1 00:00:00,280 --> 00:00:02,400 Hi, and welcome to the Neil Ashton Podcast. 2 00:00:03,080 --> 00:00:05,920 In each episode, we explained some of the fascinating ways 3 00:00:05,920 --> 00:00:09,080 that science and engineering are changing the world around us. 4 00:00:09,800 --> 00:00:12,760 We talked to leading engineers from elite level sports like 5 00:00:12,840 --> 00:00:16,920 cycling and Formula One to some of the world's top academics to 6 00:00:16,920 --> 00:00:21,000 understand how fluid dynamics, machine learning, supercomputing 7 00:00:21,360 --> 00:00:23,000 are bringing in a new era of discovery. 8 00:00:23,960 --> 00:00:27,080 We also hear some of their life stories, their career advice, 9 00:00:27,640 --> 00:00:30,080 the lessons they've learned on the way that I hope will be 10 00:00:30,080 --> 00:00:33,840 helpful to you too. So sit back and enjoy this 11 00:00:33,840 --> 00:00:41,160 episode. Hi, and welcome back to the Neil 12 00:00:41,160 --> 00:00:45,000 Ashton Podcast. So today's episode is just me 13 00:00:45,720 --> 00:00:51,600 and I wanted to go through what I thought were five of the top 14 00:00:52,360 --> 00:00:57,960 important trends in CFD, let's say over the now and over the 15 00:00:57,960 --> 00:01:01,720 next 5 or so years. I'd love to know whether you 16 00:01:01,720 --> 00:01:03,800 agree with what I'm going to say. 17 00:01:03,880 --> 00:01:06,880 These are sort of more, I guess, high level trends rather than 18 00:01:06,880 --> 00:01:08,960 getting into the, you know, specific details. 19 00:01:09,720 --> 00:01:12,560 Very happy to do an episode diving a bit deeper into one of 20 00:01:12,560 --> 00:01:16,240 these, but I thought I would maybe stimulate some discussion 21 00:01:16,240 --> 00:01:18,680 and get you thinking about some of these. 22 00:01:19,840 --> 00:01:25,000 And I have to also give a, a bit of, I think that I'm coming from 23 00:01:25,000 --> 00:01:28,280 a certain angle here. I guess my experiences in 24 00:01:28,280 --> 00:01:31,160 everything I've done will bias me. 25 00:01:31,160 --> 00:01:33,560 And I just want to admit that that bias. 26 00:01:34,320 --> 00:01:37,040 I come more from an external aerodynamics background in terms 27 00:01:37,040 --> 00:01:40,080 of CFD, you know, I currently work for a cloud computing 28 00:01:40,080 --> 00:01:41,880 company. I'm obviously going to see 29 00:01:41,880 --> 00:01:45,960 things through the conversations I have and the lens I see, but 30 00:01:45,960 --> 00:01:49,600 I'd like to think that I see enough of what's going on in 31 00:01:49,600 --> 00:01:52,880 different sort of verticals and and different areas that gives 32 00:01:52,880 --> 00:01:55,920 me a reasonable understanding of what's going on. 33 00:01:55,920 --> 00:01:59,960 So yeah, let's see if you agree. So yeah, five things that I 34 00:01:59,960 --> 00:02:05,120 think are important for CFD over the next five years or so, not 35 00:02:05,120 --> 00:02:09,360 necessarily in chronological or like in in important order, but 36 00:02:09,360 --> 00:02:16,920 number one for me is GPUs and based processes. 37 00:02:17,600 --> 00:02:22,400 Why do I say this? It was the case when I let's say 38 00:02:22,400 --> 00:02:26,200 back in 2013 or something like that, I remember going to a path 39 00:02:26,200 --> 00:02:29,200 CFD conference. So this they still run parallel 40 00:02:29,200 --> 00:02:31,320 CFD. It's a great conference to go to 41 00:02:31,840 --> 00:02:34,720 sort of that blend between high performance computing and CFD. 42 00:02:35,040 --> 00:02:37,640 And I remember there like 10 years ago, this was at Barcelona 43 00:02:37,640 --> 00:02:40,600 Supercomputing Center, really great venue right on the 44 00:02:40,600 --> 00:02:43,120 Ramblers sort of strip. Anyone who's listening who went 45 00:02:43,120 --> 00:02:45,600 there will know it was great conference organised by the 46 00:02:45,600 --> 00:02:49,520 Barcelona people and I remember seeing presentations showing 47 00:02:49,520 --> 00:02:53,320 about GPUs and I can't believe that was like 10 years ago now. 48 00:02:54,560 --> 00:02:59,760 And there was a real sense in the room of I remember someone 49 00:02:59,760 --> 00:03:02,440 said that you can only get good performance out of it if you 50 00:03:02,440 --> 00:03:06,960 write the code from scratch. I remember someone saying that 51 00:03:07,120 --> 00:03:09,680 and there being a bit of a debate and I remember there 52 00:03:09,680 --> 00:03:15,480 being sort of sensational claims, you know, that GPUs are 53 00:03:15,480 --> 00:03:19,120 this much faster and, and a little bit of a, oh, you know, 54 00:03:19,280 --> 00:03:23,360 ignore them sort of thing. And obviously at that time 55 00:03:23,720 --> 00:03:28,360 NVIDIA was doing a lot of work to invest in in CUDA and sort of 56 00:03:28,360 --> 00:03:31,280 giving funding to research groups. 57 00:03:31,320 --> 00:03:35,800 And I remember seeing this and yeah, it, it, it was, it 58 00:03:36,320 --> 00:03:39,600 registered on my radar. But a lot of people were, were, 59 00:03:39,640 --> 00:03:42,360 were debating whether it could be possible, whether there'd be 60 00:03:42,360 --> 00:03:46,960 enough memory on the GPU's. Was it really possible to 61 00:03:46,960 --> 00:03:49,160 rewrite a code? You know, people said they had 62 00:03:49,160 --> 00:03:51,640 millions of lines of code. I'm never going to be able to do 63 00:03:51,640 --> 00:03:56,840 this. Fast forward to now and I would 64 00:03:56,840 --> 00:04:05,800 say for me it's incredibly clear that GPU's are going to be the 65 00:04:06,680 --> 00:04:11,280 the dominant computing platform for CFD. 66 00:04:13,560 --> 00:04:17,200 But as I'll say later, things move very slowly, especially the 67 00:04:17,200 --> 00:04:20,360 industrial CFD. And so the adoption is going to 68 00:04:20,360 --> 00:04:24,400 take time. But I have seen enough to make 69 00:04:24,400 --> 00:04:26,400 me believe that and I'll give for several reasons. 70 00:04:27,120 --> 00:04:32,400 One, there are published publications out there that show 71 00:04:32,400 --> 00:04:33,520 it. I've certainly showed it in 72 00:04:33,520 --> 00:04:38,080 public settings and it's the Isvs and software companies have 73 00:04:38,080 --> 00:04:40,480 done it. And in fact just for auto CFD, 74 00:04:40,560 --> 00:04:45,280 this automotive workshop that that I created back in 2018 in 75 00:04:45,280 --> 00:04:48,160 19 with Gary Page Loughborough and now it's in its fourth 76 00:04:48,160 --> 00:04:50,280 edition. I think the slides are now up 77 00:04:50,360 --> 00:04:53,520 and the video is on the website. So if if you go to autocfd.org 78 00:04:54,120 --> 00:04:59,280 and you Scroll down and you go to the HPC presentation, Herbert 79 00:04:59,280 --> 00:05:01,280 Owen from Barcelona Supercomputing Centre did a 80 00:05:01,280 --> 00:05:08,120 great job of taking everybody's submissions from CPU and GPU and 81 00:05:08,160 --> 00:05:12,480 then looking at what type of GPU, what type of CPU, and doing 82 00:05:12,480 --> 00:05:14,680 a really good, good job of looking at several different 83 00:05:14,680 --> 00:05:17,120 metrics. So the speed, you know, per 84 00:05:17,200 --> 00:05:21,560 degree of freedom, per time step, looking at the cost. 85 00:05:21,560 --> 00:05:25,760 So we took a sort of a notional cost from AWS. 86 00:05:25,760 --> 00:05:27,120 It could have been any cloud provider. 87 00:05:27,120 --> 00:05:29,800 It's just, you know, it's the only way of getting an actual 88 00:05:29,800 --> 00:05:34,280 price for these things because obviously the cloud gives you 89 00:05:34,280 --> 00:05:37,440 the all in cost. Whereas if you were to find a 90 00:05:37,440 --> 00:05:40,560 website that would give you just the sort of the GPU or the CPU 91 00:05:40,560 --> 00:05:44,480 that wouldn't be the actual entire cost including the power 92 00:05:44,480 --> 00:05:50,840 and the the network etcetera. And what it showed is that for 93 00:05:50,840 --> 00:05:58,160 external aerodynamics of cars, GPUs were more cost effective, 94 00:05:58,680 --> 00:06:03,000 at least 23510. It depended on the code or even 95 00:06:03,000 --> 00:06:10,680 more cheaper than an equivalent CPU at the scale that it was 96 00:06:10,680 --> 00:06:13,240 being one for a realistic problem, not sort of toy 97 00:06:13,240 --> 00:06:17,680 benchmark problem. And they were faster, but that's 98 00:06:17,680 --> 00:06:19,880 a slightly harder one because it depends what the parallel 99 00:06:19,880 --> 00:06:23,440 scaling of the code is and and you know, how many CPUs or GPUs 100 00:06:23,440 --> 00:06:27,320 do you throw at it. Whereas the cost, you know, 101 00:06:27,520 --> 00:06:30,200 obviously it's going to change a little bit on linear regime. 