NVIDIA Trains Custom AI To Assist Chip Designers

AI is big news lately, but as with all new technology moves, it’s important to pierce through the hype. Recent news about NVIDIA creating a custom large language model (LLM) called ChipNeMo to assist in chip design is tailor-made for breathless hyperbole, so it’s refreshing to read exactly how such a thing is genuinely useful.

ChipNeMo is trained on the highly specific domain of semiconductor design via internal code repositories, documentation, and more. The result is a vast 43-billion parameter LLM running on a single A100 GPU that actually plays no direct role in designing chips, but focuses instead on making designers’ jobs easier.

For example, it turns out that senior designers spend a lot of time answering questions from junior designers. If a junior designer can ask ChipNeMo a question like “what does signal x from memory unit y do?” and that saves a senior designer’s time, then NVIDIA says the tool is already worth it. In addition, it turns out another big time sink for designers is dealing with bugs. Bugs are extensively documented in a variety of ways, and designers spend a lot of time reading documentation just to grasp the basics of a particular bug. Acting as a smart interface to such narrowly-focused repositories is something a tool like ChipNeMo excels at, because it can provide not just summaries but also concrete references and sources. Saving developer time in this way is a clear and easy win.

It’s an internal tool and part research project, but it’s easy to see the benefits ChipNeMo can bring. Using LLMs trained on internal information for internal use is something organizations have experimented with (for example, Mozilla did so, while explaining how to do it for yourself) but it’s interesting to see a clear roadmap to assisting developers in concrete ways.

16 thoughts on “NVIDIA Trains Custom AI To Assist Chip Designers

    1. I think that once you would be doing the hard work of designing many different working 555 timers, and collecting all the possible failure modes over 10 years, and setup the test harness for this ASIC, you could do the easy job of throwing AI at the problem. ;)

  1. If they have the documentation to train the AI what signal X does, seems that the AI could be trained to reply “RTFM”.

    Next year they’ll use it to design the test cases, and at least then if it makes mistakes the tests will still pass…

  2. ” Using LLMs trained on internal information for internal use is something organizations have experimented with (for example, Mozilla did so, while explaining how to do it for yourself) but it’s interesting to see a clear roadmap to assisting developers in concrete ways.”

    NSA might create something to deal with their “internal information”. ;-)

    1. It began with the human equivalent to a neural network, compartmentalized, each individual doing their bit, at first running parallel with, then as an assistant tool. The data still requires verification, not because the process isn’t verifiable if designed correctly, ie. As a human brain usually has a little voice in it’s head to discuss the thought process. But because pouring in data from untrusted sources isn’t entirely safe. Having a splice on fibre-optic cables and sifting through twenty four seven, pulling out keywords is one thing, inferring them is another, especially with people trying to poison the data, like the slightly corrupted images featured on HaD a while back, making the AI model infer the wrong outcome. It’s generally a trickier problem than one might first expect.

    1. Currently, can machine learning improve it’s learning algorithms? E.g. given the same training set and an initial model algorithm, and heuristics, can the learning algorithm change itself to improve its ability to learn?

        1. Next up, AI will become “evolutionary” where competing models within the model fight for survival of the fitist. It’s only a matter of time before AI becomes self-evolving while performing it’s work getting better overtime without human involvement. It’s likey this is already underway as it is a logical progression of the technology

      1. “Learning” is the name given to one step of number crunching to estimate the coefficients of a graph. These coefficients are chosen so that using the graph gives the best score to a problem automatically measurable: some height reached by a construction, the total distance traveled by some object, a distance between two points…

        It is not the same as “learning” as used for humans.

        > Can learning algorithm change itself to improve its ability to learn?

        If this is an heuristic question and it can be asked in a formal way then the answer is “yes, sure, of course, this is something the machine learning calculator can do as long as you can put the numbers into it”.

        If you can find one score that summarizes how efficiently the learning process was, and can expose a list of parameters as numbers, then you may plug a neural network in-between the two, and use machine learning to guess the coefficients, just like any other task. The difficulty is not putting the numbers into the “AI calculator”, it is asking the question the right way. :)

  3. To me, “AI” is just another few branches of mathematics, like statistics are in a way.
    Computers are used to build computers since long: https://en.wikipedia.org/wiki/SPICE#Origins

    But the marketing team wanted something more seducing than “graph-based heuristics” or something like that, so claimed practicing technomancy and called the result “intelligence”, and the public applauded.

    Engineers (and everyone) needs to not get caught by that cheap marketing trick.

    Another reading frame for this is that NVIDIA tries to take shortcuts for reducing the legacy debt… or worse: allow it going further without externally visible impact.

    1. And you would be wrong because the term “artificial intelligence” was coined by American computer scientist and cognitive scientist John McCarthy in 1955. It was defined as “the capability of computer systems or algorithms to imitate intelligent human behavior”. It’s not a specific branch of math or a particular method.

      There is plenty to blame on marketing but they had no part in this.

      1. Thank you! I did not know the origin of the name.


        > In some respects, computers are easily more self-conscious than human beings, it’s not hard to make a computer program look at its own program.

        This is not the same as saying this specie is capable of “self-consciousness” and this specie is usually not.

        An “A.I. chip” is a chip with cores optimized to perform some machine learning computation (inference).

        I welcome both the philosophical discussion and these new computation of heuristics, and even the two combined.

        But confusing them is dangerous as it becomes believing in magic to a whole new level.

Leave a Reply

Please be kind and respectful to help make the comments section excellent. (Comment Policy)

This site uses Akismet to reduce spam. Learn how your comment data is processed.