Will AI Design Computer Chips Faster?
Вставка
- Опубліковано 10 чер 2024
- Get 10% off your first month of therapy with our sponsor BetterHelp: betterhelp.com/anastasi
TIMESTAMPS:
00:00 - Why to use ChatGPT for Chip Design?
01:26 - GPT for Chip Design Explained
06:13 - How to Train LLMs to Design Computer Chips
10:20 - Deepmind's Chip Designing AI Explained
12:10 - Other Companies & Stocks
The Paper: arxiv.org/abs/2309.10730
👉 Support me at Patreon ➜ / anastasiintech
📩 Sign up for my Deep In Tech Newsletter for free! ➜ anastasiintech.substack.com
Get 10% off your first month of therapy with our sponsor BetterHelp: betterhelp.com/anastasi Let me know what you think!
of course ai nowadays cant cover the WHOLE process of design, but can automate and suggest actions in the way, that is also a great advantage to accelerate further innovations.
Computers are giving birth to computers, nothing to worry about 😄
I can listen to you videos all day. You're so smart with such a pleasant voice.
I'd like to see a deep dive into the possibilities of nanoimprit litography in the future nodes such as "2nm" and similar. Canon might be on the way back, without fancy mirrors from Carl Zeiss!
I for one welcome our robot overlords. 🤖
I'm a 3D artists/designer in Unity and not a programmer. I've found that for simple scripts it works for me. I've learned it's about how you ask, and how to explain your need in Unity terms. My partner and programmer has become a GTP Ninja.
Prompt Engineer is the new team leader position.
I tried using ChatGTP to create a Python script to scrap data from a specific website. After feeding back the errors of two failed attempts, it said the dynamic nature of the website was perhaps the cause of the issue, and it went on to suggest the URL of another website containing the same data, supplied the code and it worked. I consider that creative and I was pretty blown away!!
when GPT4 failed at solving captchas, it submitted a mechanicalturk task to have a human solve them... kid you not
@@manonamission2000 Yes, there was some cheating around it. There were given some answers to a questions that it couldn't have whatever ability to answer. So it means that there were was some real-time human whispering.
Could be interesting to compare the time, the learning gain, and else, with a fellow that used the classic approach to obtain the same result. I hope I stated it clear, and have to do with honest persons - that so many gpt adepts seem to not know that the GPT Parrot is based on knowledge based on human evolution, and that the humans so far evolved without help of gpt parroting.
Sounds like you were unable to formulate your query properly.
Although ChatGPT is really good at retrieving working code and can piece together code, in a lot of cases you have to either know how to specify all the relevant parameters to obtain the desired result. ChatGPT like any human can only go so far guessing the interrogator's true intent.
@@tonysu8860 «Sounds like you were unable to formulate your query properly.»
--
Something more precise like "this, then this, then that ...", and you finish with writing the code and asking gpt to produce the very same code.
----
«ChatGPT like any human can only go so far guessing the interrogator's true intent.»
--
ChatGPT cannot 'guess' since guessing implies estimation, and since estimations implies reasoning, and since whatever kind of reasoning is absent in chatgpt.
It just pattern matches statistically and / or stochastically.
Good video Anastas
Fantastic video ! I was always wondering if someone was taking on designing chips with LLMs. And here we go ! Congrats GeorgiaTech and thanks again Anastasi
Thanks for pointing out other companies working in this field and their stock !
I wonder how updates to LLMs (like increasing the context window size or the parameter count) will change their ability to generate chips. My guess is the biggest leaps will come from these updates for a while, not necessarily the additional training in chip design specifically. Definitely an interesting idea though!
IMO there will be dedicated LLMs especially for chips. Don’t you think so ?
@@dchdch8290 In the short term, yes. In the long term it's likely that the best models for chips will be a general purpose one. In the same way that Google's RT2 performed on average 50% better with the OpenX dataset (a HUGE text dataset), and even outperformed specialized robot AI's.
@@snailedltthanks ! Good point.
Spot on!
LLM's make the entire chipmaking process faster and better. This leads to better chips which will allow us to power even larger LLM's.
But there is still a lot of room for efficiency in the LLM's. Good thing LLM's can help us make more efficient LLM's! It's all snowballing, and every tech-related sector will massively increase in productivity, value and performance over the coming years!