102 00:06:30,200 --> 00:06:32,920 You know, if you're really badly scaling, the cost is going to go 103 00:06:32,920 --> 00:06:35,400 up relative. But essentially what it showed 104 00:06:35,840 --> 00:06:38,200 was the cost. Now I have repeated this same 105 00:06:38,200 --> 00:06:43,880 thing myself personally with startup codes and ISV codes. 106 00:06:44,280 --> 00:06:49,680 And I have seen the same thing, which means that for a certain 107 00:06:49,680 --> 00:06:53,040 classification of workloads and I can only speak what I've seen 108 00:06:53,040 --> 00:06:55,760 and personally, so I'll speak more, let's say in plain design, 109 00:06:55,760 --> 00:06:59,200 car design, but I've seen it and I believe it to be so in other 110 00:06:59,200 --> 00:07:01,840 areas. It is. 111 00:07:02,280 --> 00:07:06,880 You're ultimately going to run your simulation cheaper on a GPU 112 00:07:06,880 --> 00:07:09,960 than a CPU. So forget about speed, it's 113 00:07:09,960 --> 00:07:11,680 going to be cheaper. And why is it cheaper? 114 00:07:12,120 --> 00:07:14,200 Ultimately, these are things like power efficiency, energy 115 00:07:14,200 --> 00:07:17,240 efficiency. So even though the GPU looks 116 00:07:17,240 --> 00:07:21,040 more expensive to buy for one node, when you actually look at 117 00:07:21,040 --> 00:07:24,520 how long does that node need to run for, you know, it's easier 118 00:07:24,600 --> 00:07:27,360 like I said, to calculate with cloud because it paid, pays you 119 00:07:27,360 --> 00:07:32,800 good pricing, it is cheaper, and as soon as something is cheaper 120 00:07:34,120 --> 00:07:38,320 economically it makes sense it is going to gain widespread 121 00:07:38,480 --> 00:07:42,280 industrial adoption. Now where's the new ones to this 122 00:07:42,800 --> 00:07:50,080 is what GPU there is in some ways a divergent, well, not 123 00:07:50,080 --> 00:07:55,800 divergent, but there is a, there is a market for machine learning 124 00:07:55,840 --> 00:08:01,800 GPUs with a particular focus on, you know, certain precision, 125 00:08:02,480 --> 00:08:05,840 lower precision ultimately than what most CFD codes would need. 126 00:08:07,000 --> 00:08:10,200 And so you're going, you know, from A1 hundreds to H1 hundreds, 127 00:08:10,200 --> 00:08:13,480 H2 hundreds, etcetera to Blackwell that are driving up, 128 00:08:13,480 --> 00:08:15,160 you know, more and more performance. 129 00:08:15,760 --> 00:08:19,080 And CFD absolutely can take advantage of that performance. 130 00:08:19,480 --> 00:08:23,040 And it's great because there's a, we can jump on the bandwagon 131 00:08:23,040 --> 00:08:26,640 essentially of machine learning. But there is also another 132 00:08:26,640 --> 00:08:31,840 classification of cards that quite often are more designed 133 00:08:31,840 --> 00:08:38,400 for inference, but these sort of L40SR TX type cards that 134 00:08:38,400 --> 00:08:43,600 actually have a very, very good price performance. 135 00:08:43,600 --> 00:08:47,680 So that raw performance may not be as high as those sort of more 136 00:08:47,920 --> 00:08:53,880 top end GPUs, but for CFD, they actually are often when you do 137 00:08:53,880 --> 00:08:55,720 the maths, even better price performance. 138 00:08:55,720 --> 00:08:57,880 And that's what we showed of this, this auto CFD. 139 00:08:58,960 --> 00:09:02,840 Now the reason I'm saying that I think the GPUs will become the 140 00:09:02,840 --> 00:09:08,160 main is because if you look at the trajectory of GPU 141 00:09:08,160 --> 00:09:12,000 development in terms of generation after generation 142 00:09:12,000 --> 00:09:15,920 improvement and you look at the same on the CPU front. 143 00:09:16,960 --> 00:09:20,000 I think it's hard to deny that again, because of the investment 144 00:09:20,000 --> 00:09:24,400 from machine learning, there is a much greater boost in 145 00:09:24,400 --> 00:09:28,920 performance from generation to generation on GPUs than there is 146 00:09:29,040 --> 00:09:32,440 on the X86 CPUs. And the fact that there is 147 00:09:32,440 --> 00:09:35,280 rising competition now between NVIDIA and AMD. 148 00:09:35,920 --> 00:09:39,840 And also maybe there's room for more, you know, startup or 149 00:09:39,840 --> 00:09:42,800 emergent technologies that's only going to drive more and 150 00:09:42,800 --> 00:09:47,440 more focus. Now, I say ARM 64 as well 151 00:09:47,640 --> 00:09:54,040 because that has emerged as a, at first maybe more of a niche 152 00:09:54,040 --> 00:09:57,080 architecture that was harder to get people to convince. 153 00:09:57,240 --> 00:10:00,840 And I know speaking from being AWS, we have our Graviton ones 154 00:10:00,840 --> 00:10:03,440 and you know, we struggled for people to think, well, why 155 00:10:03,440 --> 00:10:06,680 should I bother to pull my code, you know, to, to harm. 156 00:10:07,320 --> 00:10:10,440 But now with more and more companies with, you know, the 157 00:10:10,440 --> 00:10:13,520 other hyperscalers also releasing obviously with video, 158 00:10:13,520 --> 00:10:17,200 with the Gray system sort of linking the two, there is far 159 00:10:17,200 --> 00:10:22,960 more and there's far more reason for a major code to port to it. 160 00:10:23,400 --> 00:10:26,640 And as soon as that happens, the ecosystem jumps behind it. 161 00:10:27,400 --> 00:10:30,240 And again, the the power efficiency and the performance 162 00:10:30,240 --> 00:10:33,160 numbers make it extremely attractive. 163 00:10:35,120 --> 00:10:37,960 That's not to say that people aren't going to be running on, 164 00:10:38,080 --> 00:10:40,640 you know, AMD or Intel CPUs, of course. 165 00:10:40,800 --> 00:10:45,720 And as I said, the CFD market and engineering in general or 166 00:10:45,720 --> 00:10:51,560 any use of CFD tends to be slow. And so if from 2013 to now 167 00:10:52,400 --> 00:10:55,000 there's been that sort of movement, it's not going to 168 00:10:55,000 --> 00:10:58,760 happen overnight. But what I can see is every 169 00:10:58,760 --> 00:11:04,880 single one of the major Isvs has or is porting their codes. 170 00:11:06,760 --> 00:11:09,800 The issue is that not all the features in the codes are 171 00:11:09,800 --> 00:11:12,240 available. So yeah, some of the multi 172 00:11:12,240 --> 00:11:14,840 physics, some of the options are not yet done. 173 00:11:14,840 --> 00:11:18,880 So you may not be able to use the GPU version today because it 174 00:11:18,880 --> 00:11:21,160 may not have the feature that you need if you're doing some 175 00:11:21,160 --> 00:11:24,320 more sophisticated modeling, but it's probably on the road map. 176 00:11:24,960 --> 00:11:29,680 And more importantly is the codes. 177 00:11:30,160 --> 00:11:33,840 And these are more the start-ups like luminary or Volcano or Flex 178 00:11:33,840 --> 00:11:37,360 compute. And there are others who are 179 00:11:37,440 --> 00:11:42,080 really who have really focused on a GPU native code. 180 00:11:43,680 --> 00:11:49,480 And in the case of, for example, Volcano taking advantage of GP 181 00:11:49,520 --> 00:11:52,880 US like designing it for GP us Cartesian methods and there's 182 00:11:52,880 --> 00:11:55,880 boundary, etcetera. And there are others as well 183 00:11:55,880 --> 00:11:59,880 doing this and, you know, hiring super talented people who have 184 00:11:59,880 --> 00:12:03,080 come out of universities and institutes with a deep, deep 185 00:12:03,120 --> 00:12:06,840 understanding of how GPUs work. Programming it from the base 186 00:12:06,840 --> 00:12:09,800 from the beginning with the latest, you know, programming 187 00:12:09,800 --> 00:12:12,960 languages and and and knowledge of how to extract the maximum 188 00:12:12,960 --> 00:12:17,760 performance is really only going to accelerate it further. 