 So we hear the HPE booth and who are you? I'm Dennis Floyd, I'm part of the Apollo 70 team and I'm holding the Apollo 70 compute note. It's the first armed compute note actually designed specifically for the HPU market. This is a one-use note, there's also a two-year note which allows for two presses or two GPUs. So where did the GPUs go? Where did the GPUs go? The GPUs, there's one slot here and one here and they're actually risers that they mounted to. So that could be like a regular NVIDIA 1080 or something like that? The original one is with AMD Fire Pro. So AMD GPUs that's especially made for supercomputing? Cool. So right here is the Thunder X2? Thunder X2 from Cavium. It's the next generation of the ARM SoC. And what are you talking about here and on the wall? You're talking about the whole cooperation? As part of our development model we worked with the Department of Energy Labs on a program we call Comanche which is essentially an early adopter program. We provided them with the Apollo 70 early on and they've been working to develop a software with it and to utilize it in their space. And so there's a software stack of AHPC, there's ARM Alinear Studio, AMD GPU software, everything just works together. And some stuff going on with Melanox and Red Hat, Susie. And what are you showing over there? The corner there. This is actually the two-year version of it. And the chip. Can we look under it? And on the back of it is to have all the connectors. And this one you can open? Yeah, I can open this one. So that's two. And it sits in the Apollo 2000 rack. So it's a known rack we can use it. Okay, so standard size. And let's jump over there. Can you introduce these guys what they're doing over there? This is Lisa Pilotti, I'm short on the machine. Hello, so who are you? My name is Lisa Pilotti and I'm a senior system engineer working for Helipackered Labs. And what are you showing here? Is this an ARM solution? It is an ARM solution. It is the first instantiation of the machine of memory-driven computing architecture. This is the machine? This is a prototype of the machine. It's called the machine? The machine. Why? Because the idea is that the technology that we're developing as part of the machine program can really be used in a variety of different ways. It can be used in servers for an exascale-type solution. It can also be used in smaller systems. So we called it something a little bit non-specific to re-emphasize that it's really the technology and the architecture that is the focus of the program. So there's a whole bunch of special things going on here, right? There's an ARM server here? Yes, this is a high-performance ARM processor. It's a Cavium Thunder X2. Thunder X2 right here? Correct. And then what's going on around here? So out of the ARM processor, there are ICI links that come out. This is an FPGA that provides translation from the links coming from the processor out to the other parts of the system. So this is a memory semantic fabric. It's the precursor to Gen Z. So once you're at Gen Z, you can talk to fabric-attached memory anywhere in the system. For instance, over here... Gen Z is right here. It says Gen Z. So that's a new idea? Yes, it's a new fabric. So think of Infiniband or Ethernet. It's a new memory semantic fabric. There's a consortium of, I believe, over 40 different people, different vendors, server, operating system, networking components, a wide variety of partners. So was that originally your idea or somebody at HPE, right? And then they decided to open it up? Well, it was really a group effort. We certainly are one of the 12 founding partners, but it really is a collaborative effort across all the different vent companies. So is the idea to have one ARM chip with a ton of RAM, or is that what it is? So what you really want is you want to have, with memory-driven computing, you want to have a very large pool of non-volatile memory. And then you want to have processors that talk directly to that memory. In today's processor-centric architectures, you have memory that's attached to processors, and if you need more memory, you have to scale them in lockstep. And then this processor has to talk to that processor to get to this memory. What if you had a really large pool of non-volatile memory and the processors could talk directly to that memory? And because this memory is non-volatile, you don't have to be moving your data from main memory to storage and back and forth. And what if, because you're talking to that large pool of memory with the Gen Z fabric, any type of processor can talk to that main memory. So if your application wants an x86 processor, great. If it wants a GBGPU to run on it more efficiently, great. So you get to pick the processor and the application that works best for your application. So what are those cables going on over there? So this is just an electrical interconnect from that it has the Gen Z fabric going off to this large pool of fabric attached memory. Here we have four FPGAs, and they are memory media controllers. So they're converting the Gen Z fabric to whatever memory or storage media you happen to have. In this case, it's DDR4 memory because that's what we could get for a fairly low cost. So we have 160 terabytes in four enclosures, and so we could get that in DDR4 first. How much terabytes? 