Hi, this is Wayne again with a topic “Checking Out Intel Data Center GPU Flex @ Innovation 2023!”.
Now, one of the hottest things that I saw at Intel – Innovation 2023 is Intel, is getting aggressive, very, very aggressive with their Flex gpus products for both vdi and machine learning. On the vdi side, there’s not going to be a subscription fee or a license model, or anything like that. They’Re just going to sell you the hardware which is extremely disruptive for where we are now competing products require ongoing subscriptions and they are ludicrously expensive. So I’m actually going to try to get my hands on these and what I saw was pretty exciting. So, let’s take a closer look, my trip to Intel Innovation, 2023 and this video are made possible by sponsorship from Intel. My opinion and my thoughts on this, however, remain my own, I’m here with Landon from Intel and we’re going to talk, srov virtual desktops and the Intel Flex 140 gpus. These are the server side gpus that I’ve mentioned in other videos.
They got a pretty killer demo. Here, so why don’t you walk us through it sure uh? What we have here today is. We have uh a couple of flex 140 gpus that are uh set up, so we can show the difference between what a um a virtual machine is like when you have it Hardware accelerated versus not Hardware accelerated um, and then we have uh.
You know we were talking about our licensing model where we don’t have a per user license or a subscription. You know you just buy our Hardware, you get it and it works. The the idea here is that you put a bunch of these in a server, and this is a GPU in a server, but it’s made for remote clients, there’s no output or anything like that.
But the reason that you do, that is because you can access Hardware, encoding and Hardware acceleration, even if you’re not playing back a movie. The visual display, that is, an application running on the server, is sent to the client in a compressed efficient, low, latency format, and so the user experience is demonstrably better. That’S right, so this is actually running in a server that uh I have back in Oregon and then we’re streaming all of that in a compressed format back to these thin clients that are that are uh here with us in San Jose. So so today we support um, I’m actually showing this demo, it’s running uh, esxi version 8.0.2, VMware, yeah, V, VMware yeah and then uh.
We support you buntu as well um yeah. So let me show you the actual demo here today. So you know what what I have is. I have uh on the left hand side over here.
This is your traditional VM that you would have. That would not have any GPU Hardware acceleration and you know as wendle. You know he moves the globe around. You can see it’s not a super, not a super smooth experience, but when you hardw accelerate that with a flex GPU, you know it’s much more smooth, it’s kind of what a user would expect when they’re, you know using Google Earth on their their client and same Thing over here you know I’ve got I’ve, got blender right, so blender I’m going to just move this around and this once again the left here.
This is not GPU accelerated, so I’m going to move the mouse around and it is doing stuff give it a second. If, if you want to experience this for yourself, all you got to do is uninstall your graphics, driver and you’ll have the same experience, yeah yeah and then you know on our Flex, GPU it’s nice and smooth. But what’s interesting, let’s see if it’s kind to me when I do this: let’s pull this up. You notice my uh CPU utilization didn’t Spike, but you know when I I did this on my CPU only VM. You can see how my CPU utilization spiked up to 100 % cu. The CPU has to do the rendering exactly so you know what does this allow you to do? You know, let’s say I had uh four uh vcpus allocated to each of my VMS uh. It’S not a great experience for my users if it’s using 100 % of the CPU to do that 3D. Rendering um and you know, I’m also eating a ton of my system uh my server system resources. So if I actually add a flex GPU in this guy right here is only running on two vcpus and you can see the the CPU utilization was still much lower. So that allows me to increase my existing VM density too.
So I could even add a flex CPU into an old server uh and you know, increase my VM density over and give my users a better experience than than what they have today. I go so far as to say that if you’re doing vdi in your organization today – and you don’t have some type of Graphics acceleration, you are literally and figuratively doing it wrong. Okay and then uh. You know.
The other last thing I want to show is uh. You know you can hook up multiple displays, so what we have here this is one Thin Client that is actually powering four displays and, like you know, while while we’re doing this, I have one video decoding here. I can start another video decoding over here, so be multiple videos. Decoding kiosk is running off of vdi lot, twist yeah yeah, and then you know I I have my GPU resources down here and I could be doing stuff in Excel as well. You know all of this is hardware accelerated and it’s it’s smooth. It’S good. It’S a great experience.
If the vdi side doesn’t tickle your fancy, there’s also the red hat integration you see. Intel has a device management plugin that would let you seamlessly shift between uh, vdi type Solutions and machine learning, type Solutions. You could go back and forth on the open shift platform and we know that the flex 140 Flex 170 gpus their their desktop counterparts.
The a770 uh that level one has already been working on off label uses for the a770, that is around sharing multiple having multiple virtual machines, share resources of a single GPU, officially sanctioned and fully supported in in Linux, open environments hyperv as well, but also more exciting. Coming news, I’m going to try to get my hands on some of this Hardware, I’m going to compose it into my new liquid fabric, but that was really exciting for Intel Innovation 2023, and I can’t wait to get my hands on some actual real hardware and explore This further because this could be extremely disruptive for vdi, I’m here with Eric from Intel hello and we’re going to take a look at Red Hat, open shift, but red hat open shift for AI, but also how easy it is to move back and forth between CPU And GPU inferencing on the new Intel platforms and also it’s a it’s a flex, 140 demo, I’m all about the flex, gpus and uh, some off Lael uses of those, and it’s really exciting – to see them in a real world use case sort of scenario here. So hello, hello, so I’m showing redot open shift data science and, as mentioned, I’ve got a flex 140 card. This card has eight XE car, eight XE cores per GPU for a total of 16.