189 00:12:18,160 --> 00:12:20,760 And I'll talk about this in the last point. 190 00:12:21,200 --> 00:12:26,240 This emergence of new start up codes is also invigorating 191 00:12:26,800 --> 00:12:32,120 because simply, if you look through the eyes of a commercial 192 00:12:32,120 --> 00:12:38,360 company, why change? If you are the main company, 193 00:12:38,360 --> 00:12:40,720 main ISV provider, why would you change? 194 00:12:41,240 --> 00:12:43,120 You're earning good money from your licenses. 195 00:12:43,640 --> 00:12:48,920 You are you've got customers and if your competitors are all 196 00:12:48,920 --> 00:12:52,560 basically doing the same thing, so none of them, let's say five 197 00:12:52,560 --> 00:12:54,880 years ago, six years ago, OK, none of them are moving to GPS. 198 00:12:54,880 --> 00:12:56,440 Well, why should we spend all the effort doing? 199 00:12:56,440 --> 00:12:58,480 Is it really that important? Is it really going to make a 200 00:12:58,480 --> 00:12:59,960 difference? Are we going to lose customers 201 00:13:00,240 --> 00:13:02,760 from it? And of course the technical 202 00:13:02,760 --> 00:13:04,840 reasons, you know, was there enough memory? 203 00:13:04,840 --> 00:13:07,080 Is there enough performance to gain from it? 204 00:13:09,320 --> 00:13:11,720 The one thing that drives a company to change is 205 00:13:11,720 --> 00:13:15,080 competition. And so with the emergence of 206 00:13:16,800 --> 00:13:21,360 alternative providers, many of whom who are specifically 207 00:13:21,360 --> 00:13:26,080 marketing their code around being GPU performance, it forces 208 00:13:26,120 --> 00:13:29,200 companies to do it. And you've seen that with, you 209 00:13:29,280 --> 00:13:32,120 know, with ancestors sort of rewriting their their solver and 210 00:13:32,320 --> 00:13:35,440 you know, showing a very, very strong performance. 211 00:13:35,640 --> 00:13:39,320 And at the auto CFD workshop, we saw pretty much all of the Isvs 212 00:13:40,280 --> 00:13:43,600 specifically focus on their GPU performance and their speed, 213 00:13:43,960 --> 00:13:47,520 which three years ago was not the case. 214 00:13:48,520 --> 00:13:51,920 So everything feels like the right momentum. 215 00:13:51,920 --> 00:13:56,680 Now the tricky thing is the whole ecosystem needs to to 216 00:13:56,680 --> 00:13:58,640 work. So is are the third party 217 00:13:58,640 --> 00:14:00,880 meshing tools going to work on GPS? 218 00:14:01,440 --> 00:14:04,560 And I mention this because we get to another point, which is 219 00:14:04,560 --> 00:14:08,760 around the sort of high performance computing, because 220 00:14:08,760 --> 00:14:13,160 obviously if you're going to procure a cluster, if one code 221 00:14:13,160 --> 00:14:16,280 works amazing on GPUs, but all the restaurant CPUs, you then 222 00:14:16,280 --> 00:14:19,880 need to figure out, well, what's my ratio of compute options to 223 00:14:19,880 --> 00:14:22,040 buy? And, and so this links it to one 224 00:14:22,040 --> 00:14:27,960 of the third points later that the other reason for GPUs is 225 00:14:29,160 --> 00:14:32,200 national supercomputers. OK, what do I mean by that? 226 00:14:32,200 --> 00:14:41,360 Well, if you are, let's say, in Spain and you want to show to 227 00:14:41,360 --> 00:14:46,480 your government that you as a country are pushing the 228 00:14:46,480 --> 00:14:49,440 boundaries of science, of R&D, and you're attracting 229 00:14:49,440 --> 00:14:52,480 entrepreneurial entrepreneurialism, 230 00:14:54,160 --> 00:14:56,160 supercomputing has traditionally been a great way. 231 00:14:56,160 --> 00:14:57,960 It's a sort of, you know, a big lighthouse. 232 00:14:58,000 --> 00:15:00,520 And the US is a more obvious one for that. 233 00:15:00,520 --> 00:15:05,640 But I'll pick a country other than the US, your supercomputer 234 00:15:07,000 --> 00:15:10,760 obviously a lot of that is then based on its performance and the 235 00:15:10,760 --> 00:15:14,360 top 500 list is often you know the metric like what is the 236 00:15:14,360 --> 00:15:18,080 total performance of the system. You know the the sort of exaflop 237 00:15:18,080 --> 00:15:20,040 race to try and get an ex scale computing. 238 00:15:21,560 --> 00:15:26,160 A lot of this has driven to GPU's because the total 239 00:15:26,160 --> 00:15:31,160 performance of the system is much higher for a given power 240 00:15:31,160 --> 00:15:38,080 output with GPU's, and arguably it is also newer technology. 241 00:15:38,200 --> 00:15:41,360 So more novelty, more differentiation can enable 242 00:15:41,360 --> 00:15:44,160 better science because there's more, you know, throughput can 243 00:15:44,160 --> 00:15:46,480 be produced. And secondly, even more 244 00:15:46,480 --> 00:15:49,160 importantly, secondly, is because of machine learning, 245 00:15:49,880 --> 00:15:53,600 because machine learning is now being integrated and is a key 246 00:15:53,600 --> 00:15:55,520 research topic as well. And so if you're a national 247 00:15:55,520 --> 00:15:58,800 super computing centre, you want to be able to do ML research and 248 00:15:58,800 --> 00:16:01,440 you want GPU's. What does that mean? 249 00:16:01,960 --> 00:16:06,320 It means that if you're an academic and you want access to 250 00:16:06,320 --> 00:16:09,200 these supercomputers to do your research, like people do in 251 00:16:09,200 --> 00:16:15,040 Spain, let's say, you will have to write your code to fit on the 252 00:16:15,040 --> 00:16:19,000 system. And so sure enough, that 253 00:16:19,000 --> 00:16:22,080 motivated researchers in Spain. And now let's broaden it to the 254 00:16:22,080 --> 00:16:26,440 UK, to the US, to port their code to GPUs. 255 00:16:27,120 --> 00:16:30,080 Essentially, they can run on the systems that they need to. 256 00:16:31,160 --> 00:16:32,960 And so there's been a push and pull. 257 00:16:33,640 --> 00:16:35,880 There's been a push because they think it is the right 258 00:16:35,880 --> 00:16:37,760 architecture to deliver better performance. 259 00:16:38,600 --> 00:16:41,000 Or maybe that's the pull. But there's also a push because 260 00:16:41,000 --> 00:16:42,920 the system they want to run on has GPS. 261 00:16:43,320 --> 00:16:47,560 So I should say that has also been the drive towards GPUs. 262 00:16:47,720 --> 00:16:51,320 There is a, once a decision is made that the big supercomputer 263 00:16:51,320 --> 00:16:54,440 are going to have GPUs, you almost are then having to, well, 264 00:16:54,600 --> 00:16:57,560 I'll just put my code, you know, to, to go to it. 265 00:16:58,000 --> 00:17:02,960 And so that's more the balance to the industrial side and the 266 00:17:02,960 --> 00:17:05,680 academic side, why both are now meeting. 267 00:17:05,920 --> 00:17:10,400 And again, remember, many of these startup codes emerged out 268 00:17:10,400 --> 00:17:13,800 of more academic or government work where they had moved to 269 00:17:13,800 --> 00:17:18,119 GPUs because one, they're just more aware of the the science. 270 00:17:18,359 --> 00:17:20,560 And I would say this is what companies like NVIDIA have also 271 00:17:20,560 --> 00:17:23,400 put a lot of groundwork in investing in helping people to 272 00:17:23,400 --> 00:17:26,079 understand CUDA and helping them to pull their codes. 273 00:17:26,520 --> 00:17:28,640 And that has now moved through into the startup. 274 00:17:29,200 --> 00:17:33,240 And as I'll speak about later, almost inevitably many of these 275 00:17:33,240 --> 00:17:35,400 startups are going to be acquired by the larger 276 00:17:35,760 --> 00:17:37,160 companies. There's an interesting sort of 277 00:17:37,200 --> 00:17:47,040 ecosystem and flywheel effect. So the next one is on AIML And 278 00:17:47,040 --> 00:17:50,880 you know, again, you're, you're typically really for AIML or 279 00:17:50,880 --> 00:17:54,360 you're really against it. This tends to be the case of any 280 00:17:54,360 --> 00:18:00,000 new technologies. The reason I'm more bullish on 281 00:18:00,000 --> 00:18:03,720 AML and that's through my own personal experience actually, at 282 00:18:03,760 --> 00:18:06,600 least I would encourage anybody else on, you know, looking at 283 00:18:06,600 --> 00:18:11,160 code, writing code, doing experiments, publishing, 284 00:18:11,160 --> 00:18:13,520 organizing, really trying to immerse themselves. 