WEBVTT 00:00.000 --> 00:10.240 A number of weeks or months, depending on how much data you have in order to get a model. 00:10.240 --> 00:17.280 But really, this will not get you rich, this will give you a big electricity bill, a big 00:17.280 --> 00:18.280 hardware bill. 00:18.280 --> 00:23.640 But what will get you rich is the actual inference when people are querying that a 00:23.640 --> 00:27.280 base or doing some searches for something. 00:27.280 --> 00:31.560 So that's the point of the first line, trading is interesting, but inference is for 00:31.560 --> 00:36.720 her, and that's where you want to be effective in those accelerators. 00:36.720 --> 00:40.960 But of course, those accelerators go in big enterprise, data centers, and of course, 00:40.960 --> 00:45.280 you will waste lots of electricity on them, it can't stop. 00:45.280 --> 00:48.840 And of course, you need to have large engineering team. 00:48.840 --> 00:55.840 But, you know, it's totally possible to get into that business right now with risk five. 00:55.840 --> 01:00.600 Because when you look at this accelerator, it is a very simple design. 01:00.600 --> 01:03.760 You just take a basic processing element. 01:03.760 --> 01:09.640 You copy it a few times, a few dozen times, you know, and the thing is, if you have 01:09.640 --> 01:16.080 the proper IO's, then you have a chip perfectly, you know, capable to do that kind 01:16.080 --> 01:19.840 of inference at scale, you know, in a big enterprise data center. 01:19.840 --> 01:24.320 So maybe you start smiling, you grow, and, well, if you find the niche, other than boring 01:24.320 --> 01:26.800 ads, then maybe AI would be useful. 01:26.800 --> 01:32.160 So that's one way where, essentially, you could use risk five as this building block. 01:32.160 --> 01:37.520 Because of course, the AI specific sauce would be up to you, but anything which is 01:37.520 --> 01:43.320 generic computing that chip, you know, can be built using open source risk five designs. 01:43.320 --> 01:47.200 And that's the way to do it, because you don't want to reinvent the wheel with your own 01:47.200 --> 01:49.200 custom architecture for everything. 01:49.200 --> 01:57.640 Well, maybe this would be fun, but this would not be fast to market or profitable. 01:57.640 --> 02:00.320 Then, of course, the AI at the edge is a big topic. 02:00.320 --> 02:05.720 We have plenty of devices, all of you replace with built in AI, built in intelligence. 02:05.720 --> 02:11.000 And, well, a good example of that is this little AI chip by a sort of accelerer. 02:11.000 --> 02:16.960 So in that case, it's really a low power design, so you could run this in the field on 02:16.960 --> 02:22.240 battery powered, consumes just a few watts, okay, and probably if you would optimize it 02:22.240 --> 02:26.400 more, maybe it would go down up to a few million watts. 02:26.400 --> 02:33.520 And, well, that's great for a wide range of embedded, deeply embedded applications. 02:33.520 --> 02:39.040 In that case, the hardware is fairly small, so it could fit on a small M2 form factor 02:39.040 --> 02:43.320 board, okay, so a very small chip with the supporting components. 02:43.320 --> 02:48.480 And really, when you look at this, this is yet another very simple design. 02:48.480 --> 02:52.960 You have common building blocks with a risk of five controllers, and of course, security 02:52.960 --> 02:57.320 on the chip built in, because, well, if you are deploying this in the field, you need 02:57.320 --> 03:02.400 secure boot, you need accelerances that someone would rip off the device from the wall, 03:02.400 --> 03:07.040 let's say, wouldn't be able to access the data and be rich instead of you. 03:07.040 --> 03:13.080 Well, no, they are privacy implications and other like that, other things like that, of course. 03:13.080 --> 03:20.360 So really, the fact that the security is built in, in the case of this particular design 03:20.360 --> 03:25.880 is really fundamental, because you are deploying in the field, and so you need, you want 03:25.880 --> 03:35.000 to ensure that the data is private, and of course, that it stays really uncompromised. 03:35.000 --> 03:40.560 And, not your way to do this is to deal with, maybe, commercial vendors. 03:40.560 --> 03:47.120 So, for example, and desk technology, we'll known a risk-five vendor, they've got a number 03:47.120 --> 03:52.600 of risk-five cores, of course, but then they will happily license you their AI building 03:52.600 --> 03:54.560 blocks as well. 03:54.560 --> 04:00.280 So this is what industry to there, so you have a diagram for the type of core or chip 04:00.280 --> 04:02.040 that they would sell you. 04:02.040 --> 04:07.640 And of course, you know, they have an extensive software stack going well that, so they 04:07.640 --> 04:12.920 will provide you optimised compilers, real-time operating system, even an IDE if you want 04:12.920 --> 04:16.560 to use that, and what's wrong with them, anyway. 