Five key trends for CFD revisited

0:00 Hi, and welcome to the Neil Ashton Podcast. In each episode, we explained some of the fascinating ways that science and engineering are changing the world around us. We talked to leading engineers from elite level sports like cycling in Formula One to some of the world's top academics to understand how fluid dynamics, machine learning, supercomputing are bringing in a new era discovery. We also hear some of their life stories, their career advice, the lessons they've learned on the way that I hope will be helpful to you too. So sit back and enjoy this episode. Hi, welcome back to the new Ashton podcast. I thought I'd do something

0:43 interesting today, which I was. I was looking back at the previous episodes and I saw that almost a year ago today I did an episode called the Future of CFD Five key trends to watch. It was season 2 episode 2 and I thought I will try and do another version of that and basically see a year on, was I right or was I wrong? So what I had as the five were the rise of GP, us in CFDAI and machine learning, shift to cloud computing, digital certification, and I think I had the last one which was mergers and acquisitions. OK, Now this was before I joined NVIDIA. Oh, I should just say. So obviously Nvidia's you know,

1:33 creates GPUs. So the rise of GPUs in CFD. Can I answer that with an being completely unbiased? Probably not, but it's true. Basically, there has been an unbelievable rise in GPS for CFDI. Can say that even if I wasn't working for NVIDIA, I think people in the community know that to be the case. Has that accelerated in the past year? Yes, most definitely it is. Yeah. Just the number one question. And again, I know I work for NVIDIA, so I'm going to say that, but it's true. I go to conferences and I see it without me even talking to them. You know, they're showing slides saying, OK, it's 10 times faster on a GPUI did not get involved in that study whatsoever.

2:18 I'm just sat there listening to it. So it it is absolutely the case. And as anything there was, it was slower at the beginning. People were like, OK, is this really true? You know, is this marketing? I think now there's been so many independent studies showing it that people are fully getting behind. And I think what's also made the difference is 1, the hardware's obviously keeps getting better and better. The pace of GP us, because obviously AI has helped. This is far out pacing. CP US in terms of like generational generation, you know, like the improvement that you get each generation is, is like this compared to CP US,

2:54 which is, you know, a bit slower. So you're, by moving to it, you keep getting better and better. So the reason to move has probably accelerated. I think the, there's been a healthy competition between different CFD companies each seeing GP used to be part of that differentiating thing. And that has created a, a sort of an acceleration, you know, because there's one vendor releases AGP version number one wants to make sure they also have one. So there's been a little bit of that, but for good reasons. It's not purely, you know, just for the sake of it. It's because people are genuinely seeing a speed up customers liking it and and

3:33 therefore they're doing it. And I think most interestingly, which is I think something I commented on is there's a lot of start-ups now who are specifically focusing their on GPU native. And that has also definitely played a ship because no surprise if you write code from scratch with a sort of software, hardware code design, you're going to create something really good. And so because of those start-ups, they have shown in some ways the like most optimum solution and has really shown some very interesting performance. What I find more at a technical level interesting is where the whole like unstructured implicit finite volume, which I guess was

4:20 the industry standard does that. Is that the most optimum numerical approach on GPS? It could be, but because they're so much more powerful and because of that, you can start to do more transient high fidelity LES, then people are going, well, if I'm going to do an LES and I'm going to have to run lots of time steps and I've got this GPU, should I maybe now go back to explicit that previously wasn't maybe the optimum way of doing it, but now with the GP it is. And then maybe could I do a, a war modelled approach if I'm doing a war model, LES, and does that mean I should maybe use the most boundary method because I don't need to resolve the wall?

5:07 And so that it's this interplay between the hardware and the software is creating some really interesting changes. And and finally, one of the interesting was a high order side, You know, I think high order, without going into too much of the technical details, there's some interesting stuff around like tensor cords, etcetera. So the whole high order thing is coming back in to play. So yeah, was my prediction of key trends for CFT correct, GPUs and CFT, I'd say yes. And is it still for 20/25/26? Definitely. It's only accelerating and I think you'll find that it continues to accelerate whilst there is so much more learnings to be had in terms of optimal

5:56 architectures. And there's like a crossover point when your compute gets higher, you can go to a different modeling paradigm. So for example, and not wanted to speak about this for too long, you've gone from, let's say a wall resolved RANS like a low Y plus RANS where you wanted to resolve the ball and where you need a certain gridding strategy to be able to resolve the wall to a wall modelled LES where you say, oh, well, I don't need to resolve the wall because I'm modelling it. So maybe I can go to a different gridding paradigm and I can go to explicit. But if you go to the next stage, which is the war resolved Elias,

