 And we're here at the Lunaro Connect and this is the world's first ARM developer box. Now, it's third. It was third. The third. Or maybe the fourth. I heard something. I heard the fourth as well, yes. Somebody said that it was a Seattle base, right? No, no, not the Seattle, but there was a PC after the RISC PC, yes. So, how are you? I'm Daniel Thompson. I work at Lunaro in the support and solutions engineering team. And it's been my privilege to be playing around with the developer box. So I've had one of these on my desk. I've had it for six months. I've been using it as my day-to-day computer for the last four months. This has been your main desktop machine? Yes. Unless I'm in a coffee shop, I've done every bit of work I do for Lunaro on this board for the last month. When you're home? Yeah, okay. Obviously, it doesn't work well in a coffee shop. Is it not yet the laptop? No. But it could be, you're waiting for a laptop? It would be quite a pokey laptop. Let's get a laptop soon. You want a laptop that's ARM powered? I would love a laptop that's ARM powered. I mean, I like this one a lot because it can take a lot of RAM. So I run mine with 8GB, but you can put up to 64GB of RAM into this box. And for me, I don't want a device with less than 8GB regularly. Why? Why do you need with 8GB? I find that you see swapping if you run 4GB and you have Chromium open, and I also run other applications. Chromium at home will survive, but you open the next application and you tend to see swapping. Swapping, that means it goes into the main memory inside the RAM? It's the process where when you start to run out of memory, it puts memory on the disk instead. And even with modern fast SSDs, the delays of getting stuff on the SSD and back again are enough to make it feel lumpy. And I don't want to work on something that feels lumpy. So you say you're in the what support group? The support and solutions engineering. So what does that do? We look at a whole bunch of things. The short version is we help members exploit their membership with Lunaro. So the core engineering teams will enable pieces of hardware occasionally, but they want to enable technology. And our members want to sell chips. What's that? The club and core members want to sell chips. And so you need to help bridge the technology and help the members exploit the technology that Lunaro has created. So how awesome to have an ARM developer box compared to have an Intel thing. Yes. How does that change? Surprisingly little. I've run it for four months and you don't get up in the morning and think, oh, I wish I had this this program or this program. Everything has, what I say everything, but it's a few tiny, you know, two tiny problems. Almost everything has just worked. And it's always been, you know, quite obscure developer centric tools that haven't quite been made to work yet. What's the most important thing you do? Usually you open a terminal and you start typing stuff or what do you do? I have some managerial responsibilities. So the most important program I open is a mail reader. A mail reader. And that's actually true for job-earning engineers as well. I mean engineering is now a social discipline. So open source is a social process. So for every developer, I think mail is probably their most important tool. But email is maybe not very multi-threaded, right? You need a single thread performance to run an email program? It depends which programs you run, of course. Yes, so the one I liked or the one I was using before I went to the Swiss Thunderbird and that didn't feel very snappy. That's a single-threaded app. I don't know how threaded it is, but it certainly wasn't very snappy. And so that's the only change I made to my set of applications. All my other applications that I've used, you know, Chromium, things like that, they've all come across fine. But Thunderbird just didn't work for me. I switched to MUT, which I used years ago. Is that a good one? I don't know if I've been installed on this device, but it's a terminal one. Terminal-based email? Yes, security people are very fond of it, because it's harder to attack. It's got a smaller attack surface than a GUI browser. So we're in a terminal right now? Yeah, I should imagine it's not installed. Cool stuff. Well, it is merely a terminal. It will do the same thing as every other terminal on every other Linux machine. That's the point. We want it to be like using any other device. We don't want surprises when you look at the terminal. So what would you do, for example? The nice things you can see is H-Top. So H-Top has this nice representation of the number of processors running. So when you can see here sitting up, they're not really doing a great job. I hope opening Chromium is usually a good way to get six or eight of the cores running. Have we gotten Chromium? Internet? Oh, we haven't, no. This hasn't got Chromium installed yet. So you need to do an app get or something? Yes. If that works, I'd be rather... The Internet is pretty fast here. It's nice in Hong Kong, right? How's it in your home? It's not making it through, so I think it's not good it plugged into the net. Okay, but how's your Internet at home? How's my Internet at home? It's modern broadband. I live in a city, so I get good broadband, yeah. So you can do anything you want to get stuff downloaded? So we had some cool tricks on it. So we got Minecraft running on one of these. Because it's written in Java. And because it's written in Java, we had to change a couple of shared libraries, but it just ran Minecraft and it worked. It was fine. My summer's very happy. Is the stuff about using an ARM developer box the most important that you can compile on native architecture? Yes, so the very best thing for me is when I need to spin up a virtual machine to test something on. I just type QMU and then I say, use the kernel that's in slash boot slash vmlinus. So I didn't even have to remember where I left the kernel because the kernel that booted the box is also the same kernel that runs the virtual machine. And that just makes it, it just reduces the amount of remembering things you have to do to start a QMU session. But is it also better for like testing ARM on ARM? Yeah, I mean that's very much yeah. So by spinning up that virtual machine that is all about testing ARM on ARM. It's allowing me to make kernel changes and spin up a kernel and see what happens if it works. I also did a bunch of stuff with containers similarly. Being able to create native ARM containers to test if you've done it right, if you've generated your root of S correctly, all these types of things I can now do without having to spin up a developer board, a small, singable computer. I can do it all on the one platform so much more easily. If you were doing that on an Intel, would there be some issues that you wouldn't notice because it's not native? Well, the first thing