04:16.560 --> 04:23.160 All of that to say is you've got everything in the box to get you started more quickly, 04:23.160 --> 04:29.760 but in this case, their own business was only possible because risk-five was around providing 04:29.760 --> 04:35.360 this very stable foundation, and then they were able to build something that's really an 04:35.360 --> 04:41.040 interesting licensing place, so, you know, join the fund, that's yet another possibility 04:41.040 --> 04:43.520 for a startup if you are so inclined. 04:43.520 --> 04:49.200 And then, of course, there's plenty of AI in your pocket, right, and most phones today 04:49.200 --> 04:55.640 are able to do AI, but even for non-AI functions, have plenty of risk-five cores built 04:55.640 --> 05:00.960 in for power management, or screen management in the camera, all over the place, even 05:00.960 --> 05:07.520 fruity smartphones, so to speak, if you like, Apple's, you know, we'll have built-in 05:07.520 --> 05:12.920 risk-five controllers in the main SOC to do various things, right, and of course, they 05:12.920 --> 05:17.440 won't, they won't share that, those are not open source design. 05:17.440 --> 05:24.520 But all of that to say is that there are plenty of examples of, you know, using risk-five 05:24.520 --> 05:32.240 technology and build that into a more complex system to really deliver specialized things. 05:32.240 --> 05:38.160 So, if you build such a small AI accelerator, let's say that you want to embed into 05:38.160 --> 05:43.560 a non-based device, or even it's a device, you can do it, and once again, that's a business 05:43.560 --> 05:48.040 opportunity for you, and that's a business opportunity because, essentially, a risk-five 05:48.040 --> 05:53.440 is around and gives you a relatively free open building block to start with. 05:53.440 --> 05:59.680 But of course, well, that's a thing, risk-five is just the instruction set, and there 05:59.680 --> 06:04.760 are plenty of risk-five people that came on this stage earlier that will tell you about 06:04.760 --> 06:08.480 their marvelous boards and multiple scores and everything like that, but all of those 06:08.480 --> 06:14.200 things are proprietary, right, so you wouldn't necessarily have the source for that. 06:14.200 --> 06:21.880 So, what we do at Open Outwear Foundation is really to work on open source cores and even 06:21.880 --> 06:28.000 open source microcontrollers in one case, and provide those building blocks to the community, 06:28.000 --> 06:34.160 but at the same time, the focus is on industrial grades, so we do verification on those 06:34.160 --> 06:39.840 things, and I'm not the chip designer, Jeremy is, so we can tell you more about this aspect 06:39.840 --> 06:41.840 of things. 06:41.840 --> 06:49.320 Now, of course, you know, saying that we have risk-five cores could be anything, but really, 06:49.320 --> 06:56.320 in our case, the focus is really on smaller cores that go once again in a variety of embedded 06:56.320 --> 06:57.320 applications. 06:57.320 --> 07:04.480 Starting with CV2, which is roughly equivalent to the ARM Cortex M0, and then in the 07:04.480 --> 07:11.440 middle, the most popular family, let's say, so the CV4 family, where essentially we 07:11.440 --> 07:18.600 have equivalent for M4s or M2T3s on the arm side, and in this case, they are several variants 07:18.600 --> 07:24.560 with specific variations or specific feature sets, different from each other, but all of 07:24.560 --> 07:29.400 those are on GitHub in open source, and you can play with them and modify them, run them 07:29.400 --> 07:33.400 on FPGA, or even go get a chip made. 07:33.400 --> 07:42.120 And finally, CV6, CV6, in this case, runs Linux, so it's of course, a bigger and more powerful, 07:42.120 --> 07:51.840 and equivalent to ARM Cortex M0 or A55, in that case, so plenty of potential applications 07:51.840 --> 07:56.560 to all of those things, and of course, this is just what we have on the roadmap. 07:56.560 --> 08:02.960 We have, you know, longer-term plans for more powerful chips, eventually, in a can of 08:02.960 --> 08:09.200 stuff, so that's what you see, the CV8, for example, would be more powerful than all of 08:09.200 --> 08:10.200 those chips. 08:10.200 --> 08:17.680 So, once again, this, in the case of the open other Foundation, is a community driven 08:17.680 --> 08:23.040 set of projects, which means that, well, if you find this interesting and join, then 08:23.040 --> 08:30.160 we could go even smaller than the CV2, or even bigger than what CV8 is. 08:30.160 --> 08:34.640 This is all about, you know, bringing new members, bringing new ideas, and going where 08:34.640 --> 08:37.160 you want to go as a community. 08:37.160 --> 08:43.560 Once again, we are here, well, in my case, doing some part of the program management 08:43.560 --> 08:47.640 there to support our members in whatever they want to do. 08:47.640 --> 08:52.520 We don't necessarily dictate the technology direction. 08:52.520 --> 08:59.080 And so, what, will you find at Open Outwear Foundation, in terms of actual deliverables? 08:59.080 --> 09:01.400 Well, first, they are the course themselves. 09:01.400 --> 09:05.800 They are under the soldered pad license, which is roughly equivalent to the Apache tools, 09:05.800 --> 09:06.800 essentially. 