6:33 that maybe say, well, actually now my gridding paradigm that was good for war modeled LES may not be good for war resolved LES. Maybe I need to do something slightly different. So it's interesting how these waves happen and the, like I said, doesn't interplay. So what I find most interesting is a year ago or two years ago, the most bold predictions was more towards Hybridrans, Elias, war modeled and everyone said our war resolved is too far away. I'm actually seeing now people be a bit bolder and say, well, you know what with the latest GP us maybe a war resolved Elias, it started to become possible. And that's interesting. I think not for everyday

7:13 applications perhaps, but now that that is coming up, some other stuff starts to get interesting like just the sheer number of time steps you need to do. And yeah, so still important. OK, AI, machine learning. Well, duh. Yeah, that's clearly still a key trend. Has things moved up in the last year? Absolutely. I'd say one of the big breakthroughs I personally seen is probably a year ago or a little bit four year ago, one of the bottlenecks seemed to be volume prediction using AI surrogates. Stuff was coming out around like graph neural Nets and being able to predict things, but all the examples were always on smaller cases.

7:54 And one of the things that I did with some colleagues was to create that drive ML data set, which was really put out there as a data set to give realistically size meshes and hopefully encourage people to see, right. Can my method take the entire volume and a year or certainly two years ago, it looked like not many people could do it. Everyone was just doing the surface or maybe a few slices and every time you saw something, you know in marketing online, it was always like the surface of the car or a slice of the car. What I found was breakthrough in the past year is peep various methods have come out and I guess the most, most most recent will be like Transformers that

8:40 seem to be the most computationally efficient. Efficient for doing full volume predictions as in training on a full volume and then doing an inference and being able to get a full volume out. I feel like that really accelerated the realization that some of these AI methods could give you back a flow field that was more similar to a traditional one. And yeah, that transition from like graphs to maybe neural operators and Transformers for me has been a really interesting 1. And I'm certainly excited to see what happens in the next year. I think we'll continue to see new architectures, new evolution of, of AI. So that is definitely a trend

9:18 that's continuing. Yeah. And they're obviously linked, right. So if you have a faster solver, you can generate data more easily. You can then do the training. I think one of the big discussions now is on the right code architecture. Sorry, like programming environment, because online training is definitely a hot topic. You know, how can you take advantage of the transient nature of the flow? So at the moment, I'd say the state-of-the-art is normally that you would just generate those data and even if it was with a hybrid brand or well modelled, you'll just take a time average and dump that and train on that. But I think what's emerging now is this interesting, well, can I

9:56 use some of the transient data that's being created anyway, you know, dump them out at checkpoint and train on that. So I think the whole AI will continue and continue. The one thing that I would probably add to that AI bucket is the agentic AI side. So let's say you've got surrogates that's in one market and you've got LLMS for just general productivity gains and coding on the other. I put a gentic AI somewhere in the middle, which is how can you have multiple agents working autonomously through some sort of master agent that can go off and do tasks for you. So instead of you having to manually click run this simulation or write a Python code that goes and runs, you'd

10:41 said, give a text prompt saying I would like to go and run this simulation with these parameters and please create me a report. And you'll have a workflow where an agent will interpret that request will then go and let's say create a simulation setup for you. But crucially also then run the simulation, let's say an HPC cluster, use a separate, you know, agent could be a surrogate model essentially, you know, to go run it or to do the meshing, then another LLM, maybe do the analysis of the results. And the point being is that you have to kick that all off with a textbook. That is the most cutting edge example. And that's what I feel is probably in the next 12 months

11:19 will be the next big jump. And some start-ups are already doing it as some companies, but I feel that's probably one of the big next steps. I should add, by the way, for GP US, I didn't mention, I think precision is a very interesting topic now. I think there's huge gains if CFD can use lower precision for for obvious reasons, you know, that the so much flops in in the lower precision because of AI generally only needs, you know, 48 or 16. So it doesn't mean that you can't do 64 or 32, but if you can find a way to harness that power, then you can really unlock the next big accelerate. So the first one I had was the

12:00 shift to cloud computing. Now, you know, people who know me or listened to this before know that you know, I currently work from video, but I used to work for Amazon Web Services, which is obviously a massive cloud computing company. And there is no doubt whatsoever that cloud computing has totally transformed the IT sector and the HPC sector. When I left AWSI definitely saw an increasing momentum in people adopting the cloud cause of the fact that they could essentially focus on the stuff that mattered to them and not on the sort of manual data centre stuff. I still see that increasing, but probably the bit that maybe I was not brainwashed because I'm

12:49 joking aside a little bit. But you I definitely, I'm seeing I'm still a more hybrid sense where on one hand people are absolutely seeing the value of the cloud, you know, through ADBS or GCP or Microsoft or you know, all the cloud plates. And by the way, there's an interesting angle now with all these neo clouds like core weave, etcetera, which is mainly I guess to the AI some GPU angle, but I'm still seeing a hybrid used. You know, there's still quite a lot of on Prem, there's cloud and and I still think there's a lot of room for improvements in the way to access cloud resources. You know, I think SAS from ACFD point of view now SAS products are still not fully there.

