Open Compute Project

Author: tomoakisato

325

@@ -1 +1,32 @@
-(Negotiations are underway with the YouTube channel concerned.)
+
+Okay, you guys, the day's almost over, right? We're getting to almost the end. Does everyone need to get up and stretch? I can support that. If you guys want to get up and just stretch for a second, let me make sure I know where I'm going on the foils while we do this. Let's see. Okay, so everybody back up. So, I've done this presentation a number of times. There's not a lot of changes, but I'm here just to give you kind of what's going on really in the world. So, we've seen a lot of presentations that get really into deep dives on the technology. I'm not planning on doing that. We've done a lot of presentations on where CXL is and why CXL is needed. I'll briefly touch on that, but what I'm going to do is talk to you about an actual use case today, and I'm actually going to talk to you about the market. Where is the actual market? Where is the market today? Where is the market tomorrow, and where is the market in the future? So, that's the real focus of my presentation.
+
+So, who the heck is Montage? Does anybody else? I can't see you. It's like blinding. So, I'm assuming all of you guys have your hands up that know who Montage is, right? I think I see one person waving at the back of the room, maybe another person back there. So, Montage itself has been in business for about 20 years. We are really good at not telling people about us. So, our website is not very good, I'll be honest. We sell to engineers, and our customers are Samsung, Micron, SK Hynix, and other people. So, you guys all know who those guys are. We sell them parts that go on their memory modules, and we feel we're like number one in that market. The second thing that we are really good at is listening to our board of directors and our investors, and one of those is Intel. And Intel told us to get into the PCI market a number of years ago because they said, "Hey, that market is accelerating year on year." And we think the market is prime for re-timers. You're really good at memory products. You should be really good at interface products. So, we make PCI re-timers. Well, what is CXL? It's the merging of memory products and PCI backbone. So, of course, what did Intel do? "Hey, you guys should get into CXL." So, we listened to them, and we got into CXL, and we feel we're doing really good in that space. We feel we're number one in that market right now with design wins at Samsung and Hynix. But did Samsung and Hynix, when they presented? Did they tell you whose controller they're using? No, but they're using my controller in their products. So, again, we kind of keep below the radar, but we make very good products in that space.
+
+Okay, you guys have seen this slide time and again. I'm going to spin it slightly differently and tell you kind of what's going on in the market today. So, the market today is two areas. One is an E3 module, which Micron did a stellar job talking about. That market is for people who instantly want more bandwidth and more capacity. Now, that can either be a DDR4 instant add-in or a DDR5 instant add-in. I need more bandwidth. I need more capacity. I don't have any slots left on my memory channel. I'm going to plug in a DDR4 E3 module, or I'm going to plug in a DDR5 E3 module. Works today, not rocket science, and it's the way to go. The market is fairly small because there are available slots. There's lots left, and DDR5 pricing keeps coming down, so that market is more for "I don't have any room left. I need more capacity. I need more bandwidth. I'm going to plug it into the E3 module."