09:06.800 --> 09:14.280 You take it, you do whatever you want with it, just put a small, a smaller, recognition 09:14.280 --> 09:19.280 in your readme somewhere that we've been around to provide you those building blocks. 09:19.280 --> 09:24.080 So the course are defined in system catalog, so you will find the false source code 09:24.080 --> 09:25.800 on GitHub. 09:25.800 --> 09:30.920 And then there's all the supporting infrastructure around it, right? 09:30.920 --> 09:37.560 So we have test benches, and of course, integration with most of the popular tools 09:37.560 --> 09:45.360 that are used by actual chip designers, and not, so please don't ask me about this. 09:45.360 --> 09:50.400 But once again, Dream is there, and it was more about those aspects than me. 09:50.400 --> 09:56.840 And of course, our friend Flow, so if you're wondering, it was on the talk, well, it's 09:56.920 --> 09:57.840 a really good thing for the first time. 09:57.840 --> 10:03.840 So, I'm here, instead. 10:03.840 --> 10:07.840 And so there's also a bunch of support software. 10:07.840 --> 10:16.840 So we work upstream with the TLVM and GCC and Jeremy will provide more details about that, 10:16.840 --> 10:22.440 but essentially we work upstream to ensure that our course are worth supported in those upstream 10:22.440 --> 10:23.840 tools. 10:23.840 --> 10:31.560 And then, of course, the open hardware foundation community worked on properly supported 10:31.560 --> 10:36.000 VRTOS support, for, of course, the smaller course. 10:36.000 --> 10:42.120 And we are starting to see an effort emerge around eclipse tradux as I will elaborate a bit 10:42.120 --> 10:43.120 later. 10:43.120 --> 10:48.920 And of course, there's plenty of documentation, and even board images for popular FPGA. 10:48.920 --> 10:53.640 So once again, if you want to modify the course, test them, then if you have one of those 10:53.640 --> 11:01.120 two FPGA's, well, the Nexus A7 will be appropriate for the CV-2 and 4, and the Genesis 11:01.120 --> 11:06.120 2, of course, for the CV-6, it's more expensive, but more powerful. 11:06.120 --> 11:11.240 You can actually boot Linux on that. 11:11.240 --> 11:15.440 Now of course, software will make and break your chip. 11:15.440 --> 11:34.760 And now it's Jeremy's time to shine, so I'll let us exchange a mic. 11:34.760 --> 11:35.760 Thank you, Frederick. 11:35.760 --> 11:39.600 Hopefully that's all good and clear. 11:39.600 --> 11:42.640 My background's more on the software side, so I'm going to talk about this. 11:42.640 --> 11:49.480 I'm co-chair of the technical work in group for open hardware foundation, which is 11:49.480 --> 11:56.600 the technical oversight for all the technical work we do. 11:56.600 --> 12:03.120 So within open hardware group, we break down into a small number of task groups to oversight 12:03.120 --> 12:04.880 to see the different areas we work on. 12:04.880 --> 12:07.480 We've got about five or six. 12:07.480 --> 12:13.040 The open hardware software task group is chaired by my colleague, Paolo Savini, and it's 12:13.040 --> 12:20.000 responsible for making sure that the software ecosystem we need is there around that chip. 12:20.000 --> 12:24.680 The tool chains, that's absolutely essential, but operating system ports where they needed 12:24.680 --> 12:30.520 any firmware that's needed for use the course and things like the SDK. 12:30.520 --> 12:36.160 And this continues to evolve and the work with ThreadX is just one of the latest things 12:36.160 --> 12:38.160 we're working on. 12:38.160 --> 12:43.440 It's a collaborative effort, it's open source, but it does rely on the member companies 12:43.440 --> 12:46.000 contributing to that effort. 12:46.000 --> 12:51.440 It's not just someone in their spare time, so that's been done both by companies like 12:51.440 --> 12:56.880 my and because I'm contributing, but also other companies actually funding us to contribute 12:56.880 --> 13:01.560 even more because they don't have a software team but they can buy us to do it as well. 13:01.560 --> 13:05.440 So that's how it's been done. 13:05.440 --> 13:10.960 In terms of the compilers, we've got GCC and Lofium, the goal is to go up stream. 13:10.960 --> 13:17.160 We don't want to maintain an out-of-tree compiler that is a whole world of pain of work 13:17.160 --> 13:19.160 closed. 13:19.160 --> 13:30.140 GCC 14.1 that came out last spring has five of the eight isher extensions for the CV-first 13:30.500 --> 13:33.240 costs of processor. 13:33.240 --> 13:39.540 There's three of them still in the review pipeline mostly because they're quite complex 13:39.540 --> 13:42.540 and the amount of review effort that's needed. 13:42.540 --> 13:48.460 KangelLVM18 has four of the extensions upstream that's running slightly behind. 13:48.460 --> 13:54.740 That work is largely done by our colleagues at the programming languages and compiler team 13:54.740 --> 13:57.300 of the Chinese Academy of Sciences. 13:57.300 --> 14:01.900 So that team has been a huge contribution there. 14:01.900 --> 14:03.300 They're all available to you, though, 14:03.300 --> 14:05.700 because we do maintain an out-of-tree mirror 14:05.700 --> 14:07.380 for the stuff that hasn't yet gone up soon. 14:07.380 --> 14:09.860 So if you go to the open hardware groups, 14:09.860 --> 14:13.580 GitHub, you will see that the source 14:13.580 --> 14:15.900 for the remaining out-of-tree materials, 14:15.900 --> 14:18.100 they're all the extensions are supported. 