mead published his book “Analog VLSI and Neural Systems” in 1989. I heard many people making extravagant claims that analog computation would soon replace digital (like the time one of Carver Mead’s students who gave a talk at my university said, “Oh! Carver says that no one’s working on digital any more”).
The hype never manifested because of the issues you and I have both discussed. Another issue is that the design and validation cycles for complex analog chips are so long that by the time you get your product to market, your digital competitors have moved on a process node, or two. Maybe that runs out at some point with the much-discussed end of Moore’s Law, but there’s still a long way to go before lower power digital innovations run out of steam.
I’m not saying this approach can never yield benefits (evangelists would always point to the brain being analog and only consuming around 12W), but extraordinary claims require extraordinary evidence. You can’t just claim a low-power analog multiply-accumulate (again) and extrapolate that to a lower power real-world system. You need to demonstrate that end to end power efficiency.
One third of electricity expenditure in data-centres is used on cooling, so the argument for analog chips starts to become more appealing given how little heat the little buggers produce.
@@locinolacolino1302 it only becomes appealing if they actually use less power. The power gets converted to heat. If the full system including ADCs and DACs uses more power than a completely digital implementation, then it means it creates more heat.
Analog to digital converters in the chip architecture. Boy these are great videos! Thank you
Thank you so much for your videos! Right now I am studying semiconductors and organic electronics and your channel gives huge motivation to keep working. We really need more beautiful and cheerful people talking about technologies and how they can make our world better. People think of science as somethiing boring and dull, and science is difficult indeed. However, if a pretty woman or a good-looking man explain it in a charismatic way, it will help us a lot. So thank you and I wish you all the best!!!
P.S. Только сейчас узнал, что вы из Москвы) я из Зеленограда (где как раз разрабатывалась советская микроэлектроника), еще раз желаю вам удачи на поприще блогинга, у вас это отлично получается!
I like the way Anastasi explains things...
Great vidéo...thank you..
The information is excellent. It really sounds so familiar.
It's great that all this is moving along at pace, whats not great is that IBM exists as an entity.
This is awesome stuff! This analogue compute is just for inference though, correct? Not for training?
Hi John, right at the moment it’s just for inference :)
Fantastic content, thank you.
Q: Given Moores law is there any argument for CPU design where all internal registers get replaced by a single static RAM area where any address can act as register (and as fast) but also acting as the L1 cache?
Thank you!
My first thought when i heard "phase change" was about temperature sensitivity. If the ambient temperature changes in the computer (such as from the waste heat of nearby digital components), will it affect the phase of all bits (and will they be affected uniformly)? Does this need to be cooled with special equipment as quantum processors do or can it run at room temperature?
In any case, thanks for the new topic(s) to explore. 😄
The analog components are made of a ferroelectric material that can be constructed in such a way that a single component can act as a capacitor resistor or transistor
I think the output is calibrated to take into account a scale of temperature.
The low energy consumption of chips like this would be great for mobile devices, and are a necessity for advanced AI entities like androids.
Great topic. The "analog computing" is digital but not necessarily binary. It can work on base n (n= integer). So one "bit" in base 256 for example , is equivalent to 8 bits in base 2. As long as we can distinguish all 256 states without noise, similar to how we distinguish 0 from 1 as -5 V and + 5 V. We can compute faster and we can store more info per bit ... We should have been doing this a long time ago to compensate for Moore's Law limitations. As allways Anastasi explained this very clear. Now I understand it much better ...
This is very relaxing to watch. It's hard to find good copy these days. About half way through I started to rub the top of my head quite a bit. Thanks Anastasi for calming me down quite a bit.
As far as the accuracy of analog computers, could that be trained in similar to how a biological does? it would slow down the production process at first and produce a lot of highly specialized chips, but in time we'd figure how to develop the training protocols to broaden the applications and efficiency.
Fascinating channel with excellent production quality. I hope to see you grow quickly to millions of viewers, as you are talking about the most interesting new tech, like the huge Cerebras chip and this analog chip. Thank you for explaining these new technologies in an interesting way, without making me feel dumb for not being an engineer myself. ❤🔥
Thank you 😊
Wish your videos had more views. Glad to see more EEs with software experience.