+
+The second market, which we really thought was going to be huge, is turning into reality today, not being very big; that's the add-in market. So, it was mentioned earlier that Intel, AMD, and all the ARM products today only support DDR5 on their platform. Well, DDR4 has been shipping for over 10 years now. There are millions and millions of DDR4 modules out there. But if you want to use the newest, latest Intel, AMD, or ARM-based server products, you can't use that memory anymore. This was the first time Intel, AMD, and ARM did not have dual support in their controller. They were thinking that the DDR5 prices were going to be rock bottom when they launched the platform, and so I'm all in for DDR5. The industry will transition hard, and everyone will go right over to DDR5. Reality, it's been one of the slowest transitions ever on a memory technology because AI came out, and everyone said, "Where's the HBM play? Where's that market?" So, that market transition to DDR5 has been very slow. There's still a lot of DDR4 being sold in the industry, so people are saying, "Let me reuse my DDR4." But now this is where the complexity comes out. We montage. I've shipped over 1,000 add-in cards. Most of them were free to the companies around the world. So, they've got them. They've been playing with them. They plug in the DDR4 module from Micron, or Samsung, or Smart Module, or Kingston. I'm trying to hit every company I see out here that makes these DIMM modules onto the DIMM, then I plug it in. Where do I plug it in? How do I cool it? How do I verify that the DDR4 module is going to work and the longevity on those parts? So, that's what's kind of happening in that space right now. Is it possible? Could I use it today? You bet. Has anybody implemented it in full mass production? No. But the possibility is there, and it does have opportunities to be used in the market today. So, that's the two markets for CXL today. One, I'm going to put my controller on an E3 module, and I'm going to give you DDR4 instant capacity and bandwidth. Or I'm going to put it on a DDR5 module and give you instant DDR4 capacity and bandwidth. That's the E3 modules that Samsung, SK Hynix, and Micron are pursuing. The other option is the add-in card option, which Marvell talked about earlier. And again, you can get that in a DDR4 or DDR5 add-in card, adds compression. As you add all these features, you're adding dollars, and you're driving up the cost. Where are they going to cool that? How are they going to cool that? How are they going to cable that? These are all the challenges we need you smart guys out in the room here to figure out, because that's where we're at. We're in the crawl phase of CXL transitioning to the walk phase. So I've done my whole presentation. Any questions?
+
+No, just kidding.
+
+Let's go to the next slide. Okay. So, this gets into the nitty-gritty. A number of you guys have probably seen these presentations. A couple of years ago, I worked with Yule Technologies, Yule, the consulting company, and we put together a market forecast for CXL. At that time, three-plus years ago, we did not take into account a number of key technologies that were in the process. One, 32 gigabit, which got accelerated by the DRAM industry because of the big push for AI and high-capacity DIMMs, and the other technology that took off was MR-DIMM. So, this chart is very different from the Yule chart that was presented a number of years ago because this is more changing and morphing on what's really happening with CXL because we continue to be in the crawl phase as we get companies developing the hardware and as we get companies developing the software to optimize for the usage. Micron did a stellar job of talking about some of the use cases. So did some of the prior presenters. Now, we need the rest of the industry to start optimizing their software for it. So, this is broken into two areas. You can see CXL expansion. Got an E3 module. I plug it in. I get capacity and bandwidth. You can see that the market is extremely small, probably 10,000 to 20,000 units this year. Next year, we're hoping to 5X that, being 100,000 to 200,000 units. If you look at the old Yule presentation from a couple of years ago, it was forecasted to sell over a million units this next year. So, not happening. Again, I love the technology. I'm leading the controller industry. Again, you don't hear about me, but I'm leading the controller industry. But I'm also doing very well on memory. So, it doesn't hurt that the memory is still taking off and doing well. But I'd like CXL to start taking off. So, it's not like I'm Mr. Negative and I don't want this market to happen. I definitely want this market to happen. But it's only going to happen when you and I, and others, get together to figure out the right use cases and the right value propositions in the TCO.
+
+The next market, next data point, is to the left; it is the expansion option. And the next is the pooling option. So, these, when you combine them both, you'll get the next charts showing forward. So again, the expansion option is, "I just need more bandwidth. I need more capacity. I've got an E3. Let me plug it in and take advantage of that." The chart to the right is the one that's only being talked about a little bit right now because, with pooling, ideally, you need Diamond Rapids, which I don't know if I'm supposed to say that word or not. It's Intel's next platform. Also, you need the follow-on deterrent from AMD. Those aren't supported until 2026. That's when CXL 3.0 will be readily available at that interface level. But you're going to need, next year will all be about providing samples, proving out the technology. Then it comes out in '26... Well, the price points of the switches are probably going to be too high. You'll have to wait until the second generation of the switches. Thus, you see the big influx from '27 to '28. Now you've got the second generation of the switches. The price points will come down. So now the switches aren't just at the very high end, but they're at the next tier and the next tier down of customers. That's the hockey stick that will really take off. Would I want all this to happen sooner? Oh, please, let's have this sooner. But this is kind of the reality of what's going on. Crawl phase, walk phase, run phase. And so, we need you guys to start doing the software optimizations. We need to figure out from a hardware standpoint, where do we put those add-in cards? How do we cool those add-in cards? How do we cable those add-in cards? Those are definitely a number of challenges that the industry can solve to pull in these opportunities. But you will not be able to use pooling truly until Intel and AMD support that, and that doesn't take place until 2026.