285 00:18:13,920 --> 00:18:18,800 And I would actually make it similar to the way that I saw 286 00:18:18,800 --> 00:18:22,840 GPUs back in, let's say the 2000 and 10s, where again, I saw a 287 00:18:22,840 --> 00:18:27,000 similar dynamic of some people really being against it and 288 00:18:27,000 --> 00:18:29,280 saying, oh, it's just all hype. You know, there's no way they're 289 00:18:29,280 --> 00:18:31,200 going to be able to do it. I'm not going to be able to 290 00:18:31,200 --> 00:18:32,880 pull. It's not, you know, the 291 00:18:32,880 --> 00:18:35,960 technology is not going to be there to others who are like, 292 00:18:36,160 --> 00:18:41,920 it's just a matter of when I see the AI in a similar way, but 293 00:18:41,920 --> 00:18:48,520 with some more nuances, which is what exactly can AI be used for 294 00:18:48,520 --> 00:18:51,520 within CFD? Well, the first point I would 295 00:18:51,520 --> 00:18:55,920 say is the commercial pull. And you know you may if you come 296 00:18:55,920 --> 00:18:58,880 from a more pure background, you may hate this sort of speak, but 297 00:18:58,880 --> 00:19:00,760 this is just the way the world works. 298 00:19:01,520 --> 00:19:05,320 When the large ISV companies, many of them who are 299 00:19:05,320 --> 00:19:10,480 stakeholders or boards to please people invest in their companies 300 00:19:10,480 --> 00:19:13,480 in the current climate because of perception of how much are 301 00:19:13,480 --> 00:19:16,600 they embracing and leading when it comes to innovation, 302 00:19:16,600 --> 00:19:22,440 particularly around AI. And so a major ISV has to show 303 00:19:22,440 --> 00:19:27,160 to their board and stakeholders and shareholders that they are 304 00:19:27,160 --> 00:19:30,240 fully embracing and are doing something about it. 305 00:19:31,080 --> 00:19:33,320 And I would say in some ways, many of them have had to catch 306 00:19:33,320 --> 00:19:34,880 up. Now, I don't know that, you 307 00:19:34,880 --> 00:19:37,640 know, working to the companies, maybe they've had many proposals 308 00:19:37,640 --> 00:19:40,000 beforehand. But certainly I would say the 309 00:19:40,000 --> 00:19:43,040 whole ChatGPT moment has, you know, caught some of them off 310 00:19:43,040 --> 00:19:46,720 guard and made them really have to double down and decide what 311 00:19:46,720 --> 00:19:48,680 to do and what and if they can release it. 312 00:19:49,240 --> 00:19:52,960 That's means that there's a lot of sort of R&D and focus within 313 00:19:52,960 --> 00:19:55,680 these major CFD and CE companies. 314 00:19:57,760 --> 00:20:02,560 But this is also fed a lot of startup market. 315 00:20:03,360 --> 00:20:10,200 So for a similar reason that many of the startups and there 316 00:20:10,200 --> 00:20:15,280 are many many to list, you know the Navistos, the Physics X 317 00:20:15,280 --> 00:20:17,920 beyond math. I'm probably going to insult 318 00:20:17,920 --> 00:20:20,840 people by not mentioning them. I can't think I'll probably list 319 00:20:20,840 --> 00:20:25,040 them more out newer concepts. Yeah, I can't remember. 320 00:20:25,040 --> 00:20:27,680 There's, there's loads of them. So I, I, I'm sure I'm offending 321 00:20:27,800 --> 00:20:32,040 some of them. The point is that or keyword as 322 00:20:32,040 --> 00:20:34,880 well. And yeah, OK, I'll try and make 323 00:20:34,880 --> 00:20:36,840 sure I mention most of them so I don't offend people. 324 00:20:37,920 --> 00:20:42,840 They have got VC funding because a lot of the VCs are seeing. 325 00:20:42,840 --> 00:20:45,040 Could this be the next big thing? 326 00:20:46,440 --> 00:20:50,440 And so there are to my Alaska at least 10 start-ups looking at 327 00:20:50,440 --> 00:20:54,000 doing AIML specifically for CFD. And quite often for things like 328 00:20:54,000 --> 00:20:55,680 I spoke about card design or plate design. 329 00:20:57,120 --> 00:20:59,920 The fact that they have a lot of money does not necessarily mean 330 00:20:59,920 --> 00:21:03,760 that it will turn into success, but it does show that there's a 331 00:21:03,760 --> 00:21:05,920 lot of concerted effort. And when there are hundreds of 332 00:21:05,920 --> 00:21:09,040 people or thousands of people collectively working on this, 333 00:21:09,480 --> 00:21:13,240 that often leads to advancement. And when I talk about 334 00:21:13,240 --> 00:21:16,840 advancement, I typically mean in terms of surrogate modelling. 335 00:21:17,120 --> 00:21:19,400 So what some people would call reduced order modelling, but 336 00:21:19,440 --> 00:21:21,880 obviously more sort of sophisticated version of that 337 00:21:22,240 --> 00:21:24,160 were essentially, and This is why it's particularly 338 00:21:24,160 --> 00:21:27,120 interesting from an ISV or commercial point where you may 339 00:21:27,120 --> 00:21:32,640 essentially not need the CFT solver during the solve step. 340 00:21:32,960 --> 00:21:36,600 So you train a model on some data and then once you have the 341 00:21:36,600 --> 00:21:38,520 model, you do inference and that's completely outside the 342 00:21:38,520 --> 00:21:43,680 CFT code, which obviously is worrying for an ISV that thinks 343 00:21:43,680 --> 00:21:47,880 we could be losing some of our, you know, potential license or 344 00:21:47,880 --> 00:21:50,600 revenue or, you know, customers because they could be trying to 345 00:21:50,600 --> 00:21:52,400 use this other tool that we don't have. 346 00:21:55,160 --> 00:21:58,880 So and just speaking commercially now that means 347 00:21:58,880 --> 00:22:02,120 those IS VS really need to figure out is this an important 348 00:22:02,120 --> 00:22:05,440 technology that could transform? And if so, we either need to 349 00:22:05,440 --> 00:22:08,360 develop it ourselves or we need to go and buy a start up. 350 00:22:08,960 --> 00:22:11,880 And that's really the the key thing that's driving a lot of 351 00:22:11,880 --> 00:22:14,680 these companies to then, you know, have a look at these 352 00:22:14,680 --> 00:22:19,560 start-ups. Why economically does it work as 353 00:22:19,560 --> 00:22:25,800 well is because on the one hand, the end user just as GPUs. 354 00:22:26,000 --> 00:22:28,200 So, so what's the reason for the end user to be interested in 355 00:22:28,200 --> 00:22:32,440 AIML for GPUs? The main reason was lower cost 356 00:22:32,880 --> 00:22:38,840 computing and potentially faster computing, you know, GPUs if if 357 00:22:38,840 --> 00:22:41,880 the pace continues, you know, we're going to see potential of 358 00:22:41,880 --> 00:22:45,040 the next five years 5 to 10X speed up in in stuff. 359 00:22:45,040 --> 00:22:48,880 So it not only is a lot faster and also talk a minute about 360 00:22:48,880 --> 00:22:51,640 this leads to more like high fidelity transit methods to 361 00:22:51,640 --> 00:22:57,400 become possible even in a day or less than a day for AIML. 362 00:22:57,640 --> 00:23:03,240 It's that the inference can be both cheap and fast and that is 363 00:23:03,240 --> 00:23:07,600 something that brings up a lot of potential for for companies. 364 00:23:07,600 --> 00:23:13,240 So for example, it means that real time CFD has long been 365 00:23:13,240 --> 00:23:17,440 spoken about and GPU's were sort of billed as potentially 366 00:23:17,440 --> 00:23:21,120 enabling real time CFD if you go back 5-10 years ago. 367 00:23:21,720 --> 00:23:24,520 But I would argue that that's not really turned out to be the 368 00:23:24,520 --> 00:23:27,760 case because some problems are not. 369 00:23:28,040 --> 00:23:31,600 They can be parallel in space, but maybe not parallel in time. 370 00:23:31,920 --> 00:23:34,520 So doesn't matter how much compute you throw at it, 371 00:23:34,520 --> 00:23:36,720 sometimes it just takes a while, particularly for transient 372 00:23:36,720 --> 00:23:39,360 methods which you may need to do for accuracy purposes. 