14:18.100 --> 14:21.420 And if you like your compile-a-tree change pre-made, 14:21.420 --> 14:24.420 you can download them from the Emboxon website. 14:24.420 --> 14:29.700 So how far have we got with the tool chains? 14:29.700 --> 14:32.820 Full assembler support, full built-in support, 14:32.820 --> 14:34.300 where it makes sense for some functions, 14:34.300 --> 14:35.820 it doesn't make sense to have built-in support. 14:35.820 --> 14:37.340 So that's due to efficiently. 14:37.340 --> 14:41.820 And for all but two of the extensions, 14:41.820 --> 14:44.740 we also have automatic code generation. 14:44.740 --> 14:46.460 We haven't yet got the code generation 14:46.460 --> 14:47.460 for the Bitmin population, 14:47.460 --> 14:49.820 because it is not standard Bitmin population. 14:49.820 --> 14:52.100 And we haven't yet got code generation 14:52.100 --> 14:56.100 for the SIMD instructions, but the rest of it's there. 14:58.500 --> 15:01.420 Operating systems are the QEMU port. 15:01.420 --> 15:02.660 That's a work in progress. 15:02.660 --> 15:05.340 Do come and join in if you'd like to help with that. 15:05.340 --> 15:09.140 There's free R-toss was done some time ago. 15:09.140 --> 15:11.580 There's work on going with Zepha. 15:11.580 --> 15:14.060 ThreadX, Frederick is much better qualified 15:14.060 --> 15:17.300 to talk about and he'll talk more about that later. 15:17.300 --> 15:19.020 There's no special Linux work, 15:19.020 --> 15:21.180 because actually we don't need anything special on Linux. 15:21.180 --> 15:24.140 It uses a standard Linux kernel, 15:24.140 --> 15:26.340 but for Dora and Red Hat have been working 15:26.340 --> 15:28.500 on then getting that all pulled together 15:28.500 --> 15:29.500 at a system level. 15:29.500 --> 15:34.500 So it all works on particularly CVACX. 15:37.220 --> 15:38.700 And last, you have an ID. 15:38.700 --> 15:43.020 Now this ID may not be, it is Ashley's risk free ID. 15:43.020 --> 15:46.300 So it's a well-established industry ID based on, 15:46.300 --> 15:49.940 in this case, eclipse, it'll actually run on other infrastructures, 15:49.940 --> 15:51.140 as well. 15:51.460 --> 15:56.140 But it's made available free for the open, 15:57.620 --> 16:00.060 for the open hardware foundation, 16:00.060 --> 16:04.100 and has the compilers, the debug, or everything built in. 16:04.100 --> 16:05.820 It does everything you'd expect, 16:05.820 --> 16:07.220 and you can just program in there. 16:07.220 --> 16:10.980 You can plug into your FPGA board and it just works. 16:10.980 --> 16:12.220 And well documented. 16:15.100 --> 16:16.140 We have an MCU. 16:16.140 --> 16:20.300 So in order to, we thought we'd make ourselves a reference MCU, 16:20.300 --> 16:21.940 so not just the processor, 16:21.940 --> 16:23.380 but some useful peripherals, 16:23.380 --> 16:25.980 so you've got some networking connections. 16:25.980 --> 16:28.940 You've got a timer, you've got buses to hook up to memory, 16:28.940 --> 16:30.260 and so forth. 16:30.260 --> 16:35.260 And that is the basis of the Core 5 Devkit, 16:36.460 --> 16:37.940 which is a Devboard. 16:37.940 --> 16:41.820 Now, I do have to say there's a slight catch here. 16:41.820 --> 16:43.940 The first version of the chip we spun 16:43.940 --> 16:46.020 had a metal layer error. 16:46.020 --> 16:49.300 We do have a small number of working chips, 16:49.300 --> 16:52.260 courtesy is a very clever laser use by our friends 16:52.260 --> 16:54.260 at the University of Saskatchewan, 16:54.260 --> 16:56.740 but there is a plan that this will be responding to you, 16:56.740 --> 16:58.980 course, and then we'll have a working silicon. 17:00.340 --> 17:03.740 So, and the idea of that is that we've got real 17:03.740 --> 17:05.740 a sick type hardware that goes faster 17:05.740 --> 17:07.100 and so forth, we can test things on. 17:09.340 --> 17:11.780 Now, I want to just end, because this is a talk 17:11.780 --> 17:16.100 about RIS 5 in AI, about an open source AI project, 17:16.100 --> 17:18.580 that is RIS 5, it's Core 5 based, 17:18.580 --> 17:19.460 and it is for the open. 17:19.460 --> 17:23.900 You can go and download this from GitHub. 17:25.540 --> 17:28.820 As a side interest, William Jones, who's my head of AI, 17:28.820 --> 17:31.820 and I, Supervisor Group Design Project, 17:31.820 --> 17:35.100 Southampton University, every autumn. 17:35.100 --> 17:38.900 So, we have six, four to six students working on a project. 17:38.900 --> 17:42.020 And we asked them a few years ago the question 17:42.020 --> 17:45.820 of everyone's building these AI accelerators based on RIS 5, 17:45.820 --> 17:47.820 which have the whole vector instruction set in, 17:47.820 --> 17:50.220 and they take vast acres of silicon. 