I really liked the robot suit on the last video:)
Not heard of "Colossus" then? It used valves to perform Boolean functions, therefore it was digital. That was around in 1943.
Analogue chips. i need read up on this ❤ Great video.
Thank you.
This technology is interesting taking an old idea of. Analog and improving on it. But it seems to me like they need much more testing.
The older I get, I the more complex and unfathomable all this stuff seems to be. I sure hope the next generation will be smart enough and educated enough to maintain what this generation is creating!
AI will be doing that for us in a generation.
Not maintaining but advancing things further at a scale millions of times faster than humans ever could.
what are those blue dots simulation at 3:30 ? Is it a simulation of a process ?
Very interesting!
Amazing explanation! I would be thrilled to know more about the Design and Architecture challenges in ADC and Readout operations and how are these the current bottleneck for this tech.
The issue is that they need many ADCs to scale the MAC for larger models. Usual area/power/performance trade-off.
ENOB of ADC is the key metric. Higher ENOB more energy goes into ADCs, longer conversion time. Lower ENOB lower target accuracy. The number of ADCs is typically 1 every 8 memory columns, but could be also less up to 1 in 64 columns.
However, the most critical global thing to optimize is the size of the memory array. Larger arrays, more efficient ADC usage, better peak TOPS/W (key global metric, not TOPS/mm2 like in the paper since that depends heavily on the technology node and used memory cell technology). However, larger arrays are more difficult to fabricate and have lower utilization reducing average TOPS/W.
IBM already outsourced everything, thus can sit on it as that's what they do for all their new innovations.
It must be the case that the temperature of phase change memory will have to be controlled accurately.
This accuracy could be supported with a phase change material environment.
Obviously the phase change environment will have to have just enough delay so that it does not swamp the signal.
Very interesting.
Hinton's Forward Forward algorithm paper talks about AC and importance of it in Mortal Computing... good to see progress in it...
This is amazing. Working the memorized data in the same place. I was thinking about it for two years already but this is so advanced I will drop development.
I believe she mentioned it in another video, but it's important to recall that not only the performance and energy utilization of accessing memory is a bottleneck right now, but static memory is not scaling as the logic is doing. So all seems to point we will have even more problems in the future.
Thank You
i have an odd question. what watch are you wearing?
keep up the great work and excellent content. your videos are well done, well researched, and you explain everything so well for interested laymen like me.
I find it interesting the analog cells containing weights only need to change during the training phase, but IBM's example shows the inference phase is really simple as a sum of currents. I wonder if at some point in the far future we'll have mass-produced 'hard-coded' chips for TRAINED networks that simply use fixed resistance elements for each of the weights.
The difficulty is that your trading the flexibility of changing the weights for a speed gain. One would have to send out a new processor with every revision of the trained system. The chip foundries would love it.
Uplifting development. ✨👏🎉
Amazing and very interesting! as a software developer , I really try to follow that field for some time too !!!
Phase change memory does seem like it could be a good way to reduce the cost of running pretrained AI models, but I'd be surprised if it will be production quality anytime soon. In the near term I think moving digital memory closer to the processor using chiplet designs will be more practical. The Samsung cache DRAM looks quite interesting. I'd like to see the age of separate sticks of RAM go away and just put all the memory on top of the CPU chip. The challenge there is getting all the added heat out, so that will also take a while to reach production.
Thank you for the insightful comment.
Let’s see which tech gets multi-bit first.
Agree, we will see more stacking and near memory computing in the near future
That would be a disaster for the consumer though as far as flexibility for RAM amounts.... There is no way manufacturer's would allow you the same flexibility of just swapping out RAM sticks or installing more RAM.
Love your knowledge delivered so elegantly and, easy to grasp! Big fan glad I found your channel...❤
I read on IBM's webpage about analog AI chips that this type of hardware uses drastically less power to do AI tasks. A recurrent neural network transducer 'speech-to-text' was implemented in analog AI chip, tested, and compared with the same function implemented with digital AI chip: 17 times less power required. Power consumption of biological organic neural networks must be insanely low, but impossible to implement in silicon. I think we are still a decade away from affordable General Purpose Robots (GPR's), even with analog AI chips.
Pretty soon, my smartphone will have these, and one day it'll climb out of my pocket, grow legs and walk away, singing redemption songs....