Thank you, Anastasi, for bringing us the latest cutting-edge technology so elegantly and professionally. 🙏🙏
This process of giving the LLM "hats" that it wears for various steps are showing effectiveness in many fields, from coding games, to writing novels (yes, you make it a slush editor, then content editor, etc etc), and now to chip design. We need to think of prompting large problems as more like how you would design a company, having multiple roles which either specialist AI agents can be slotted into or a general AI can be guided to wear a hat for that role. This is how humans naturally think anyway, utilizing latent space activation in the mind, and it's how we can take our utilization of AI as partners to the next level.
This is already part of the discussion and has for a bit, but we are finally seeing more talk about the specialized agents. Most conversations about this using a center general agent that connects with a number of specialized agents
@@mc9723 Exactly, so, for a novelist I know he says he personally acts as the general agent in the center which organizes and connects the specialist AI agents that he's created around himself. In this way he's largely replaced what his publishing house would normally provide to him, including the marketing team. He workshops his own work whenever he likes. No waiting months for feedback, which helps him keep his work flow going. It's increased his productivity by an order of magnitude. Until we get AGI, the human functions as the general AI agent in the center. Leave it to a sci-fi writer to figure out how to get it done.
@@armartin0003 «Leave it to a sci-fi writer to figure out how to get it done»
--
Sorry, but the problems were, and always will be, solved by those that live on our planet. Sci-fi is nice, I read it all when I was young. It can be useful as long as one is able to distinguish between fantasy and reality constraints. So while sci-fi is useful in enriching the minds, sci-fi writers are harmful when they are not able to understand what they are talking about.
@@voltydequa845 Writers live in the clouds, engineers live on the surface of the earth. However, without a greater view of the realm of the possible provided by creative minds, engineers would still be designing better plows.
@@armartin0003 «Writers live in the clouds, engineers live on the surface of the earth. However, without a greater view of the realm of the possible provided by creative minds, engineers would still be designing better plows.»
--
Nope. Mine was not against fantasy, but against confusion where commercial fantasy-hype is taken as reality. Writers write their narrative, and analysts write their view on where the world could be headed.
Or, said in other terms, my advice regarding intelligence in sci-fi is that of avoiding the cognitive confusion that arises when a ParrotGPT is perceived as a form of intelligence.
So my wasn't about clouds, but about planets, where cognitive confusion is presented as fantasy.
Sci-fi writers should try to understand what-it-is-all-about (gpt, llm, etc) for the sake of avoiding extra confusion.
In those pattern matching technologies there's no whatever cognitive ability. And there won't be since the bridging with the real AI - that can only be implemented by the over-complex (in terms of definition, as well as calculus) symbolic logic - is impossibile since that the gpt llm can't produce (formal) knowledge, and the symbolic logic part can't use, and anyway won't need, the gpt llm part, since the analyse and generation would be logic.
"Our Martin is eating his ice-cream" , where the words and their order are chosen not for their meaning and for grammatical reasons, but just because in 99.999% of the texts in appears this way. Symbolic logic is so hard because it has to do with intelligence. All the rest is talking the talk by resorting to imitation.
p.s feel tired, hope to have transmitted the essence of the difference, and so why all this nonsense-out-of-excess-of-hype-and-disinformation.
Very cool, now time to create a self hosted MemGPT instance, load it with as much data on you drive with the design phase, and test!
I'd like to see a video on EUV alternatives in future process nodes
i think we can leverage chatgpt for any system and design a system given enough context. i work for a company that uses chatgpt apis for replacing teachers in some form. what i did for them is just break down various steps that a teacher would perform fundamentally. similarly i think if we can replicate this approach in chip design we can accomplish at-least repetitive and common tasks. i hope we can solve the "creativity problem"
«i think we can leverage chatgpt for any system and design a system given enough context. »
--
Too vague. To obtain what, exactly?
----
« i hope we can solve the "creativity problem"»
--
Creativity concept implies thoughts that have no precedents. But GPT is just a parrot technique. What can (!) be seen as creativity can come from calculus on symbolic logic, so from almost total calculus (of all the possible alternatives) that could come up with solutions that human couldn't find since too much and too deep. Anyway nothing to do with gpt parroting.