is, even with the relatively slow single-formance of this device, it's still far better at being an ARM than a PCS, right? So if you get the PC to pretend to be an ARM, it will never be as fast and you will never expose the memory ordering. So PCs are perhaps over-aggressive in memory ordering. They've got really strong constraints on how the multiple processors can interleave when they work in memory. And ARM has what's called a much more relaxed model. And that relaxed model has benefits and it also has some testing needs. So for all the interior engineers, all people working on ARM software, it would be better to use an ARM device to develop ARM. I'm not going to preach to other people they should do, but yeah, for me it's a great way to do ARM development. And the only thing that might be missing is just single thread, right? So this device is a kind of performance per watt design. So it's designed to offer a quite incredible performance per watt. And when this machine is red-lined, it only uses about five watts more than it does when it's idle because the CPU only ever draws five watts. So yeah, for sure. So it's a performance per watt machine. It's very wide and when you can exploit that width, everything's fine. But when you can't get all of those CPUs lit up, then you do sometimes see slowdowns in particular single-threaded activity. And here they're doing a demo. Let's just check this one out first. This is the box that people get when they order it. Yes. So there's like the desktop here. And what is in here? Actually, it comes as a kit. So the kit contains the motherboard and graphics card and memory and a disk. And then you assemble that yourself into a case. And there's a really nice assembly guide for people to follow. Anyone can order. Anyone can order. Anyone needs to own a screwdriver. And the price is good. The price I'm not quite sure of at the moment. The price is like, is it $1,000 or? $1,250 for the whole like this. For the kit, yeah. Shipped all of the world. All of the world. So it has CE, FCC and all that, I guess. So it just comes smoothly through the... And this is an AI demo right here. Can we check it out? This is an example of what happens when you have multiple PCI slots. So this device has a lot of PCI capability. So it has a graphics card in it. It also has... You've talked to one person some extent, but it has a really powerful AI accelerator that made a big splash at CES. What AI accelerator is that? I don't know. It's written on the wall. It's a neural processor. So it's there to help do all the matrix multiplications. So that's on the SoC, the AI accelerator? No, no, it's attached to the PCIe. Ah, it's a PCIe. Read. So it's simply an example of what happens when you have PCIe slots that have lots of them. So actually if you need some of those singles, could you add that on PCIe somehow? Could you add another big armchair? In principle, yes, but somebody would have to make such a board. And you might have some challenges with the memory benefit, not memory benefit, making sure they can access memory efficiently. So this is going to launch up a pattern recognizer, which is a train neural network to try and classify images. So it's now playing videos, and it's able to recognize each of these images and classify it for what's contained in the image at an impressive speed and at an even more impressive power cost. So that was reading 32. 32 is kind of the overhead of power supply and refreshing the RAM and all the other things you have on your board. And on a PC, a conventional PC, which you do that if you start to use those cores that the PC has, you'll see the power level triple quite often, whereas this has gone up by three watts. Three watts to do all this stuff? So it's like a 22.4... Terra-flops, I believe. No, not flops. Terra-operations. They're not flops. Terra-operations per second. Powerful CNN. Is that to do with TV channel? No, something else. It's a type of neural network. So I come with what the C stands for, but the N and the N are neural network. 200 frames per second system. Just a three-watt increase in the AI. So that's just another A accelerator. That's actually a Socialnext chipset. It's not. Socialnext partnered with what they call Greifonkin. All right. Partnered with the... Yeah. Somewhere. Yeah, so it's on an M2. The actual accelerator is on an M2 format, which allows, in theory, you could drop it into other devices. But that means it's ready to go into small-scale embedded systems because M2 is nice and compact. But in this case, it's been attached to an adapter card and placed in the PCI slots. And that's the PCIe you can... Usually, in yours, you put an NVIDIA? Yes, I run NVIDIA GT10. GT10 is quite elderly, and the reason we picked it was the least power. So it is... It's actually one of the highest power-consuming components in the box is the graphics card. Which, you know, would be nice to find... If we could find a lower-powered PCIe graphics card, then the numbers for the base idle at 32 would go down quite substantially. There's just one PCIe slot, right? No, no, there's three PCIe slots. There's three. So you can have a GPU, an AI accelerator, and what else? Well, sound card. I'm an audio guy, historically. I used to do GSP, so plugging in a sound card is a nice thing to do. And you can get a very high-spec sound card to slot in PCIe slots. At home, I've got HDMI running out, and then audio comes out the side of my monitor. And there's all kinds of other things, potentially, there could be more memory. Fast memory, fast networking, fast all kinds of things in PCIe. It's got a good bit of ethernet. It's got a fairly performant RAM. Like I say, you can put quite a lot in it. It goes up to 64 gigs. And how many other linear engineers have been able to use it constantly for four months? Are you like the record holder? Or do you have some other guys, too? I think I'm the only one doing arm-on-arm on developer box at the moment. There's a couple of people in the server groups who are doing arm-on-arm on more expensive boxes. Like they have the big server chips, maybe? Yeah. Or in some cases, some harder-to-find boxes that they don't make anymore. Do you have to have one of those? In my garage, but not in my study. The other thing about these things is they're pretty much silent. They're passively cooled. The graphics card is passively cooled. There's not a fan. It's passively cooled. You have no noise. And that's quite special. And how about your electricity bill has gone up a little bit? My electricity bill hasn't gone up. I can't really measure the effect of this on my electricity bill. There are many, many people in my family. A few cents, right? I mean it's been up like a pound or something over the last four months. Well, it will have been using less power than what I was using before. So I would have saved something, but I couldn't tell you what.