160 terabytes of RAM. In four enclosures, but that's 24 terabytes. How many on processors run that? So there's four terabytes and one processor per node, 40 nodes. So four terabytes of RAM? Just on the node. Correct. Four terabytes is a lot. People usually talk in gigabytes, right? Yes. Four terabytes is a lot. 160 terabytes, exactly. But that's the idea behind this memory-driven computing architecture because now if you have that large pool of memory and it's non-volatile so you don't have to worry about moving it out to storage because it costs too much to refresh that amount of memory continuously, you can just keep it there. And so it's readily accessible by all your CPUs when you need it. And you're out paying to refresh it if you don't happen to be using it. And the Thunder X2 is great at managing four terabytes of RAM? Yes. So the idea again behind the architecture... It has no issues with that. It can just work with so much RAM. Well, we are up to four terabytes, yes. Once you get to a certain level, you run out of physical addressing. So we have a scheme where we have aperture windows. So only so much of the physical memory is visible to the processor and then we switch apertures if what you need is outside of that window. And what are you talking about here? So we have a Linux team that has made changes to support a memory-driven computing architecture. We call it Linux for the machine. And so they submitted over 4,000 submissions to Open Source because we're kind of building the ecosystem around this memory-driven computing architecture. That's why we're part of the... This is a logo right here. The machine. That's correct. Sounds like a movie or something. Well, we were in a movie. We were mentioned in The Ghost. Really? This one was Scarlett Johansson. Really? The Ghost? My colleague, my school friend was acting in that one. Oh, yeah. It was the second role. So in The Ghost, they're talking about how much RAM using per ARM CPU are. They're not mentioning... They just very fleetingly mentioned the machine. They're not like... They don't go on and on. Yeah, it's a great way to add some RAM to the ARM processor. But they didn't mention that. They did not. So the machine is eventually going to take over everything? Is that what happened? Well, because there's an explosion of data, right? I mean, there's still going to be need for your traditional processors, but they're estimating that the explosion of data is coming right at the time where our ability to keep up from a computer point of view is running out because Moore's Law has served us very, very well for many, many years, but it's starting to run out of steam just at the same point that the amount of data that we're having to deal with because we all have our six or seven devices and more is occurring. So that's why we think a revolutionary new architecture is needed like memory-driven computing to deal with that amount of data. Is this video also talking about that? So this video, yes, is on the machine. So here we're running an application on it. Here you can see some of the video of the machine. And here this node is getting plugged into an enclosure. And we have people talking about it. So how soon is this going to take over the whole supercomputing market? I think it will take some time. It takes some time to build a completely new ecosystem. Why? Can't you just get a few more extra help and stuff and get it done? Well, that's why we have submitted all the changes we've made to Linux to open source, to GitHub. We have a machine users group so that people can learn more about the machine and think about how they can make modifications to their applications to take advantage of it. We're part of the GenC consortium. We're working with vendors to support their processors on GenC and in memory-driven computing. And a more basic Apollo 70. Is that a stepping stone? You could think of it as a stepping stone. So this has the same processor as the Apollo 70. But, again, here's our product with the ARM processor. And why is it shaped like this over here? What's over here? So these are switches. There's four switches. You have a whole bunch of those and this is a different place in the rack. What happens is that this is a single node. So it has one processor and up to four terabytes. This is the processor, sorry. You flip it up vertically and ten of them plug into this set of switches. And these switches are GenC and GenC out. So they allow any processor to access fabric-attached memory anywhere in the enclosure. Do you talk about, is this FPGAs too? These are FPGAs. Do you talk about what FPGAs you're using or are they secret? So we have not told which vendor of FPGAs that we're using. But the vendor that you're using is probably very excited about this. I would hope that they are. Or they're trying to fight to be the one that provides the best solution for this. Well, they're obviously very high-performance state of the art FPGAs. So we're helping them push the extremes of their technology. And what would you be able to demand? Because the Thunder X2 is great, right? It's fantastic. Right. But you have some demands for Thunder X3. What do you think would be great for the next ARM chipset? Would it be a support more than four terabytes? Oh yeah, I think that we would always love to be able to support a larger set of physical address memory space. That would be great. It would be wonderful if processors had the GenZ links incorporated into the processors or itself, because then we wouldn't have to go through a bridge chip and we'd have even less latency to talk to that fabric attached memory. So you just ask them, they will add it, right? They're very good at adding a lot of stuff in their SOC. I am sure that Capium is part of the GenZ consortium and so I'm sure there are a lot of discussions happening. I'm not part of the consortium, so I don't know. Where's the bridge chip? This is the bridge chip. Is it an FPGA then? Yes. So you use an FPGA, it would be nice to have it. There's some other FPGAs around. Right. So there's not only those, there's some more. Right. I mean, you can only have so much because what you want to do is with your memory media controllers, you want to go from GenZ and you want to be able to support different types of memory or storage media, right? Cool. And so the biggest cost right here is around, right? The arm chip is kind of relatively low compared to everything else. When you're using really large 128 gigabyte DIMMs, yes. That's the highest cost. So hopefully the cost of RAM is going to go down or something. I could help this. The cost of memory does go down over time, right? So you use the largest capacity DIMMs available at this time.