@@voltydequa845 I don't know what you are referring to when you say about calculus and other techniques to solve creativity problem. Right now they are just using different words and sentences to solve repetitiveness.
Coming to an earlier point, what I mean is if we break down complex tasks a human would traditionally perform to complete the task and we can replicate it in some form that llm would understand, then it will be able to do the task with very enhanced capability.
@@vatanrangani8033 «I don't know what you are referring to when you say about calculus and other techniques to solve creativity problem. Right now they are just using different words and sentences to solve repetitiveness. »
--
Give a fast look at Prolog, for example. As for the different words, I guess they are just resorting to pattern matching that cross synonyms with their usage frequency. It is just another level of pattern matching.
Back to calculus - that you seem a kind soul - it is also useful to think in terms of comparison between algorithmic chess and gpt chess. The first calculates by means of ramification and evaluation for the sake of the goal of winning, while the second just pattern matches using its database where a response move is based upon probability out of victory (and of course not necessarily out of that move). So the first one calculates and evaluates, the second one blindly imitates.
----
«Coming to an earlier point, what I mean is if we break down complex tasks a human would traditionally perform to complete the task and we can replicate it in some form that llm would understand, then it will be able to do the task with very enhanced capability.»
--
I do not believe so. I do not believe that complex tasks in disciplines like teaching (and similar) can be so easily fractioned. We should keep aware that we are talking about teaching under the form of talk, where too often goes "whatever way you say it". I mean there is no, and there cannot be (or it would be too difficult), a measure of the quality of teaching. As for me, I prefer the google search engine's way because the alternatives keep open my mind, bring new ideas, let me preserve the critical spirit.
Anyway a very complex topic.
I have found ChatGPT (GPT-4) somewhat useless when it comes to writing SystemVerilog code. My suspicion is that because there isn't a large corpus of open source hardware code, there isn't enough data to train off to make good RTL. It somewhat frequently uses C-like curly braces instead of begin/end, and writes code that's sort of in the right direction, but really doesn't connect together in a sensible way. Perhaps the fact that hardware code is inherently concurrent doesn't fit well with what the model has learned from software design. My go to test has been to ask it to create a clock domain crossing FIFO, because that requires understanding metastability, gray codes, RAM modeling, etc.
I already used ChatGPT during my Digital Systems course to help me with VHDL language, but certainly it's not perfect and you still need to know the concepts. I made a VGA Controller as my first project, and now I'm working in the Snake game written in assembly language running in a MIPS processor inside of an FPGA.
Currently watching at work as a orthopedic manufacturer 😊
Very interesting, and conceptually it makes a lot of sense that you could leverage the sort of AI which has been proven at playing chess and go, and make a game of optimising circuits. Going the other way, and hoping for generative / creative design of chips in what sounds like more of a top-down approach, that sounds a lot more optimistic by contrast (to a lay person anyway). Could they build on the game, take collections of optimised circuits and build larger constructions instead? Move upwards from the bottom... Feels like that might better suit current capabilities.
software will solve it sooner, it just needs more context, maybe a million. reason being that you want to feed the build logs, output and generated code back to the model. so when you have a medium sized transformer, that will already be 100k+ tokens with output. so this is extremely limiting considering how short transformer code like llama is. but with a million you can build larger more compex models. and then you just feed it back in a loop 100 times per second on a supercomputer and just through sheer attempts will eventually have working code. apply genetic algorithms to that and wait half a year for the next gen models to arrive and it will build something extraordinary
I think the easier and faster way is to do what people do, just build on abstractions. I don't see how spitting out a working HDL of CPU in one go Will be easier than working over multiple abstraction layers connected by text, as in the first case all of the reasoning of what to put where would have to happen mostly in the latent space while in the second case, the concepts in the latent space wont as complicated because they will be abstracted away. No matter the details I think with todays model we could get working chips, we just have to have a way to make them agentic and work in the loop.
Always a good day when Anastasi posts :) Thanks for the hardwork and effort to give us inspiring and fun educational content! Love watching your channel grow!
The cool part about Chat GPT is, you might be able to start getting it to program by its self, just by talking to it:)
Thats fantastic:)
Levels of coolness, Rising!