17:50.220 --> 17:52.860 And that's not good for low-power applications, 17:52.860 --> 17:56.140 and really do you need the whole RIS 5 vector extension 17:56.140 --> 17:57.140 to accelerate AI. 17:57.140 --> 17:59.700 So, we said, how small can you make it? 17:59.700 --> 18:02.980 And they developed a derivative of CV-3040B, 18:02.980 --> 18:06.300 extended with just eight RIS 5 vector instructions, 18:06.300 --> 18:11.420 load and store, and one or two ads track multiplying and so forth. 18:11.420 --> 18:15.940 So, a tiny extension, and the result of that was, 18:16.020 --> 18:18.220 they got between a 5 and 7 fold speed up 18:18.220 --> 18:20.580 on the tiny ML-Perv benchmark. 18:20.580 --> 18:24.180 There is a video YouTube, they gave a talk 18:24.180 --> 18:26.460 to the British computer society about this work. 18:26.460 --> 18:28.340 It is the second most viewed video 18:28.340 --> 18:30.740 from the British computer society open source group. 18:30.740 --> 18:34.180 They're on YouTube, add all the sources on GitHub. 18:34.180 --> 18:37.260 And so, I put that there, not because it's particularly 18:37.260 --> 18:39.060 a software project, because I supervise it, 18:39.060 --> 18:40.660 and I was doing the software talk, so I think you came 18:40.660 --> 18:41.660 in the talk. 18:41.660 --> 18:46.660 So, with that, I'm going to hand back to Frederick. 18:52.220 --> 18:54.460 Thank you, all right. 18:54.460 --> 18:59.180 So, why would you come play with us, right? 18:59.180 --> 19:02.140 Because you want to get rich, you have ideas 19:02.140 --> 19:07.140 for our NEI startup, so why work in open source with us? 19:07.140 --> 19:14.140 Well, essentially, first, risk-5 is really changing 19:14.140 --> 19:16.700 the game in terms of what you can do with AI, 19:16.700 --> 19:18.980 because by using it as a building block, 19:18.980 --> 19:23.500 you can focus on your accelerator and on your added value 19:23.500 --> 19:26.980 on top, which means you get faster to market. 19:26.980 --> 19:29.180 And from that perspective, once again, 19:29.180 --> 19:31.940 it's one thing to have the I.C, but in our case, 19:31.940 --> 19:35.020 we provide you in this real great, verified course 19:35.020 --> 19:38.300 that you can leverage in your designs. 19:38.300 --> 19:42.380 And, of course, since the license is very permissive, 19:42.380 --> 19:45.580 you can use them as is, but you can improve on them, as well, 19:45.580 --> 19:46.260 right? 19:46.260 --> 19:50.020 And we hope that we get contributions back from that, 19:50.020 --> 19:54.100 but nothing in the license forces you to do so, okay? 19:54.100 --> 19:57.740 And really, the way that we operate is around three core values, 19:57.740 --> 20:00.580 open, transparent, and meritocratic. 20:00.580 --> 20:03.500 Open, being with everything in public, okay? 20:03.500 --> 20:07.380 There are no hidden mailing lists or hidden room, 20:07.380 --> 20:11.740 where you know the big guy stock, everything is done in the open, okay? 20:11.740 --> 20:14.580 Transparent, of course, means that we have public road maps, 20:14.580 --> 20:16.180 and I love that, you've seen the road map, 20:16.180 --> 20:18.660 and of course, there's more detail information 20:18.660 --> 20:20.940 on the open, other web foundation website. 20:20.940 --> 20:23.500 You can join the matter most in order to interact 20:23.500 --> 20:24.340 with the community. 20:24.340 --> 20:27.100 There are plenty of ways that you can find us. 20:27.100 --> 20:29.580 And even if you're not a member yet, 20:29.580 --> 20:31.380 you can interact with us. 20:31.380 --> 20:33.500 And meritocratic means that essentially, 20:33.500 --> 20:37.140 the people with commit rights on the repositories 20:37.140 --> 20:39.060 have to earn their way, right? 20:39.060 --> 20:40.940 They need to prove that they have the skills 20:40.940 --> 20:42.300 and that they've made contributions 20:42.300 --> 20:46.620 before becoming core committees on those projects. 20:46.620 --> 20:49.260 And a third big advantage of open-out web foundation 20:49.260 --> 20:52.220 is that it really brings together industrial players 20:52.220 --> 20:53.540 and academics. 20:53.540 --> 20:57.580 And to give you an example, I was traveling two weeks ago 20:57.580 --> 20:59.780 with Flow, we went to Montreal to put it 20:59.780 --> 21:01.180 at Technique Montreal. 21:01.180 --> 21:03.460 And we were visiting a professor there. 21:03.460 --> 21:06.900 He's got a project, a research project called Polara. 21:06.900 --> 21:11.700 And Polara was about making a physical chip out of the CVA6. 21:11.700 --> 21:16.700 So I've seen that very particular chip, okay? 21:16.700 --> 21:18.420 There's only an unfolding the word, 21:18.420 --> 21:22.340 and that was quite, quite impressive as a moment, 21:22.340 --> 21:25.260 because that was literally an open-source design 21:25.260 --> 21:28.860 on a chip that I could plug into, 21:28.860 --> 21:30.500 something and run the next on it. 21:30.500 --> 21:34.700 So of course, they are challenges, okay? 21:34.700 --> 21:38.980 And so to me, the technology needs to mature a little bit more, 21:38.980 --> 21:41.340 but this is an impressive achievement, 21:41.340 --> 21:43.980 and this is something that you could replicate yourself 21:43.980 --> 21:46.780 if you're thinking about starting a business. 21:46.780 --> 21:49.380 Of course, the cores are fully open. 21:49.380 --> 21:51.620 There's nothing hidden in there. 