373 00:23:39,760 --> 00:23:43,760 OK you could do maybe a Rand or lower fidelity, but truly real 374 00:23:43,760 --> 00:23:46,160 time click of a button instantly get the answer. 375 00:23:46,680 --> 00:23:51,280 You have to make so many accommodations on accuracy that 376 00:23:51,320 --> 00:23:55,720 it ends up really not being worth the sort of the balance. 377 00:23:56,520 --> 00:24:01,000 The AIML is the interesting bit because the inference itself 378 00:24:01,000 --> 00:24:04,840 could be real time, but you could spend a lot of money in 379 00:24:04,840 --> 00:24:08,360 computing time on training a model based on high fidelity 380 00:24:08,360 --> 00:24:12,320 data that could, if the ML model was accurate enough, give you 381 00:24:12,480 --> 00:24:14,680 quite a good answer in real time. 382 00:24:15,560 --> 00:24:21,240 And so this opens up a lot of potential, particularly in 383 00:24:21,240 --> 00:24:27,040 bringing CFD closer to the designer, closer to digital 384 00:24:27,040 --> 00:24:31,160 twins, more into, you know, like Nvidia's doing with Omniverse, 385 00:24:31,160 --> 00:24:35,040 into a sort of virtual world where you could be simulating 386 00:24:35,040 --> 00:24:39,480 things almost in real time. The class example is the car 387 00:24:39,480 --> 00:24:43,000 designer, you know who's not ACFD specialist who wants to 388 00:24:43,000 --> 00:24:48,280 draw a car, click a button, see it, get a sense of what the drag 389 00:24:48,280 --> 00:24:53,720 is and then continue iterating. Now I should say this is not in 390 00:24:53,720 --> 00:24:58,520 the short term or even probably the medium term about replacing 391 00:24:59,240 --> 00:25:03,120 high fidelity CFD. But I do think it has a very 392 00:25:03,280 --> 00:25:06,720 potentially disruptive role in the lower fidelity conceptual 393 00:25:06,720 --> 00:25:11,720 design phases that could really be important. 394 00:25:11,720 --> 00:25:16,080 And so commercially, this could mean basically a key tool in the 395 00:25:16,080 --> 00:25:20,560 market and this is driving a lot of start-ups because there is a 396 00:25:20,560 --> 00:25:24,280 sense that there could be acquisitions or mergers from 397 00:25:24,280 --> 00:25:28,560 some of the big Isvs or they could simply become big codes. 398 00:25:28,560 --> 00:25:33,960 And the CA market is five, $10 billion at least and growing. 399 00:25:34,880 --> 00:25:37,680 And it's growing because there is a desire for digital 400 00:25:37,680 --> 00:25:41,560 certification. There's a desire to go more into 401 00:25:41,560 --> 00:25:44,400 the virtual world because it's cheaper and faster than winter 402 00:25:44,400 --> 00:25:49,920 testing or physical testing. So now there is still, and I 403 00:25:50,120 --> 00:25:55,200 want to make this very clear in my mind, many challenges and 404 00:25:55,200 --> 00:25:59,240 many areas that will most likely change dramatically in this 405 00:25:59,240 --> 00:26:02,560 space. I do not believe that any of the 406 00:26:02,560 --> 00:26:08,520 current methods published today will be the method that we end 407 00:26:08,520 --> 00:26:13,160 up using in five years time. I could be wrong and maybe it'll 408 00:26:13,160 --> 00:26:16,440 only just be iterations of it, but the field is moving so 409 00:26:16,440 --> 00:26:19,880 quickly that I just find it hard to believe that what we have 410 00:26:19,880 --> 00:26:23,480 today is the best we can do. Essentially, and I say this 411 00:26:23,480 --> 00:26:26,520 because if you look at the CFD world, actually a lot of the 412 00:26:26,520 --> 00:26:30,000 breakthroughs in terms of the physics or the modelling were 413 00:26:30,000 --> 00:26:33,320 sort of done 20 years ago. You know, most of the warmer LES 414 00:26:33,320 --> 00:26:36,560 or hybrid random LES or random modelling, it's actually 20 415 00:26:36,560 --> 00:26:39,440 years old. And what's the novel bit is more 416 00:26:40,000 --> 00:26:43,000 the computer science, the use of GPU's, etcetera. 417 00:26:43,000 --> 00:26:46,280 But the actual breakthrough modelling, mathematical 418 00:26:46,280 --> 00:26:48,640 modelling was done 20-30 years ago. 419 00:26:49,000 --> 00:26:51,840 There's obviously some new developments, but most people, 420 00:26:52,320 --> 00:26:55,920 if you're using DDS, you know, what's that 2004 or something, 421 00:26:56,080 --> 00:26:57,680 depending on which version you're using of it. 422 00:26:58,120 --> 00:27:02,240 If you're using K Omega, SST RANS, it's 1994. 423 00:27:02,240 --> 00:27:06,920 So what's that 30 years ago? So it's a bit like going back to 424 00:27:06,920 --> 00:27:09,520 1994. The K Omega comes out and it 425 00:27:09,520 --> 00:27:11,080 gives good results. I'm thinking that's it. 426 00:27:11,800 --> 00:27:15,360 I think we're in similar time, the AML that yes, there are many 427 00:27:15,360 --> 00:27:18,040 startups and companies, but I don't think the actual modelling 428 00:27:18,040 --> 00:27:20,560 or the accuracy today is anything like we'll see in the 429 00:27:20,560 --> 00:27:22,480 future. And so that means there could be 430 00:27:22,480 --> 00:27:24,040 a lot of change, a lot of disruption. 431 00:27:24,040 --> 00:27:25,960 And that's why I think it's a really important area to get 432 00:27:25,960 --> 00:27:29,600 into because there's no doubt there has still been major 433 00:27:29,600 --> 00:27:34,760 issues improving scale. Can these AI methods, you know, 434 00:27:34,760 --> 00:27:37,160 work on hundreds of millions of cell meshes? 435 00:27:38,000 --> 00:27:40,800 Can they work in terms of the accuracy? 436 00:27:40,800 --> 00:27:45,200 You know, have they really shown how accurate they are down to, 437 00:27:45,200 --> 00:27:46,560 you know, a drag count difference? 438 00:27:46,560 --> 00:27:51,240 There's been a lot of studies, but there's been limited real 439 00:27:51,240 --> 00:27:54,560 concrete evidence of it away from just sort of marketing. 440 00:27:55,240 --> 00:28:00,200 And the big question, of course, is around foundational models. 441 00:28:00,200 --> 00:28:04,040 So most of what the startups and most people who talk about it 442 00:28:04,040 --> 00:28:08,760 are doing is saying, we'll take the data you have as a company X 443 00:28:08,920 --> 00:28:12,200 train a model, you train a model essentially, and then you do 444 00:28:12,200 --> 00:28:17,400 imprints on it as opposed to ChatGPT like methods that take 445 00:28:18,680 --> 00:28:22,240 huge data from the Internet. And now by paying providers, 446 00:28:22,240 --> 00:28:26,200 they train a model for, you know, months and then all you're 447 00:28:26,200 --> 00:28:29,280 doing is using it as inference. They've done the training and 448 00:28:29,280 --> 00:28:31,720 they're recouping the cost from that through, you know, 449 00:28:31,720 --> 00:28:33,920 subscription model or pay as you go through the inference. 450 00:28:34,160 --> 00:28:36,520 Whereas in CFD we're basically saying you have to do the 451 00:28:36,520 --> 00:28:40,520 training with your own data. The future is could there be a 452 00:28:40,520 --> 00:28:43,240 way of a company pre training model, then you just do 453 00:28:43,240 --> 00:28:45,800 inference. But and I'm, I, I will 454 00:28:45,800 --> 00:28:48,000 definitely do a more of a deep dive on this. 455 00:28:48,000 --> 00:28:51,760 And hopefully if you look in season 1, I've spoke to, you 456 00:28:51,760 --> 00:28:55,640 know, 3 or 4 top ML people that, you know, I think they largely 457 00:28:55,640 --> 00:28:57,840 agree that this is a super exciting space. 458 00:28:57,840 --> 00:29:01,440 But there's still many questions and uncertainties around 459 00:29:01,440 --> 00:29:03,120 including physics, not including physics. 460 00:29:03,120 --> 00:29:05,880 And I go back to the statement that I just don't think the 461 00:29:05,880 --> 00:29:10,480 methods out today are the ones that we will use. 462 00:29:10,480 --> 00:29:13,640 But the breakthrough moment could be any time. 463 00:29:13,920 --> 00:29:16,640 And that's why I think it's such a super exciting space. 