+
+This is then converting it to units. So, you can see that's 1 million units, and that's 1.1 million units from a pooling standpoint. So, the market is there, but it will take a while based on where things are at now and where things will be in the future. We would like to accelerate that and move things forward, and that's why also the industry is looking at other media sources, not just using DDR architecture, but using a cut-down version of DDR, or using phase change memory, or other architectures that can set behind the CXL controller that have a cheaper price point than DDR5, or DDR6, or LP memory, or HBM memory. That's really where this technology will probably take off: when the media that's being used is cheaper than a native DIMM solution.
+
+Okay, so what is Montage trying to do to help this? We're kind of trying to help you guys out by developing an engineering vehicle. So, see me after the meeting, or go talk to Kyle if you want to, right out there. He's working at our booth, and we can get you guys access to an evaluation board. So, you can start playing with it from a DDR4 interface standpoint, or a DDR5 interface standpoint. We were first to market with the technology. We hit a very good price point, and we had support for both DDR4 and DDR5 in our first product. Our second product was, "You've got to lower the latency. You've got to lower the latency. You've got to improve the bandwidth. You've got to add RAS characteristics." So then, we created a second product that supports only DDR5, supports up to 6400, two DIMMs per channel, lowers the latency by 30% to 100 nanoseconds, and supports IDE support. So, we offer both products. We've got a price fighter. We've got one that supports DDR5. We've got bandwidth. So, again, depending on what you're needing today or your customers are needing, we can give you different options to work with from an engineering standpoint. Am I in mass production on these products? You betcha. I'm a silicon provider. I sell the controller. But to enable the industry, I developed this evaluation board. I have a number of partners that I'm working with that if you want an add-in card for mass production, you'd want to work with them both in China, Taiwan, and in the U.S. So, if you want to go to mass production on the add-in card, see me afterwards also.
+
+So, this is, again, a little bit more detail on what our MXC evaluation board provides to you. You guys can read it. I'm not going to go line by line. I know we're way behind schedule. You can whip out the cameras and take pictures. You guys can all have access to this later. Again, I'll be available after the presentation. But we have this evaluation board available today that has the hooks in there for a DDR4 solution or a DDR5 solution. So, you can actually start playing with CXL. You can start doing your software optimizations, your hardware optimizations. We have this available today, both around the world. We can ship this product.
+
+As I mentioned, a lot of people don't know about us. We're under the radar. So, I just want to show you we are actually in the SK Hynix product. We're actually in the Samsung product. Sadly, when they show their stuff, they blur out what the controller is. But, we're the controller that they're using in those products. So, we've got the design wins. Going forward, it looks like all the design wins are going to us. So, in the next generation product, it looks like all three vendors will use us going forward too.
+
+So, you saw the pyramid. Micron's pyramid was way better than my pyramid. Those of you guys who were around in the Optane days, you saw this kind of same foil from that. We just repurposed, took out the word "Optane," and shoved in "CXL." That's kind of really what's happening. In order for this, again, technology has some similarities to Optane. It's really bringing value to the table. But, you have to write your software and do some optimization around it to take advantage. The one big advantage this has over the Intel Optane days, and I'm ex-Intel, so I know, was it's widely supported. It's not just one vendor making Optane. This is the industry-wide is behind CXL technology. You're seeing Samsung presentations, SK Hynix presentations, Micron presentations, and then the whole ecosystem is behind this. But, we need the software and hardware optimizations for this to take place.
