Developing an AI enhanced chip design process Podcast – Transcript
If the only constant is change, then AI represents the next big change for many tools and industries. High-level synthesis, one of the key tools in the semiconductor design process, is set to be one of the tools that receives a big upgrade from AI. Currently, HLS processes rely heavily on manual adjustments and hard-coded heuristics yet, with integrated intelligence, a new approach to chip design begins to emerge.
Check out the full episode here or keep reading for a transcript of that conversation.
Spencer Acain:
Hello, and welcome to the AI Spectrum podcast, I’m your host, Spencer Acain. In this series, we explore a wide range of AI topics from all across Siemens and how they’re applied to different technologies. Today, I’m joined by Russell Klein, program director for Siemens EDA’s high-level synthesis team. Picking up where we left off last time, you talked a bit about how in Catapult you are starting to use AI as a way to replace hard-coded heuristics, but you also mentioned, of course, that this isn’t just the compiler in Catapult and HLS tools doing this, but you mentioned, of course, the C compiler, GCC, so do you see this being like a widespread implementation? Will we see AI-enhanced code compiling as it’s becoming the standard?
Russell Klein:
Well, I don’t work in those tools, but I would imagine that anybody who’s building a compiler today, again, they can replace a lot of these heuristics with artificial intelligence that’s simply just going to make a better decision, as you said, we’re going to replace a very simple decision tree with something that understands the context of the program, that’s looked at hundreds of thousands or millions of different implementations and figured out what… And it statistically remembers what worked well in different cases, and it’s going to be able to apply those. But I think we’ll see in the software domain, all across the hardware design flow, if we look at creating the RTL itself through high-level synthesis, we’re going to see a lot of that.
But we’ll also see it as you go from RTL down to gate-level, so that’s the traditional synthesis in the hardware design process, there’s a ton of heuristics that live at that level as well. And when we go from a gate-level implementation to the hardware layout, so the masks that are going to be used to produce, again, we’ve got a bunch of heuristics that are sprinkled all throughout that code, all of these are really key opportunities for AI to come in and make engineers more productive, make their lives easier, because they don’t have to worry about making all these decisions and make the end products come out better.
Spencer Acain:
Yeah, that makes sense. So, I think we talked a bit about the way AI is making these complex tools easier to use, HLS is not only the tool itself getting smarter, but like you said, the users themselves, the engineers, they need to know less of the arcane knowledge when it comes to working with the tool itself and can focus more on the actual design process. So, do you see this as kind of, I don’t know, a paradigm shift almost, changing the way that people are using these tools, HLS tools, chip design tools, or maybe even engineering software tools in general?
Russell Klein:
I don’t know if it’s a paradigm shift, it’s going to, if you will, democratize the use of a lot of these tools. Because as tools have more and more options to them, and more and more capabilities, you have to learn what all those things are. You’ve got to understand all of the different potential decisions that the tool can make, and how do you give it inputs to produce the output that you’re looking for. As we make the tool smarter, and we make the interfaces to them smarter, the end user doesn’t need to learn the tricks of how do I drive this tool, it can simply say, I want this end result, and here’s my starting point.
And between generative AI and smart decision-making throughout the engineering process that the tool’s going to go through, a lot of that becomes just much easier to use. So, engineers don’t have to worry about all that arcane knowledge, they can really focus on the craft and the creativity of designing things, not about how do I get it to push through this tool to get the end result that I’m looking for. So, I think it’s just going to make the design happen quite a bit easier.
Spencer Acain:
Yeah, that sounds like it could be really valuable, Russ, do you have any examples of that, maybe have you experienced any tools that are doing it this way?
Russell Klein:
This is a little bit of an aside, but it was a really interesting thing that happened to me. So, I’m taking classes at the local university, and one of the classes were using the Cadence Virtuoso tool. And Cadence, that’s a SPICE-level transistor simulator and layout tool. So, it basically goes between the gate-level and the transistor-level. It’s a tool I’ve never used before, it’s actually an area of design that I haven’t worked in before. I generally work at a very high level. One thing that was rather notable is that the Cadence tool was really rather difficult to use, because it’s like, okay, I’m editing this transistor, I want to know how to resize the length of the transistor, but I don’t want it to move side to side, so I want to stretch it vertically only.
Is there a way to do that? And you Google some things, and you look at the manual, and you try… Oh yeah, you right-click, and hold this key down, and grab it by the corner, and do this. But how do you know to do that? Again, you’ve got to Google it, you’ve got to do this. They’ve created what’s called an MCP, a model context… Oh, I don’t remember the name of it. I’ll have to look up MCP.
Spencer Acain:
Model context protocol.
Russell Klein:
There it is, model context, it’s a model context protocol that somebody had built for the Virtuoso tool. And so, then you could ask this and say, how do I do this thing? And it says, oh, here it is, you do these… And what it did was make the process of using that tool so much easier, because now I could simply say, here’s what I’m trying to do, oh, here’s what we’ve got. Now, Siemens is working on this as well, I haven’t used the Siemens one as much, but right now… And maybe you turn the recording back on. So, in Catapult, we’ve started to use this FuseAI, and one of the things the FuseAI does is when you’re using it in the context of the tool, it’s read all the manuals, it’s read all of our support cases, and you can ask it, how do I do this? How do I perform some specific operation? It’ll tell you, here’s the command to enter, here are the menu picks, here’s how you go make that happen.
The next step on this type of AI is, rather than… And we’ve got this in place with Catapult now on some of the steps that you’d go through. Is rather than asking it, how do I print this to whatever, or how do I save this to a new directory, instead of it giving you the individual commands, it will actually execute those commands for you. So, now you can say, oh, save a copy of this design to my backup directory, and let me come back to it later. That’s it, it’ll go run all those commands that you needed to do. Or can we change the target technology library that we’re using this, and increase the frequency by 10%, and tell me how big this is. Right now, it’ll give you a set of commands, for some of these commands that’ll actually run them for you long-term, it’s just going to run them for you. And so, you’ll say, I want to do this, and those types of AI are going to be able to just run the tool for you, and take care of all those commands.
This makes the UI of these very complex tools a whole lot easier, because a ton of effort goes into building these UIs to try and make all of these things easy, and it always falls short. But having a layer of intelligence between here’s what I want to do, what menu picks and commands do I use? That all gets to melt away, and it just becomes, here’s what I want to do, and the tool just does that.
Spencer Acain:
It really sounds like we’re moving, engineering and HLS are moving kind of toward results-driven, you can just say your end goal, and the technologies that you’re going to be working with, but now you don’t have to be fighting the tool so much to get to that, you just say, I want to do this, and it can do it. Yeah, I think that’s going to, like you said, democratize this a lot, really make it more accessible, and make it easier to use, and everything like that. That is all the time we have for this episode. So, once again, I have been your host, Spencer Acain, joined by Russ Klein on AI Spectrum podcast. Tune in again next time as we continue our discussion on developing next generation AI chips and how AI itself can help with that.
Siemens Digital Industries Software helps organizations of all sizes digitally transform using software, hardware and services from the Siemens Xcelerator business platform. Siemens’ software and the comprehensive digital twin enable companies to optimize their design, engineering and manufacturing processes to turn today’s ideas into the sustainable products of the future. From chips to entire systems, from product to process, across all industries. Siemens Digital Industries Software – Accelerating transformation.