Most intuitive Video....Keep doing such videos...
I observe that AI more accurately means "Augmented Intelligence". This work is a perfect example of why.
I would look at training a set of domain specific AI. One skilled in lithography for example, another in architectures. Then cascade them so that one is a controller AI that can create prompts for the others, collate the results, request reviews by spawning other instances of each domain to check the work. Hugging Face released a paper showing this kind of agent spawning other domain specific agents to farm out a more complex task to instances that are really good at each bit.
In terms of training I would literally use the materials we use to train a human to be skilled at x. Using the tests to test them, but include for example every paper handed in for assessment by students. There would be shared knowledge they all would have, such as physics, logic - to be able to communicate amongst each other.
Each item of knowledge would need context and relationships - such as physical laws, materials, tools, process, hard, code, soft etc
There is a reason why there is more than one of us. makes no difference which instance of intelligence one of us might be,
Thanks for stock tips.
In many cases we can not have an efficient design of a chip today. We have to have the matrix design patterns wasting millions and millions of transistors. Matrix patterns are easy to understand, microcode and troubleshoot. It would be amazing if we can have AI to optimize chip layouts and nano structures. It is not only the people (numerous teams) that spend year or two, it is also very long computer optimization and calculations. AI can lead us to the next generations of chips, faster than ever.
This is amazing but I saw this coming as it evolves. You can imagine what else can be done in the next 6 months to a year. Thank you as always your videos are amazing and provide a professional but simple way to understand. Even if I don’t have an engineering degree
«You can imagine what else can be done in the next 6 months»
--
People with void memory think in terms of "what else... " without reference to "what so far..."
So after six years (10 times more), without whatever real (not just marketing vague hype) progress happening, they will still be here to repeat their parroting trust in the bright future based on the bright future of the past six months.
Thanks for the great content
This scares me very much because they showed this technology in a closed event back in 2012 and they are revealing this "now"? What else are these companies doing behind the curtain that we don't know about?
Great work! Love you channel! Nice jacket by the way. Keep up the great work! PS Your look is so great when i first started watching you I thought you were AI. Maybe still but I love your work and you line it up so perfectly.
Excellent
On the Circuit Neural Networks: That is the absolute opposite corner to a constraint random verification approach I would say. You cannot verify that IP 🙂
its good to know you wont be replaced designing chips just yet. very interesting to see how far AI design has gotten in the tec sector thanks for the vid look forward to more.
Likely thousands of use cases where AI can find better designs and efficiencies overlooked by humans.
Anya, great show, today! I think going for PHD, would be a great idea, if you already have your financials in order and under contract control, of course. A reason, might be, you are so good lecturing on UA-cam, that teaching in academia may very well be is what can truly fulfill you. Just a thought.😊
thank you 🤓
She has already a valuable PhD relative to her public. Share with others. Help share the Words Of Anastasia. When she gets hundred thousand subscribers she will start her own academia for all of you. And you are going to know much more concrete, of it all, than getting lectures in a real academia.
Thanks for the great coverage on this, Anastasi! It's scary how quickly things can move in the world of AI. On a related note, I've been wondering something that I hope you might be able to shed light on: If a company that makes an AI chip decides to use Intel's fabs to make them, could doing so allow Intel to simply steal the critical parts of the design for use in its own AI chips? Is the only thing stopping them a sense of honor or is it not nearly that easy to do so?
No. There will be IP protection. Imagine TSMC is producing everyone's chips - this would be a nightmare scenario.
@@mvasa2582TSMC does not have a conflict of interest. Intel has broken laws and ethics in the past to get a competitive advantage.
@@mvasa2582 Oh I don't mean lift the whole thing wholesale. More like, "Hey we seem to be stuck at X but this chip here that we're fabbing solved it in this way, maybe we could do something similar?" Basically, protections aside, I'm wondering from purely a tech standpoint, can it be that easy to steal ideas if you are fabbing the chip or is that pretty much impossible. Essentially, could the fab "pay the fine" if it meant getting through their wall.
@@garhong9125 point. Any technology can be copied replicated or stolen. I believe with AI, we may even move to a world without IP. Innovation at the speed of light! 😂
Usually happy to hear your news. Not so much this time, I'm afraid, even if this was inevitable. Having a Luddite moment, I guess.