464 00:29:16,640 --> 00:29:20,480 So, so the second one is AML. And of course, why do I mention 465 00:29:20,480 --> 00:29:21,600 AML? Because it's linked. 466 00:29:21,600 --> 00:29:26,840 The first AIML needs GPUs and as a company if you're being asked 467 00:29:26,840 --> 00:29:31,680 to train ML models then GPU is the only way to do it and 468 00:29:31,680 --> 00:29:35,520 therefore it makes double the sense to procure a GPU type 469 00:29:35,520 --> 00:29:39,800 cluster or predominantly GPU cluster because it can be used 470 00:29:39,800 --> 00:29:44,520 for your ML and your CFD. Now the third one, which again 471 00:29:44,520 --> 00:29:51,000 in I believe takes its evidence from these first 2 is cloud 472 00:29:51,400 --> 00:29:59,720 computing or SAS type products. And again, I have to be, you 473 00:29:59,720 --> 00:30:02,120 know, honest I, I'm clearly working for a cloud computing 474 00:30:02,120 --> 00:30:07,000 company, but my belief is independent of working for, for 475 00:30:07,000 --> 00:30:09,720 for that company. And I think, and I hopefully I 476 00:30:09,720 --> 00:30:11,680 can give some evidence to back that up. 477 00:30:13,160 --> 00:30:19,400 The reason I'm saying this is it is undeniably true that the 478 00:30:19,400 --> 00:30:23,880 hyperscalers have huge investments in compute and are 479 00:30:23,880 --> 00:30:30,640 able to give access to compute that most companies may not be 480 00:30:30,640 --> 00:30:36,520 able to get at the size that they want and the frequency and 481 00:30:36,520 --> 00:30:41,440 the access and the variety. And now there may be times when 482 00:30:41,440 --> 00:30:43,840 there's a mixture of the two. And again, it's an evolving 483 00:30:43,840 --> 00:30:45,840 thing. If you already have, you know, a 484 00:30:45,840 --> 00:30:51,000 lot of on Prem work, then maybe the cloud for you is a sort of a 485 00:30:51,000 --> 00:30:55,960 burst option or you know, a backup option, or it could be a 486 00:30:55,960 --> 00:30:57,320 primary option. There's, you know, there's 487 00:30:57,320 --> 00:30:59,680 different, different models and this will evolve over time. 488 00:31:00,440 --> 00:31:05,800 But the fact that major, major companies around the world are 489 00:31:05,800 --> 00:31:10,280 moving to the cloud because the way they see it is less of a 490 00:31:10,280 --> 00:31:13,680 what's the cost of the compute, what's the cost of the hardware, 491 00:31:14,240 --> 00:31:16,240 but more what can people do with it? 492 00:31:17,960 --> 00:31:22,200 The point being is that there's a lot of the biggest 493 00:31:22,280 --> 00:31:25,520 misconception of the cloud is that it's just service to rent. 494 00:31:26,440 --> 00:31:29,560 That's maybe the way it was 10 years ago. 495 00:31:30,160 --> 00:31:33,280 But now there's so many high level services added by all the 496 00:31:33,280 --> 00:31:37,680 cloud providers that actually offering is not only to access 497 00:31:37,680 --> 00:31:42,760 the compute, but many services that makes it easier to do 498 00:31:42,760 --> 00:31:45,400 things, whether it's machine learning, whether it's, you 499 00:31:45,400 --> 00:31:49,040 know, high performance computing or whether it's databases or 500 00:31:49,040 --> 00:31:54,360 virtual or end user computing, many, many, many things that 501 00:31:54,440 --> 00:31:59,200 makes it easier for your end users to do stuff or for your 502 00:31:59,200 --> 00:32:01,920 sort of IT teams to do things that they would have to have 503 00:32:01,920 --> 00:32:04,160 spent more time doing. So it's making them more 504 00:32:04,160 --> 00:32:06,280 efficient. So often when people look at a 505 00:32:06,280 --> 00:32:09,880 sort of a cost thing, the cost only really works when you take 506 00:32:09,880 --> 00:32:14,440 the value of doing everything. And that's why many, many 507 00:32:14,440 --> 00:32:17,840 companies are driving this. And the reality of what I've 508 00:32:17,840 --> 00:32:23,920 seen at least is the top down is really pushing this. 509 00:32:23,920 --> 00:32:28,240 So the CI OS driving this. And so it's becoming, you know, 510 00:32:28,760 --> 00:32:31,000 a reality for many, many companies. 511 00:32:31,240 --> 00:32:35,360 But like one and two cloud for many companies is still a 512 00:32:35,360 --> 00:32:38,080 journey. And so they're not fully moving 513 00:32:38,080 --> 00:32:40,560 to it. But the reason it it's appealing 514 00:32:41,400 --> 00:32:46,960 is because if you want to get access to quite a different 515 00:32:46,960 --> 00:32:51,000 number of GPUs or you want to test out new GPUs, the reality 516 00:32:51,000 --> 00:32:52,360 is it's probably the easiest thing. 517 00:32:52,360 --> 00:32:55,160 If you want to go and test GPU, it's probably easy just to, you 518 00:32:55,160 --> 00:32:57,960 know, get an account on a cloud provider and do some testing. 519 00:32:58,320 --> 00:33:00,040 Now, of course there are challenges sometimes with 520 00:33:00,040 --> 00:33:02,920 capacity and things like that, but that's more that I guess a 521 00:33:02,920 --> 00:33:05,680 transient effect. Given the still the things of 522 00:33:05,680 --> 00:33:10,120 COVID and and supply chains, it makes sense to procure it 523 00:33:10,240 --> 00:33:12,000 through a, through a cloud system. 524 00:33:12,000 --> 00:33:15,560 And you're seeing that, you know, with, with most companies. 525 00:33:16,560 --> 00:33:20,040 Now, the reason I'm sort of linking that to the others is 526 00:33:21,600 --> 00:33:24,760 there was also a time when many of the Isvs and companies 527 00:33:24,760 --> 00:33:26,960 start-ups said, oh, the cloud, you know, it's just fad. 528 00:33:28,120 --> 00:33:31,960 But now many of them have turned around because they see it as an 529 00:33:31,960 --> 00:33:34,560 easier way of delivering their software to you. 530 00:33:35,200 --> 00:33:37,760 And this is where the SAS comes in, Software as a service. 531 00:33:38,040 --> 00:33:40,440 Now we're used to software as a service because many 532 00:33:40,440 --> 00:33:44,760 applications we use through an app, we don't, we just go to a 533 00:33:44,760 --> 00:33:47,240 website and we use it. That is software as a service. 534 00:33:47,560 --> 00:33:51,160 So, you know, one option used to be that you would download and 535 00:33:51,160 --> 00:33:56,320 install something on your computer or you can go on to 536 00:33:56,320 --> 00:33:59,160 something like, you know, over leaf I use for like writing 537 00:33:59,160 --> 00:34:01,920 papers. It's a web-based platform where 538 00:34:01,920 --> 00:34:04,640 you can write Latex documents and it's all stored there in the 539 00:34:04,640 --> 00:34:06,040 cloud. You don't have to download 540 00:34:06,040 --> 00:34:10,000 anything locally. Now that makes it a much easier, 541 00:34:10,000 --> 00:34:12,960 nicer experience and it gives access to everything without you 542 00:34:12,960 --> 00:34:15,760 having to have it. The analogy would be for the CFD 543 00:34:15,760 --> 00:34:18,920 side is the current way of doing it is you would take a piece of 544 00:34:18,920 --> 00:34:23,679 software, a binary, let's say, or open source and you get the 545 00:34:23,679 --> 00:34:26,800 software from them, but you have to do all the the hardware. 546 00:34:26,800 --> 00:34:29,239 So you wouldn't store it on your local computer, on your 547 00:34:29,239 --> 00:34:30,960 hardware. But then what if you want a 548 00:34:30,960 --> 00:34:33,520 bigger computer? What if you need more memory? 549 00:34:34,120 --> 00:34:36,960 Well, you'd have to have to go and buy a bigger computer. 550 00:34:38,080 --> 00:34:42,440 What if instead you could go through some website, access 551 00:34:42,440 --> 00:34:47,560 that same CFD software or some, you know, through website or 552 00:34:47,560 --> 00:34:51,159 application, and it sorted out the compute for you. 553 00:34:51,320 --> 00:34:53,800 So you just ran the simulation and in the back end it was 554 00:34:53,800 --> 00:34:58,040 provisioning a bigger GPU or a bigger CPU, depending on what 555 00:34:58,040 --> 00:34:59,320 you. You never needed to worry about 556 00:34:59,320 --> 00:35:04,120 that. That is the value of the SAS and 557 00:35:04,120 --> 00:35:06,120 it links to the cloud because they're all running on the 558 00:35:06,120 --> 00:35:09,360 cloud. So it gives the cloud gives 559 00:35:09,360 --> 00:35:11,920 those companies the ability that they don't need to go and buy 560 00:35:11,920 --> 00:35:14,280 data centres, They don't need to go and buy a compute. 