13:38 You know, if you ask most people are they doing it software service? No, they may use the cloud to have a cluster, but are they doing it in a true SAS like manner? I think that it's slower than I thought people still seem to prefer and there's good reason. So we'd have to get into of having a local experience in that you're not just doing through a web browser. And I think that's probably just decades of engineers being used to just SSH ING and stuff and and having like, yeah, running scripts. And it's something that you think would change, but it, it does seem remarkably stubborn. And yeah, so I, I definitely see the cloud continuing, but I'm

14:27 still waiting for some company to come out there and really make it easy from a workflow point of view in like a true hybrid way or to just simplify maybe the cloud experience. You still have to be quite an expert to do it. So I am, I'm not changing that prediction. I'm just seeing that it seems to be a bit slower than I fully appreciated. And maybe in hindsight, that's because I was seeing all of the positive and success stories at ADBS and maybe I was not hearing the bits where people weren't, you know, doing it. The 4th 1 I had was on digital certification to a higher fidelity methods that, as I mentioned on the 3rd of it is definitely the case.

15:15 You know, you're having a huge increase of people going to higher fidelity methods because of #1 you know, the fact that you can access to compute more easily. And that is definitely a topic that I think will continue to increase sector by sector. So the automotive has definitely pioneered that and all have moved to higher fidelity methods. I think now, like I said, they're, they're looking how can I go from hybrid RANS to war resolved earlier. I think the aerospace sector is probably still trying to get from RANS, you know, to hybrid random warm up LES and the work that you know, we did in the last high lift pitch workshop has definitely helped to push

15:55 that forward. And there's some interesting work going on now to try and push it even further again to like more resolved LES. But I see that continually, you know, increasing and now it's more around building up best practices and understanding of those methods, you know, robustness that everybody knows where they go wrong, where they go right. So it probably will take time for that and for people to see the value because it is undeniably more expensive to do with the low fidelity. So I think there needs to be success stories that have been created. But I think probably one of the biggest challenges is transition.

16:31 Certainly I'm speaking more from ACFD and external aero point of view, but one of the challenges still with even going to war monthly yes to hybrid is transition modelling. And I, I feel that that's why the war resolved LES is still very tempting because it doesn't fully solve it, but it, it definitely helps to overcome some of those challenges. So I think that will continue, but I would look more at how do we make war resolved LES approaches more, more affordable. I am going to be doing an episode where we're going to talk more on the combustion side of things. I do realise that this podcast just tend to be a little bit focused on external arrow.

17:10 That's basically my background. So I am purposely trying to, yeah, diversify a little bit and make sure that we're covering other areas. And the final one was mergers and acquisitions and innovations. And this has definitely kept going. We have seen ANSYS and Synopsys, so two big companies merge. We have seen, you know Siemens, Altair, Cadence, Beta, CAE. So this has kept going and was a trend that I predicted and is still true today that there is so much interest in the community aid in engineering and semiconductor businesses that,

17:58 you know, it's a growing business, really important. And big companies are realizing the value of acquiring or merging with others to, you know, bring some of their technology in and speed up. You know, that that time. I think that will continue. I would say the prediction for next year, it's not there's not many other big companies that can merge now. You know, it has consolidated a little bit to Cadence, Siemens and, and synopsis, but where I do see that being definitely move is there's a lot of start-ups and I'm sure that the many of those start-ups will be acquired by some of these bigger companies. So they can start to really bring those to a more enterprise

18:42 level and integrate it. So I, my prediction would be in the next 12 months that some of the start-ups that you, you know, you may know of, I won't mention all names, but the start-ups, you know, I think you'll see some of them being acquired by high profile companies just because that is, that's a tried and tested route. You know, when I spoke to the CTO of Ansys and he openly admitted it, he said, that's what we do. We look for great start-ups. We may even, you know, encourage them. And once they reach a certain level, we might acquire it. And it, that's a great way of developing new technology. Sometimes it's better to let a start up do it.

19:15 You have a, you know, faster way of doing than a, let's say, a big enterprise. But then they acquire them, they bring them, and that's how customers get to use them. So I think they also still continue. So where do I think things are going? Well, I think actually those five topics will still continue to be a thing for 25 and 26, and I'm excited to see how they accelerate. So I hope that was interesting. I thought I'd keep it short. I know these podcasts tend to go to an hour or two hours, so I thought I'll keep this one short and got a couple of really good guests coming up for the next two episodes. So yeah, look out for those.

19:52 And for now, I hope you enjoyed this episode.