+
+Okay. All major SOC suppliers are on board. We, again, have a very large lab presence in Atlanta for people who want to work here in the U.S., where you can come into our lab. We've got the latest Intel, latest AMD, latest Ampere, latest ARM systems, and we can co-work with you to do any of the work right there. We're also the lead vehicle for Intel and AMD checkouts on the CXL controller to check out their products. So, that's one of the advantages of working with the leading company. You'll get the advantage of working, also get the exposure there. So, we're doing that from a standpoint. We have the interop lab capabilities. We have access to all the various platforms, and we do try to do the heavy lifting so it's easier for you. Once you've purchased the product from us, we've already done the evaluation and qualification for you.
+
+Okay, demos. We have this outside, but how many of you guys have been by a server in the server room? Again, I think you're all raising your hand, right? They're loud. So what we did instead of bringing it here, because it sounds like a plane being taken off, we've got pictures and videos of it. We've got pictures of it in our office in San Jose. So, we actually have this system running, and what we're trying to show is that it's real. It's not magic. This was a Granite Rapids System. We plugged in a DDR4 module from Micron. Oh, no, that doesn't work. It, right, it doesn't work. We had to put an add-in card that had a CXL controller on there. We plugged the DDR4 module from Samsung—I mean Micron—into that add-in card, and now it works. Then we plugged in SK Hynix's latest E3 module. That was a DDR5 product. Then we plugged in Samsung's E3 DDR5 module. Then we plugged a Micron DDR5 module in. So now, I've got DDR4, DDR5, Micron, Samsung, and Hynix all running in the same system at the same time. So, again, you guys are like, 'Oh, wow, that's cool.' That is cool if you know and understand how complex each of these memory suppliers are. It's supposed to be commodity memory. It's supposed to be interoperable. That doesn't always happen. So, this did not take any heavy lifting from us, or blue wires, or special coding in the background. It works. It's running. We instantly got performance, bandwidth advantages, and capacity advantages. So, that's a real-life example.
+
+Second example is performance. What does the performance look like? Again, you saw some great stuff from Micron. I'm not a performance guy. We tried to throw this together to show you what the performance could be. But, again, we ran spec. From 2012, we ran four or five different workloads. We got a 32% to a 72% performance improvement. But, again, I look to you guys to do this kind of stuff. I just make controllers. And I try to make the best controller in the world. And I want your input on what features do you want? How important is compression? How important are those reliability features that you need? What is the price the most important thing? Is latency the most important thing? So that we can develop the best product for you to meet your needs today and tomorrow.
+
+I think that's almost it. Yeah, call to action. Now is the time to invest. We're in the crawl phase. Let's get to the run phase quicker, sooner, faster. How can we do that? Okay. We have a controller. Okay. I can develop, get you the controller part. Or I can get you an evaluation board. And you can do the development work there yourself. You can get parts from my customers, Samsung and Hynix today. Everything you buy from Samsung and Hynix is one more product that I sold to. So that's good on that one. We have demonstrations available for bandwidth and capacity outside. We can show you that demo. We can bring it into our lab in San Jose. Or we can bring you to our interop lab in Shanghai or in Atlanta, depending on your location where you're at. Multiple demos are out there. Again, this technology is real. It's available today. And it's changing and morphing and improving as we transition from CXL2, which is a good product. But then we get to a better product with CXL3, which supports the pool disaggregation. And that's really why you have the hockey stick. But unless we start getting behind the technology today, writing the software optimizations, we won't get to take full advantage of that pooling when it becomes available. So, I'm all done. I think I kind of got us back. Hopefully, you found this educational, beneficial. I try to come at it from a different angle than what you heard today. I'll be outside if you have any questions or comments. Thanks, everybody.