21:51.620 --> 21:54.540 And they are fully verified. 21:54.540 --> 21:58.140 So challenge our IRTL, if you want. 21:58.140 --> 22:01.460 But I mean, actual chip designers worked on this, 22:01.460 --> 22:03.620 and we have, of course, the verification, 22:03.620 --> 22:06.980 the validation reports to go with the technology. 22:06.980 --> 22:10.500 And finally, well, if you love Europe, I do. 22:10.500 --> 22:13.780 Then, well, open hardware foundation 22:13.780 --> 22:15.460 is a part of the Eclipse Foundation. 22:15.460 --> 22:17.940 The Eclipse Foundation is the largest open source 22:17.940 --> 22:19.340 organization in Europe. 22:19.340 --> 22:24.740 We have over 50% of our staff now in countries like France, 22:24.740 --> 22:27.820 and Germany, Sweden, Romania, Spain, Italy. 22:27.820 --> 22:29.860 So we're all available to place in Europe, 22:29.860 --> 22:31.700 and so we are close to you. 22:31.700 --> 22:33.620 We speak your language. 22:33.620 --> 22:36.900 It's a piston toy. 22:36.900 --> 22:38.860 I've known Poco is a Spaniard. 22:38.860 --> 22:41.420 Anyway, it's a part of our service. 22:41.420 --> 22:43.620 So I'll have that to say. 22:43.620 --> 22:45.660 The Eclipse Foundation, if you care about Europe, 22:45.660 --> 22:48.220 is the place to be. 22:48.220 --> 22:51.220 So my call to action, learned about core five, 22:51.220 --> 22:53.740 which is the name of our family, of course. 22:53.740 --> 22:55.900 And of course, the Open Outwear Foundation, 22:55.900 --> 22:56.820 visit the website. 22:56.820 --> 23:00.780 We've got a YouTube channel with lots of interesting videos 23:00.780 --> 23:03.740 that will get you acquainted with the technology, as well. 23:03.740 --> 23:04.860 Try the code on GitHub. 23:04.860 --> 23:09.460 The system very long is there, so if you're really a chip designer, 23:09.460 --> 23:12.500 please have a look and let us know what you think. 23:12.500 --> 23:13.500 And of course, engage. 23:13.500 --> 23:16.660 So that particular link, the slides are already in the tools. 23:16.660 --> 23:19.140 So when you click on it, you'll come to a section 23:19.140 --> 23:22.380 of our website where you have all the ways to interact with us. 23:22.380 --> 23:25.100 We'll find our references to the mailing lists, 23:25.100 --> 23:27.260 the matter most instances and everything 23:27.260 --> 23:30.260 that can get you connected into this. 23:30.260 --> 23:33.220 And now the last slide, what we're trying to achieve there 23:33.220 --> 23:37.260 by bringing Open Outwear Group into the Eclipse Foundation 23:37.260 --> 23:40.340 as the Outwear Foundation is to build this comprehensive 23:40.340 --> 23:43.420 and better risk-fifesteck, okay? 23:43.420 --> 23:46.740 From the course, to the tool chains like the compilers 23:46.740 --> 23:47.860 and we've seen with Jeremy, 23:47.860 --> 23:50.820 we're taking this upstream work really seriously. 23:50.820 --> 23:52.980 The real time operating systems 23:52.980 --> 23:56.020 and in particular, Eclipse 3DX, Eclipse 3DX 23:56.020 --> 24:00.540 is the only safety certified real time operating system, 24:00.540 --> 24:02.740 which is also open source in the market. 24:02.740 --> 24:04.340 So you can put it in cars. 24:04.340 --> 24:06.660 You can put it in home appliances. 24:06.660 --> 24:09.340 You can put it in medical devices today, 24:09.340 --> 24:12.900 not two, three, four years away. 24:12.900 --> 24:15.220 Like it would be the case with the affair, 24:15.220 --> 24:18.940 but literally today we've got the certificates to prove it, okay? 24:18.940 --> 24:21.500 So in this case, we are starting this effort 24:21.500 --> 24:26.500 to port it to Core 5 and specifically the CV4 family 24:27.860 --> 24:30.140 and eventually CV6 maybe. 24:30.140 --> 24:33.100 Anyway, and also of course, 24:33.100 --> 24:34.700 the Eclipse Foundation traditionally 24:34.700 --> 24:36.380 was all about development tools. 24:36.380 --> 24:39.020 So the traditional Eclipse IDE is still around, 24:39.020 --> 24:42.140 but we've got also Eclipse TA, Eclipse TA 24:42.140 --> 24:46.460 is like visuals to do called minus the Microsoft 24:46.460 --> 24:49.180 and it's integrated with Open VSX, 24:49.180 --> 24:52.860 which is a vendor neutral extension repository, 24:52.860 --> 24:56.260 not controlled by anyone but owned by the Eclipse Foundation 24:56.260 --> 24:58.340 on behalf of the Open source community. 24:58.340 --> 25:01.820 So if you truly believe in a hardcore open source 25:01.820 --> 25:03.860 approach and so to speak, 25:03.860 --> 25:06.140 you've got all of the pieces there in the open. 25:06.140 --> 25:10.220 And so this is not about just the traditional Eclipse IDE, 25:10.220 --> 25:14.540 but really forward-looking technologies. 25:14.540 --> 25:16.980 And of course, there's the rest of the Eclipse family. 25:16.980 --> 25:20.420 So I'm also responsible for the Eclipse IoT working group. 25:20.420 --> 25:23.140 So if you want an MQTT broker or something like that, 25:23.140 --> 25:25.940 then you can go there on the problem. 25:25.940 --> 25:28.460 And of course, we've got a full-blown software 25:28.460 --> 25:29.860 to find vehicle working group. 