561 00:35:14,280 --> 00:35:16,080 They can just pay a cloud provider. 562 00:35:16,640 --> 00:35:19,960 And so you are are getting that and they're using the cloud. 563 00:35:20,680 --> 00:35:25,080 And that's essentially how most the two ways that we see and I 564 00:35:25,080 --> 00:35:29,680 see of using the cloud, you can either use the cloud and it's 565 00:35:29,680 --> 00:35:36,400 essentially just a mechanism to get compute in a in a remote 566 00:35:36,400 --> 00:35:38,080 fashion. And you don't really notice it. 567 00:35:38,080 --> 00:35:41,040 You just SSH one to machine. You don't know whether it's in 568 00:35:41,040 --> 00:35:43,440 your cloud account or where it's in some physical data centre 569 00:35:43,440 --> 00:35:45,640 that you own. And of course you can use high 570 00:35:45,640 --> 00:35:48,680 level services. And the advantage of that one is 571 00:35:48,680 --> 00:35:50,520 you keep everything, let's say in your account. 572 00:35:50,520 --> 00:35:52,800 So you have a native account, all your data's in there. 573 00:35:52,800 --> 00:35:57,240 You're a big enterprise, but you have to know how to use Atbus, 574 00:35:57,240 --> 00:35:58,680 You have to know how to do all the cloud. 575 00:35:59,320 --> 00:36:01,640 The alternative option is you just say, well, I want to use 576 00:36:01,640 --> 00:36:04,920 ISV software. A that's just, I'm going to say, 577 00:36:04,920 --> 00:36:08,320 let's say, you know, star CCM has a SAS like, you know, a 578 00:36:08,320 --> 00:36:13,200 version, you can just go in and then use that or could be any 579 00:36:13,200 --> 00:36:15,200 other ISV. I'm just using it as an example. 580 00:36:16,880 --> 00:36:20,000 You can go in and pick the platform that you want, the 581 00:36:20,000 --> 00:36:23,360 compute that you want go in and they provision it for you. 582 00:36:23,360 --> 00:36:26,400 And so you've not had to know any ADDF knowledge to get access 583 00:36:26,400 --> 00:36:33,360 to that compute that is going to become more and more the 584 00:36:33,360 --> 00:36:36,360 mechanic you look at luminary cloud cloud is in the offering, 585 00:36:36,400 --> 00:36:40,120 right. They are a cloud first provider 586 00:36:40,120 --> 00:36:44,560 of CFD with the exact thing I just said that it makes it 587 00:36:44,600 --> 00:36:46,920 easier for you to run stuff without having to worry about 588 00:36:46,920 --> 00:36:49,560 the compute. There are many of like SIM scale 589 00:36:49,560 --> 00:36:53,040 or flex compute or there's, there's more and more, whether 590 00:36:53,040 --> 00:36:57,360 it's startup or an ISV that are offering cloud based back end. 591 00:36:57,360 --> 00:37:03,480 And that's why it's can be, in my mind, a very positive thing 592 00:37:03,480 --> 00:37:06,600 because it's also, let's say, forget about the commercials. 593 00:37:06,600 --> 00:37:09,680 I want you an academic code. If you want someone to use your 594 00:37:09,680 --> 00:37:13,400 code or to give access to it, actually it might be great if 595 00:37:13,400 --> 00:37:17,040 you build your own sort of SAS platform, you can give people 596 00:37:17,040 --> 00:37:18,760 access to your code or collaborate. 597 00:37:18,760 --> 00:37:20,680 And it could be private. Remember, it doesn't have to be 598 00:37:20,680 --> 00:37:22,120 fully public. You could just do it through a 599 00:37:22,120 --> 00:37:24,640 username password and they log on and you can collaborate. 600 00:37:25,120 --> 00:37:28,560 There are many sort of openings to this that haven't been fully 601 00:37:28,560 --> 00:37:31,720 realized. People can create codes and 602 00:37:31,720 --> 00:37:36,480 products and serve them to people without being a sort of 603 00:37:36,720 --> 00:37:39,160 delivering ACD to someone, you know, that they have to install 604 00:37:39,160 --> 00:37:43,480 on the computer, the cloud we use everyday. 605 00:37:43,480 --> 00:37:46,200 And it, it is a way of delivering things to you, 606 00:37:46,200 --> 00:37:51,360 software to you that CFD has to date not fully embraced. 607 00:37:51,360 --> 00:37:55,200 But that is something a bit like GPS and AML, which will only 608 00:37:55,200 --> 00:38:00,000 increase. And I personally think it's a 609 00:38:00,000 --> 00:38:04,040 good thing and it's something that you can explore yourself as 610 00:38:04,040 --> 00:38:08,280 again, not saying that people buying their own desktops or 611 00:38:08,280 --> 00:38:10,480 their own data centers is going to stop. 612 00:38:10,480 --> 00:38:14,200 Of course it's going to continue, but I think the 613 00:38:14,200 --> 00:38:18,040 percentage and the data shows us out that the growth of HPC for 614 00:38:18,040 --> 00:38:22,320 the cloud is something like 15% year over year, whereas HPC on 615 00:38:22,320 --> 00:38:25,040 premise like 8%. And I think that's Hyperion 616 00:38:25,160 --> 00:38:28,440 third party analyst. The data shows it is growing. 617 00:38:28,440 --> 00:38:32,160 And so yeah, in this episode it's more making you aware of 618 00:38:32,160 --> 00:38:33,560 the trends. Whether you agree with them or 619 00:38:33,560 --> 00:38:37,600 not doesn't really matter. It's it's they are in my opinion 620 00:38:37,600 --> 00:38:41,400 anyway, the trends that that are emerging now. 621 00:38:41,400 --> 00:38:46,480 What do these three things mean? GPU technology, AIML cloud. 622 00:38:47,320 --> 00:38:54,800 In my mind, it means that companies can move to higher 623 00:38:54,800 --> 00:39:03,160 fidelity transient methods. Or if your use case doesn't 624 00:39:03,160 --> 00:39:07,320 allow you to do, let's say you know, scale resolving type 625 00:39:07,320 --> 00:39:10,480 simulation, it could just mean bigger meshes. 626 00:39:10,520 --> 00:39:14,000 If it's still steady state it, it could mean running the 627 00:39:14,000 --> 00:39:16,520 simulation for longer conversion to even tighter tolerance. 628 00:39:16,520 --> 00:39:19,680 It could use more equations for your chemistry model. 629 00:39:19,680 --> 00:39:24,120 You know more physics in it. As soon as you have faster, more 630 00:39:24,120 --> 00:39:27,960 compute and access to more of it, let's say through a cloud 631 00:39:27,960 --> 00:39:31,040 like platform, you are going to see more use of it. 632 00:39:31,120 --> 00:39:33,560 And why does this matter? Because ultimately it's about 633 00:39:33,560 --> 00:39:35,760 accuracy. The biggest problem for CFD and 634 00:39:35,760 --> 00:39:41,960 trust for CFD by very senior peoples, people in companies or 635 00:39:41,960 --> 00:39:45,440 by academic research is, is accuracy compared to physical 636 00:39:45,440 --> 00:39:47,440 testing, right. That's always going to be the 637 00:39:47,440 --> 00:39:49,440 put down. The CFD looks good, but how 638 00:39:49,440 --> 00:39:51,840 accurate is it? Well, as we all know, accuracy 639 00:39:51,840 --> 00:39:56,360 is heavily linked to compute. And that's why I think this 640 00:39:56,360 --> 00:40:03,640 really will set up a new wave of move to digital certification. 641 00:40:04,080 --> 00:40:06,920 Digital certification, or certification by analysis has 642 00:40:06,920 --> 00:40:09,000 long been a goal for aircraft and. 643 00:40:09,520 --> 00:40:13,680 Car companies in any industry, but I would argue that compute 644 00:40:14,400 --> 00:40:18,360 and the method you can use is a limiting factor. 