Very cool stuff
Fascinating 😊
Wow great report of IBM chip Northpole opening doors to new memory efficiency.
Fascinating, thanks. Bet von Neumann would see the humor. All the best
I love this. Can wait to start using this tech for helping people own, utilize and monetize their personal data.
Is there not algorithm to differential eqution compiler available yet? Why?
Also 11:45 is there any work into hex computer chips?
Lovely videos, always! Thank you :)
I love watching your vids as inspiration for making compute circuitry in minecraft 😅
Thank you Anastasi. Groet, Arjan
I don't think it's a problem if much more efficient analog chips don't achieve the same accuracy due to analog noise, as long as you can subsequently finetune the network on conventional hardware to similar accuracy as networkes trained entirely on conventional hardware.
Maybe in the future AI will help with designing analogue chips that speeds up how it designs analogue chips.
I also been doing a lot work with analog computers, but not in the way they are used in this video, I found to be very useful in analog signal processing applications and direct conversion without the need ADC or DACs. I used this technology with RF modulators and demodulates, that make part of signal processing block, that able to work with noisy input signals. Applications so far: are in Navigation and Television decoders, with ideas to expand this Radar signal processing to analyze differences in phase shift and Doppler.
Interesting, have you published your research ?
@@AnastasiInTech Electronic magazines and on linked in, once I have a basic working prototype I end up moving on to doing something else. This is life been outside of the academic system, I am always happy to pass on my research for other to work with.
It seems to me that a broader area encoding can make these analog chips much more resilient. The problem is the digital small area controller to phase change such a large area. Thus the query sent to the area can be checked and rechecked for veracity. When a higher return is discovered then the answer yes or no can be binary output.
The complexity of it all! 🤯
so the this analog chip is a dedicated tensor core for doing matrix multiplication?
The energy savings of compute in memory to my mind is the most important point. Given the energy costs and heat dissipation problems of current compute in large scale projects, any advancements that cut energy consumption are useful. Analog per se, may not be the solution for more general computational purposes, but the concept of elimination of the time and power waste in data movement is something to look at across the board. From a different angle, I cannot help but think that in the process of attempting to solve the fabrications problems inherent in designing analog elements the folks at IBM and elsewhere may well solve other fabrication difficulties dealing with other unconventional materials and microelectronic structures. I keep hoping to see more advancement in non silicon semiconductors and exotic designs such as graphene transistors.
Just want to know how do they measure the accuracy to 92.8%?🤔
Probably by collecting all the numbers resulted dévided by the times done. This is just simple way to calculate accuracy so I can’t say they did that exactly
excellent.
Why not a huge and fast FPGA with memory built into the one wafer, and have the FPGA allocate logic structure based on an individual application being driven. In turn, if you wish to have several applications running at once, multiple cores could be utilized to facilitate this.
I don't see why desktop processing has to always be limited to one expensive CPU, I would prefer to have an array of 4-8 CPU sockets on a motherboard, or should I say multipurpose sockets for CPUs, SOCs, GPUs, AI chips, music synths by brands like Korg and Yamaha, PhotoShop accelerators, Camera and video related accelerators, FPGA, etc.
I like sockets better than cards because the bus running to slots is always anemically slow. We need a bunch of circuits feeding into a master socket which rules them all and connects to RAM. I like the idea of RAM baked into the wafer, but having slower ram in huge quantities can also be great;.
Hmm.. IBM, here are my immediate thoughts:
Instead of phase change cells, it would be faster and cheaper to use capacitive memory.
That is, the weight stored as an actual variable charge/voltage on a capacitor.
Changing the charge on a capacitor (during training) is almost instantaneous, and can be done a practically unlimited number of times.
And then, instead of sequential ADC (which is slow), use old school resistive ladder conversion, which is practically instantaneous.
One of the challenges would be, of course, to make the capacitors sufficiently low leakage to hold a somewhat stable charge over time.
Hope this reaches You guys.. Perhaps Anastasi would relay it?
Cheers / Love.
Anastasi, can you explain why we yet to see light being used as a fast data line more often in technology?
Brilliant giving the chip noise to train it. Then retrieving clean accurate info.😅 This is fun. is there IBM in Kenya?