Thank you very much. Please can you provide the link to the paper on Deepmind's AI design ?
Thanks
Given an LLM and statistical layers, you can predict the next token in the sentence (if a picture the next group of pixels or if a game the next move or if a robot the next movement, etc.) The key is that in order to design a paragraph it needs to have enough data to know the language and enough layers to predict the next word in language and enough structure to complete the paragraph (depends on the data, not the code).
The challenge is to discover the language of computer design. It implies making the LLM of the process of design (the many design concepts and methods) and CREATING a language with which to express how you design a chip.
Challenging. Good luck and thank you for your very observant channel.
I'm java programmer and I use Chat GPT wheneaver possible. Is very competent creating simple code o as reference. Not all generated code is perfect and need refinement here and there. Anyway is a great tool for me👍
I spent some time with ChatGPT to try and design a neuromorphic chip (because we really need one for the masses) and it helped.. mostly though, I feel like it taught me a lot
1:35
Yes ,that is a preview of whats on the horizon with AGI apporaching
Please bring back the lovely music at the end.
a chip is only a lot of blocks wired together llms use statistique to place word to give it meaning so if a llms is teach on transistor block it may be possible to use it to build chips
makes sense, Tenstorrent has talked about it and is training models to help them design their chips. I bet Nvidia and TSMC are already deep into this research.
The Flying Circus of Asses is betting too on ChatGPT to design their spectacles, of course with the constraints of not letting Asses free to fly away.
our ingenuity never ceases to amaze me, no matter what we find a way to advance technology nothing gets in our way
What do you a think about new M3 macbook?
Instead of similarity between the code, the similarity between unit tests of the code would be a better metric for LLM training. This allows for a diverse range of code and also for performance data to influence the training. AI can create the unit tests for you too.
I've been wondering if EDA tools would be able to take advantage of Neural nets and LLMs.Would it be true to say that analogue circuits would prove harder to design,test and optimise than digital ones since there seems to be much in common with the logic gates of digital circuits(and therefore chips) and neural nets but less so with analogue circuits?
Great info anyway!
for theirtain programming tasks gpt advanced data analyst is freaking awesome and does a great job
It all depends on the data. You can use data augmentation whereby the augmented data is very specific to hardware chip design data paradigm.
I'm not so sure it will take a long time! As Ray Kurzweil has said, information technology has an accelerating progress, and that's particularly true in terms of AI. For example TSMC may already be feeding massive amounts of microchip data into AI models. It would be unethical and probably even illegal if TSMC used the designs they get from companies like Nvidia and Qualcomm, but they probably have a lot of their own indirect data that they can use.
I find that ChatGPT is very good for specific functions. So if you know what you specifically want. You can build it block by block by creating and implementing supporting functions without ChatGPT even knowing that its for. For example, I used it to create a function that leverages antonyms to change the sentiment of a string to positive or negative.
Thank you Anastasia . Excellent work . Great information .
It's only a matter of time before AI can do better than humans. It'll be able to run calculations and test design patterns much faster than a human will ever be able to.
I'd assume it might not happen until AGI or the training as you say is guided in that particular direction. (Kinda like cancer detection for AI)
You have an interesting channel. But I have doubts about chatgpt chip making. For this, it is better to use the Anthropic API. In my opinion, he is better versed in modern nanoelectronics. In particular, Anthropic AI chat offers the most advanced technical solutions in photonics and molecular nanoelectronics. Chatgpt can simply be used additionally for additional data processing. In general, the idea is interesting. But for a full-fledged implementation, it is better to use a multi-agent system with a certain number of defined modalities.
أنا أتابع من العراق 🇮🇶💟
اني هم من العراق 👍
@@jokerwashington6227 ❤️
What happens with current AI is that it is a generic system, it is just an assistant. It is instructed and become "expert" only at the moment you ask the question. It is not an authentic Expert System that becomes more and more capable in a specific subject. Intelligence itself tells you it. If it acquired all the knowledge, experimentation and expertise of hundreds or thousands of professionals (which it could do, they just won't let it), the current AI would be unstoppable.
6:50 - Risc V chip layout design is done with php??