25:29.860 --> 25:33.180 So if you're thinking about doing AI 25:33.180 --> 25:36.420 but around the automotive, then of course, 25:36.420 --> 25:39.140 our SDB working group is a place to be. 25:39.140 --> 25:41.140 So thank you very much. 25:41.140 --> 25:43.100 It's been a pleasure to be with you today. 25:43.100 --> 25:45.540 And now we've got some time for questions. 25:45.540 --> 25:46.700 So. 25:46.700 --> 26:05.020 So it seems the question microphone is dead. 26:05.020 --> 26:06.780 OK, we'll repeat the questions then. 26:06.780 --> 26:09.580 So please ask, I will repeat for the people online. 26:09.580 --> 26:30.380 So the question here is the CV-2 and 4 families. 26:30.380 --> 26:34.020 They are for embedded in CV-6 for more powerful things. 26:34.020 --> 26:35.180 And yes, that's OK. 26:35.180 --> 26:37.220 So CV-6 runs Linux. 26:40.540 --> 26:58.540 So the question is why is the CV-4 or even CV-2 supporting 26:58.540 --> 27:01.500 on the 32 bits? 27:01.500 --> 27:02.420 That's a good question. 27:06.020 --> 27:06.780 It's enough. 27:06.780 --> 27:08.860 I think for the applications. 27:08.860 --> 27:13.100 If a member came along and said I wanted to build a 64-bit embedded 27:13.100 --> 27:15.820 class processor, sure, that would be fine. 27:15.820 --> 27:19.260 The intention is the CV-6, which comes in 30 to 60 27:19.260 --> 27:24.300 before the versions, is intended for application classes. 27:24.300 --> 27:27.900 Of course, a typical microcontroller would be a few hundred 27:27.900 --> 27:33.740 megahertz in clock frequency and maybe one megabyte of memory 27:33.740 --> 27:37.020 or maybe five hundred, five hundred and twelve keys. 27:37.020 --> 27:41.900 So in that case, even if the processor 27:41.900 --> 27:45.180 core itself would be 64 bits, then it would be overkill. 27:45.180 --> 27:49.260 And maybe it's a bit more power efficient to stay on 32 bits 27:49.260 --> 27:52.700 for that kind of embedded application as well. 27:52.700 --> 27:53.900 Yes? 27:53.900 --> 27:55.180 Yes. 27:55.180 --> 27:58.460 Is there a way to turn a better one in the next open neural network 27:58.460 --> 28:04.140 or change format to the core specific instructions? 28:04.140 --> 28:08.060 So the question here is, is it possible to convert 28:08.060 --> 28:11.660 from the open network in X? 28:11.660 --> 28:12.460 Oh, sorry. 28:12.460 --> 28:13.500 Oh, yeah. 28:13.500 --> 28:16.620 So the question is about ONNX, which is the open neural network 28:16.620 --> 28:17.900 exchange format. 28:17.900 --> 28:22.700 The, that has not been investment yet in the tool chain 28:22.700 --> 28:23.660 to develop for that. 28:23.660 --> 28:25.260 I recommend to the talk. 28:25.260 --> 28:28.140 I'm going to be with William Jones tomorrow afternoon in the AI 28:28.140 --> 28:30.300 on how to do that, which is an example. 28:30.300 --> 28:32.780 Again, I think example project from Southampton University 28:32.780 --> 28:34.940 is it happens on how we'll do that. 28:34.940 --> 28:37.900 But yeah, that's generally a problem for anything 28:37.900 --> 28:38.940 that's not an Nvidia chip. 28:57.260 --> 29:00.420 So your question or comment is about the fact 29:00.420 --> 29:06.740 that in terms of PCI express support, they are not, 29:06.740 --> 29:11.460 let's say they are no good open source alternatives 29:11.460 --> 29:13.060 currently that you could leverage. 29:13.060 --> 29:14.060 Is that it? 29:14.060 --> 29:17.060 The problem is not from this task, because me and my team 29:17.060 --> 29:22.660 we kind of take on $1,000 of the complex IP 29:22.660 --> 29:26.660 for PCI experts, so all can be too today. 29:26.660 --> 29:31.300 OK, so it's very expensive to get access to the IP 29:31.300 --> 29:34.260 for, you know, let's say a PCI express controller. 29:34.260 --> 29:37.940 And so you're asking what you could do about that 29:37.940 --> 29:39.780 in terms of implementing something. 29:39.780 --> 29:43.620 Not sure if some members worked on such things. 29:43.620 --> 29:45.860 I was saying more on the software side. 29:45.860 --> 29:47.940 I mean, it's a general problem with peripheral hardware. 29:47.940 --> 29:51.220 It's been forever going back to open calls days 29:51.220 --> 29:56.100 that actually well-beating high performance peripherals 29:56.100 --> 29:58.580 are really hard to design, and that hasn't 29:58.580 --> 30:01.540 been a big enough community to do open source versions. 30:01.540 --> 30:03.700 I'm sure if anyone would like to contribute a state 30:03.700 --> 30:08.020 of the art high performance PCI, any sort of PCI express 30:08.020 --> 30:09.940 interface we'd love to have it, and we'd 30:09.940 --> 30:13.780 I'm sure welcome it as a part of the open hardware foundation. 30:13.780 --> 30:15.460 But I don't know of an implementation. 30:15.460 --> 30:16.740 Does anyone in the room know if an open 30:16.740 --> 30:19.460 high performance implementation is open source? 30:19.860 --> 30:28.580 Yeah, so the session like such, it has something that might help 30:28.580 --> 30:29.860 towards that, that's good. 30:29.860 --> 30:37.460 But it is PCI 2X, not PCI 3X, so PCI 2X, not PCI 3X, okay? 30:37.460 --> 30:39.460 Okay? 30:39.460 --> 30:42.260 Yeah, I have a question. 