645 00:40:19,200 --> 00:40:22,840 Well, once the compute starts to become relaxed, the methods you 646 00:40:22,840 --> 00:40:25,800 can use better methods, which gives you more accuracy and 647 00:40:25,800 --> 00:40:30,520 gives you more evidence to do less physical testing or even if 648 00:40:30,520 --> 00:40:32,440 you do the same physical testing, just more virtual 649 00:40:32,440 --> 00:40:35,560 testing. And that will deliver, in my 650 00:40:35,560 --> 00:40:39,440 opinion, return on investment because companies can shorten 651 00:40:39,440 --> 00:40:44,640 that time to release a product or science can be done at an 652 00:40:44,640 --> 00:40:47,280 even greater scale. So, you know, DNS can be done at 653 00:40:47,640 --> 00:40:51,720 much better resolutions largely simulations can be done for more 654 00:40:51,720 --> 00:40:54,600 realistic problems. You can add more, you know, 655 00:40:54,720 --> 00:40:57,320 multi physics modelling inside case because you have more 656 00:40:57,320 --> 00:41:01,960 compute, you can do multi physics, so many more things. 657 00:41:03,320 --> 00:41:08,640 So I think whether it's more physics or just simply things 658 00:41:08,640 --> 00:41:11,520 like small time steps, laws and meshes, all of this will be more 659 00:41:11,520 --> 00:41:16,240 accuracy and I really do think that is going to drive greater 660 00:41:16,240 --> 00:41:20,600 use of CFD and CE. So I think those markets are 661 00:41:20,600 --> 00:41:24,840 only going to increase because of 1-2 and three the adoption. 662 00:41:24,840 --> 00:41:26,440 This is not in my opinion static. 663 00:41:26,440 --> 00:41:29,600 This is a growing and will continue to grow because it's 664 00:41:29,600 --> 00:41:33,760 value to a business will be beyond just pretty pictures and 665 00:41:33,760 --> 00:41:38,720 actually meaningful improvements in, in, in design and making 666 00:41:38,720 --> 00:41:42,360 things safer, hopefully safe and faster better than the usual 667 00:41:42,360 --> 00:41:46,640 metrics. So what's my final fifth point 668 00:41:46,640 --> 00:41:52,960 to this is this means in my mind mergers and acquisitions, which 669 00:41:53,680 --> 00:41:57,560 is sort of interesting. And again, it's a slightly more 670 00:41:58,080 --> 00:41:59,960 commercial angle, but I think this is interesting for any 671 00:42:00,160 --> 00:42:03,480 people who are looking to find start-ups or you know, people 672 00:42:03,480 --> 00:42:06,560 interesting inside of space. Whenever you have destructive 673 00:42:06,600 --> 00:42:09,440 technologies like this, whenever you have the option of new tech, 674 00:42:09,440 --> 00:42:13,800 new hardware, the emergence of AIML, the emergence of SAS 675 00:42:13,800 --> 00:42:17,840 providing and the the thing that has and high fidelity transit 676 00:42:17,840 --> 00:42:22,400 methods, it means it's perfect time for start-ups and we are 677 00:42:22,400 --> 00:42:25,640 seeing the start-ups. Why now in the past two years 678 00:42:25,640 --> 00:42:28,720 what why didn't we see you know, alumina cloud of volcano 679 00:42:28,720 --> 00:42:31,600 effects, compute, etcetera 5-7 years ago? 680 00:42:32,000 --> 00:42:35,280 Because I believe because of the GPU technology, because of the 681 00:42:35,280 --> 00:42:37,600 cloud it it's set the right environment. 682 00:42:37,880 --> 00:42:41,000 Why are we seeing 10 start-ups now with AIML methods? 683 00:42:41,440 --> 00:42:43,560 Well, part of it is the technology, but it's also the 684 00:42:43,560 --> 00:42:46,160 environment because of the huge Gen. 685 00:42:46,160 --> 00:42:50,800 AI boom that's giving reason for these VCs to fund these sort of 686 00:42:50,800 --> 00:42:53,600 companies. Well, as soon as you have lots 687 00:42:53,600 --> 00:42:59,320 and lots of startups and lots of main players who own majority of 688 00:42:59,320 --> 00:43:01,840 the market, that means there's going to be mergers and 689 00:43:01,840 --> 00:43:03,720 acquisitions. For sure. 690 00:43:04,680 --> 00:43:07,920 The big companies will acquire the smaller companies to keep 691 00:43:08,440 --> 00:43:13,280 market, but there is always the potential for a disruptor that a 692 00:43:13,280 --> 00:43:17,840 company does not just get bought out but actually rises to really 693 00:43:17,840 --> 00:43:21,400 challenge these companies. I don't know which one it is 694 00:43:21,400 --> 00:43:27,000 because the to date you know the large companies like Cadence and 695 00:43:27,000 --> 00:43:31,040 Synopsis, the big EDA companies and people don't know EDA is 696 00:43:31,480 --> 00:43:33,840 probably even bigger market than than the CAE. 697 00:43:34,160 --> 00:43:37,400 They bought in to those. So you know Cadence have 698 00:43:37,400 --> 00:43:40,680 acquired a number of companies, Synopsis have acquired Ansys and 699 00:43:40,680 --> 00:43:43,800 Ansys have acquired a number of companies, Siemens obviously 700 00:43:44,520 --> 00:43:48,640 Dasseau. So you have these players who 701 00:43:48,640 --> 00:43:53,800 see the bigger picture, which is the digital certification, which 702 00:43:53,800 --> 00:43:57,360 is the digital twins, which is the, so the way that it can 703 00:43:57,360 --> 00:44:01,400 retransform manufacturing and the engineering space. 704 00:44:02,960 --> 00:44:06,040 And for them, they see that simulation is a key bit of this. 705 00:44:06,600 --> 00:44:10,040 And this is really the, the sort of the big picture that there is 706 00:44:10,160 --> 00:44:15,640 A, it's a fantastic time in my opinion, to be in CFD and C it's 707 00:44:15,640 --> 00:44:19,520 a fantastic time because there is real value that can be 708 00:44:19,520 --> 00:44:21,960 created. And the dream of someone sat 709 00:44:21,960 --> 00:44:25,040 there and, you know, virtually designing something and testing 710 00:44:25,040 --> 00:44:27,200 it and seeing it in like a, you know, through some sort of 711 00:44:27,200 --> 00:44:30,880 virtual glasses. It seems to be on the horizon. 712 00:44:30,880 --> 00:44:34,920 So I think the next 5-10 years we'll see two things. 713 00:44:35,720 --> 00:44:37,560 One, people doing the same they've always done. 714 00:44:37,760 --> 00:44:39,320 But I see it's true, it doesn't matter. 715 00:44:39,360 --> 00:44:41,760 There are some companies doing the same sort of modelling they 716 00:44:41,760 --> 00:44:44,920 did 10 years ago. And fine, if there is a listing 717 00:44:44,920 --> 00:44:46,960 technology that's nice, but I'm just going to keep doing it the 718 00:44:46,960 --> 00:44:50,000 way I'm going to do it. For some that will happen, but I 719 00:44:50,000 --> 00:44:54,600 think for others there will be a dramatic change, an acceleration 720 00:44:54,600 --> 00:44:57,160 and, and I think this is only a good thing. 721 00:44:57,840 --> 00:45:03,360 So that was my five main trends. I would love to know whether you 722 00:45:03,360 --> 00:45:05,320 agree. It is totally just an opinion. 723 00:45:05,320 --> 00:45:08,240 I'm sure I've missed many things out and I'm happy to go into 724 00:45:08,240 --> 00:45:11,640 more details and other things and be corrected if I was wrong 725 00:45:11,640 --> 00:45:13,440 and some. But I hope you found this 726 00:45:14,040 --> 00:45:16,880 interested and interesting and it's got some of your brain, you 727 00:45:16,880 --> 00:45:20,800 know, thinking of does this apply to me and my thing is this 728 00:45:20,800 --> 00:45:23,320 just a more niche thing to certain segments. 729 00:45:24,800 --> 00:45:27,400 But yeah, I hope you enjoyed it. We've got some great guests 730 00:45:27,400 --> 00:45:31,440 coming up over the next couple of months, which I'm super 731 00:45:31,440 --> 00:45:34,840 excited by. So I don't really like to say 732 00:45:34,840 --> 00:45:37,360 this too much, but it apparently it helps. 733 00:45:37,680 --> 00:45:41,360 So if you are watching this on YouTube, if you sort of like and 734 00:45:41,360 --> 00:45:44,680 subscribe, it helps the YouTube algorithms and also lets you 735 00:45:44,680 --> 00:45:47,680 know when there's something. But also if you're on Spotify, 736 00:45:47,680 --> 00:45:50,440 Apple, you can I think you can click one of the buttons to to 737 00:45:50,440 --> 00:45:54,160 follow the podcast so that when something new comes out, you'll 738 00:45:54,160 --> 00:45:58,360 get you'll get notified. All right, thanks very much. 739 00:45:58,640 --> 00:46:00,280 Hope you enjoyed it. See you soon. 740 00:46:23,320 --> 00:46:23,400 The.