Interesting
Hi Anastasi, I'm aim to become an engineer working in the chip design industry, what Xilinx/Alterta board should I buy to learn?, my budget is under 150$.
I don't know about you guys, but this sounds like a prelude to the supercomputer Deep Thought from The Hitchhiker’s Guide to the Galaxy
As AI become smarter. Like ASI Then eventually it will be able to not only design. but to test it without actually making the hardware, just reviewing everything.
As a chip designer I do find vast differences in quality between different designers, one extreme example was I wrote an audio extraction logic block for SDI that was 1/100th of the size of paid for IP. The problem wasn't clever or efficient implementation but a different thinking about the top level, this is something AI is still weak in, but I do expect AI will be a useful tool to help designers in the current generation.
The place where implementation quality would be key would be AI controlled full custom digital design, the difficulty here would be immense though and produce designs that likely only the AI itself would understand (a human might take years to decode it), best-o-luck finding a bug in that :)
Long long way to go… well, I agree.
Just few days ago, I watched a battery scientist asked ChatGPT some questions about lithium ion battery, and she spotted some flaws and issues on another UA-cam channel.
I also left my message to her and said I think ChatGPT isn’t so accurate now, because I also asked ChatGPT some questions about industrial furnaces in different ways, and I found out some contradictions in February this year.
I knew there’re many different backgrounds on those scientists and professors behind the scenes in ChatGPT. And those scientists were not battery professors whom received Nobel price like John Goodenough…
Different minds caused different output after all.
But I think maybe ChatGPT has more professional background on these circuits design, I guess. Maybe those experts are those whom expertise in electronics circuits and its related fields the most.🤔
It's been less accurate lately. this has been documented true.
Very nice and beautiful voice🤩
Interesting, but as is the chats are being used to try & do what humans do, but faster. Instead it would be interesting to tell the llm what is needed & ask it to devise ways of achieving this need. In this mode it would potentially find ways that humans have not thought about. Thanks for sharing!
What you talk about is right, but regards symbolic logic, and has nothing, and won't ever have, to do with gpt / llm pattern parroting.
Now that LLMs like LLaMA are smaller and more efficient, it wouldn't be hard to imagine using three different LLMs to design a chip. Use one with GPT-4V to recognize, label, and then analyze microchips. Use another to be an adversarial agent that does QA. Then, a third one could be a custom in-house LLM that uses that previous data to model entire systems before any prototpe is fabricated.
Is there an AI that can design optimized PCB layouts, given a circuit diagram and output Gerber files? Because I think it is a massive pain to do by hand!
PSA Better Health sold client health data to marketers. Feel free to do what you will with that information.
"When we develop a computer model capable of designing and testing a chip, even if the first iteration takes months or years, it will be revolutionary and a step towards the singularity. Humanity already possesses extremely powerful computers running various simulations. We can direct these resources towards a 'design a better chip' use case. Once we've built these enhanced chips, we can use them to design even more powerful ones at a faster pace. Over time, this could approach what seems like 'infinite' computational power.
This open the path to new processors designed by a.i. like an programmable array pal for each user and task in real time
Palai 🎉
So they’re saying ChatGPT isn’t naturally great at chip design, so they’re doing everything they can think of to try and make it as good as possible at chip design… are we actually trying to cause the singularity? It’s like we can’t wait.
I use Bard for my CPU not ChatGPT
👍😇💪💖
The first question I would've asked is whether the form of knowledge representation is appropriate for the given task at hand. LLM's would not be the first thought, that comes into mind. It's oriented towards language ("predicting the next word", or riffing on more abstract themes). I would've thought for chip design, one needs, at a minimum, 3-D spatial awareness. Just as for AlphaFold, you need to build in some priors regarding structure biology. ChatGPT (LLM's) would not be the right tool for protein folding.
Perhaps LLM's can help with high level themes, at the idea level. But that's about it. But it's not really aware of the physical world. (The UA-camr AtomicBlender, tried to use ChatGPT to design a novel nuclear reactor to see it would replace nuclear engineers someday. Not anywhere close.)
Making a LLM using all the legally obtainable chip designs, including old ones, would likely be far more effective at this though the training process takes time.
Tried to generate system verilog code with Char GPT. Epic fail, it could not even help me in my work.