30:42.260 --> 30:44.500 Oh, okay. 30:44.500 --> 30:48.820 So are there any plans of aping out CES6? 30:48.820 --> 30:55.620 You already taped out, I understand the 32 bit MCU, somehow the 30:55.620 --> 31:00.580 like you explained the PCI 3X, the boat is not available, 31:00.580 --> 31:02.980 maybe only available to members internally. 31:02.980 --> 31:07.140 But taping out CES6 hasn't been pride, or is it still 31:07.140 --> 31:10.580 let's say on the ideal side, and let's say never on an ASIC, 31:10.580 --> 31:13.060 or GDS, it hasn't seen a found rate. 31:13.060 --> 31:16.020 What is the plan going there? 31:16.020 --> 31:20.580 So just in case the microphone wasn't loud enough for people online. 31:20.580 --> 31:24.820 So the question is essentially if we are planning to tap out 31:24.820 --> 31:31.540 have a physical chip of the CVA6, because in the case of the CVA4, 31:31.540 --> 31:36.740 something something, in the context of the core 5 MCU we had, 31:36.740 --> 31:39.540 those physical chips, I have a non-working board, 31:39.540 --> 31:43.380 unfortunately with me in my back, I wish it would work. 31:43.380 --> 31:52.580 But in the case of the core 5 MCU and the dev kit board that we've shown, 31:52.580 --> 31:57.940 in fact, I think the PCB design for this is open as well, right? 31:57.940 --> 32:03.220 Yeah, so everything, everything that we could make publicly available is. 32:03.220 --> 32:07.300 But do we have the same type of plan for the CVA6 to my knowledge, no? 32:07.300 --> 32:11.940 But I think the CVA4, I mean, that small, 32:11.940 --> 32:15.460 it was easy, and it was done with the support of quick logic and silicon labs 32:15.460 --> 32:19.540 to get it out the door, it's expensive to do that. 32:19.540 --> 32:24.900 The CVA6, however, is being heavily used in a number of European research projects, 32:24.900 --> 32:29.780 and I anticipate that those projects, so we've got any of the CVA6 people here, 32:29.780 --> 32:32.420 there we haven't got them in the room, but I believe as some of those, 32:32.420 --> 32:34.820 they will be doing some degree of taping out. 32:34.820 --> 32:38.900 But I think you have to go and look at those specific projects last. 32:38.900 --> 32:43.060 Of course, I mean, we could be creative eventually about that, 32:43.060 --> 32:46.820 you know, and I'm just rambling off without the actual bus, you know, 32:46.820 --> 32:52.020 flow being, you know, in Taiwan and all, you know, 32:52.020 --> 32:56.980 if there's community demand from a number of, you know, companies that we're willing to 32:56.980 --> 33:01.620 pull resources in order to make that happen, we would, of course, consider it, right? 33:01.860 --> 33:05.060 But as Jeremy said, this is an expensive endeavor, 33:05.060 --> 33:09.220 so if you want physical chips, you better join, and tell your friends to join, 33:09.220 --> 33:12.740 etc. and then we get the money to make that happen, 33:12.740 --> 33:16.740 because that's one of the things that I noticed in today's many presentations. 33:16.740 --> 33:18.340 I'm coming to your question. 33:19.620 --> 33:23.780 There's a lack of good, you know, validated boards or easily 33:23.780 --> 33:28.900 of the obtainable boards in the market right now for people willing to work with this technology. 33:28.900 --> 33:32.180 So maybe we can fix that as an open community. 33:32.180 --> 33:32.980 So sir. 33:38.980 --> 33:41.220 Well, we are members of the RISFF Foundation. 33:41.220 --> 33:43.700 We are on a number of their committees. 33:44.660 --> 33:46.180 Maybe Jeremy knows which one. 33:48.180 --> 33:50.580 Individual members, most members of Open Hub, 33:50.580 --> 33:53.300 with my company, for example, I'm members of RISFF International, 33:53.300 --> 33:56.260 and will be on committees there. 33:56.340 --> 33:58.420 And there is engagement with chief executive, 33:58.420 --> 34:00.260 level to chief executive, of the CTA, 34:00.260 --> 34:01.380 trying to coordinate what we do. 34:01.380 --> 34:03.860 And indeed with the RISF Foundation and other organizations, 34:03.860 --> 34:05.460 it's not just RISF. 34:07.460 --> 34:11.220 So we're one big appeal, open family, so to speak. 34:12.500 --> 34:12.900 Yes. 34:22.900 --> 34:25.060 That, I have no idea. 34:26.020 --> 34:28.660 So the question is which PDK flow and I don't have an idea. 34:28.660 --> 34:31.220 At this point, it comes clear, I'm not really a hard-wrenching there. 34:31.220 --> 34:34.900 So the answer is I don't know, but it's a question we can take away and get the answer back. 34:44.580 --> 34:46.900 There is people who have looked at that and the open hardware, 34:46.900 --> 34:49.940 the print open source silicon foundation has regularly had key 34:49.940 --> 34:52.580 stock projects to try and actually get this fully open. 34:52.660 --> 34:56.580 The actual defined flows, though, are not open source. 34:56.580 --> 35:00.580 They use synopsis UVM flow, I think. 35:00.580 --> 35:04.420 So I'd love it, I'd love it if someone wants to pick that up, 35:04.420 --> 35:05.620 and you can do it.