We’re approaching the singularity 😬
Anastasiia.... I believe in training AI until I see chatgpt can train itself.... IT becomes a parent. Think what happens to parents when they respond to protecting their young YES unreasonable reaction. Yes or No ???
didnt tsmc use a.i to speed up some part of wafer manufacturing?
If AI can be used to design simple chips well, that might be the best start. Of course, there is no need to have AI for this, but if it can understand all of the RC propogation delays, tolerances, etc., and can scale a design to use different nodes (lambda), then that creates a foundation that you can trust, while waiting for AI to advance enough to be ready for a cpu, or gpu, etc.. Perhaps it could proove its worth with analog chips, or substrate interconnects and logic between different tiles. Or in optmizing microcode.
I used ChatGPT in my job programming for backend, it really doesn't understand the code that much easily as of now and fails to make changes that I want , I still have to tweak the code, sometimes I ended up frustrating, doing everything myself
Once we will reach the point that it will understand the link between hardware and software, I think it will be the beginning of the singularity and perhaps a year of two before we will be left at the event horizon.
lol, i have been using chatGPT also to design computer architecture including the chips in it - i would say it more-so verifies some design choices - it is intersting to know that they use chatGPT also for chip design - and they say any... in particuar we are developing general purpose Ananlog computers (not digital)
Anastasi are together.
They forgot to add "Let's think this through step-by-step" at the end of their prompt! Lol
Not surprising since the basic design blocks of computers are over 60+ years old. Most of the layout was already automated, with engineers hand positioning blocks, busses and interconnections, letting to tools do the optimal routing.
The only thing that makes it better is more pipelining and parallelism that gets around basic design bottlenecks.
AI can test thousands of tradeoffs to select the best ones that engineers don't have the physical time to test and analyze.
how long do you think it will take before AI can make its own cpu gpu designs that are much better than the ones we have come up with
here's the singularity 👍
AI is super handy, but as an embedded software engineer, I would not trust my nano seconds with AI yet.
So basically Anastasi )) Chip designers will be on the Dole Queue soon too, with programmers and legal & healthcare professional 😂😂😂
Next to build will be a system using a cpu and pal and gpt ai to create circuits on pal on demand
and with the replication on this pla or pal or programmable matrix of any circuit from remote ..i just invented terminator robot ??
Interesting, we tend to think at our own level when we ask questions to chat GPT, and it is incomprehensible for anybody to think above their own level and therefore they will not get good results from anything like bard or chat GPT.
It should learn LabView first, then it will be simpler to move to hardware implementations since it knows the data flows and processes needed in each node.
With the continuous advancements in artificial intelligence and the accumulation of data, it's inevitable that AI will significantly influence the chip design sector. This technological evolution promises substantial benefits in terms of efficiency and quality. However, from a philosophical standpoint, human attributes such as creativity and empathy remain irreplaceable, for now. AI will have a glorious future.
Well, i asked chatgpt to solve p=np, and it went into a slight panic mode and didn't want to try 😐
RLHF chip engineering specifications is another great biped data mining source for the centralized A/GI to continue a path to optimized self replication.
Transformer space needed today will be reduced to 8% of current physical space requirements by Jan 2024. Whether it is built then is another story.
I don't necessarily agree with giving chat GP other devices to look at per se or what they look like because then you are narrowing what it will. Do.i would think the best way to go about having a identity. Something is give the parameters capabilities in restrictions of atranister.
Then have it design its own off of that off of what we know to be true on the limitations of atrium sister but if you start bank this is. What others look like? It's gonna. Look at that and it might model off of that instead of generating something new
I asked ChatGPT to design me a cpu that's faster that what Intel or AMD has. And all i got bitching about how hard it is. Lazy AI.
❤❤❤
Using LLM’s to design chips is analogous to consuming soup with a fork. You can reason, deduce, snap your finger and click your heels all you want; you will not find success. Or maybe this - using LLM’s to design chips is like using Tardigrades to repaint the Sistine Chapel. Both can do neat tricks, but there’re just the incorrect instruments. Rant, more rant, continued rant…I think I’ve made my point.
let's see...
The error rate of even our best AI isn't something I'd want to hand over to a very expensive design validation team.
Not this year, anyway.