That also means with two gpus. You get twice as many transcoding units supports the av1 video Codec, which is the open source alternative the h.265. We also have the flex 170. Now this isn’t in My Demo, but this is 32 XE cores for one single GPU, so we also support 31 virtual functions per GPU. So on this card you get 62 versus 31 on this one now shifting over to open shift.
What I’m actually showing is our integration into red hat, open shift using our Intel device plugins operator? If you look at the different installed operators on open shift, the IDP is where we are running all of our Upstream device plugins to expose Intel accelerators. Currently, we support Intel, gpus, the flex series and Intel software guard extensions, which is the uh Enclave page crash, the encrypted memory, where you can run your applications in that’s been around for a few years. This will expose the memory resources, but I’ve got a demo coming up of the secure en clave stuff and it is really cool what Intel has in their software Center that you can just download and start using immediately and it plugs right into open shift.
This is Intel’s device, plug-in operator plugin for open shift to make it as easy as possible. Exactly that’s exactly what we’re doing now within Red Hat: open shift data science, I have these different, open, Veno notebooks and if you look across you’ll see, I have one that says: CPU one that says GPU and if you scroll up just a little bit I’ll see If I can find it, this is always hard to find to switch between running a model on CPU versus GPU is is easy as just changing this G to a c, and this G2 C recompile the model and rerun it. Now, I’m not going to run through the model, but I just ran through it, so you can see there’s a trailing edge of my GPU usage of one of my gpus there’s two gpus that are actually on the flex 140. One of them ran the workload and my CPU went up just a little bit.
If I were to rerun the same workload, I’m running a 30 CPU container, you would see all 30 CPUs shoot straight to the top. This is really really exciting, because this same kind of thing is also available in Intel’s Dev Cloud that they just announced well la they kind of launched a long time ago, but they kind of didn’t. But it’s sort of getting some new stuff. Dev cloud is where it’s at from Intel to be able to play with this kind of stuff immediately, but you can also just deploy it in your own open shift container and on your own hardware and you’re.
Ready to go is that I mentioned these are certified device plugins. Oh, that does make a difference because picture nobody ever got or nobody ever got fired for buying certified Hardware. So we actually have a team that validates and certifies the driver for each open shift release each major open shift release, that’s what you’re paying for well, we don’t charge the red hat support to be clear. The red hat support I’m here with Jordan from myac hi.
I’M Jordan Leong from MAAC tyen computer, we’re going to take a look at some chassis that are set up for Sapphire Rapids. Fourth generation tell us about myac, okay, okay! So U my tech was founded by a former Intel engineer and uh in 1978. He left the US and went back to Taiwan to found his own uh business and many years later, many decades later. Mch is now a80 billion company.
They make a lot of servers. The Thunder HX s 5652. This chassis is pretty interesting because the drives are on the sides. Look at this.
Yes, they put 24 3 and2 drives into you. I need this for my home lab up to uh yeah, oh up to nvme wow, exactly oh yeah! If you put them yeah, oh yeah, we have a. We have a card right there. If you put two of them, you get NV me here. Nv me there. All a day, that’s pretty cool, it is pretty cool, so you can mix and match. If you wanted to have 2 and 1 half in drive edsff, maybe someday someday, maybe someday uh and it’s a single socket, Sapphire Rapids platform today, but Dro in ready for Emerald Rapids tomorrow. This looks to me like Intel DSG, but wait a minute.
You guys were’re doing DSG the whole time. Yes, that is correct, so pretty much what it was with mych uh mych was the odm for Intel DSG products. So what happened was we were always providing the L6 versions to L6 for Intel? Then they make it to the L9 to the RX now for people that don’t know DSG like what is that? What problem is that solving so DSG? Is the data set a solution group, so pretty much uh? If you want hyper converge, if you want uh HP or HPC uh, they actually accepted, or they made the servers for that. So the ultimate and vdi servers you got to get the flex gpus so pretty much right. Now we have the two servers uh on display. We have the Fox free pass, which is the Dual Intel Sapphire Rapids. We also have the not Ser, which is the 4u or 2u four nodes on display, but it was very modular.
You could do a 22 node as well, which happened to have like the pcie acceleration and GPU acceleration, so this chassis can be 2u2 node or 2u 4 node, just depending on the density that you’re going for memory single socket dual socket. Yes, so it’s going to be always dual socket when it came to Intel, especially on these lines itself. Uh future platforms will be maybe changing over to the single socket, but as of right now, Intel has always been dual socket.
Denali pass the Denali pass nice so, and it’s also available in both air cooled and water cooled solutions that looks like it’s perfect for uh Flex, 140, two Flex, 140 gpus per node bunch of compute. This would be really highend, uh vdi setup. It was set up. Uh mainly for management itself, but then there’s nodes, there’s a GPU node as well, which is a 2u high. So you get two2 node and you can put two uh two uh full height, full width, gpus per and and then on top of that, the later the later technology of OEM modules, all right so pretty much. What this one is, is the RB 7129, their tie-in server itself, it’s a GPU server, so it is equipped with a daughter board on top which could uh put 10 gpus in it.
That is a crazy amount of pcie slots. Look at that exactly top of that. It still has the O the ocp 3.0 slot too, so you get the networking to go with it as well all right, so I can put 400 gig e internet and run a bunch of gpus on a dual socket platform that is ready to go for fifth Gen Intel zeeon scalable exactly so. You’Ve also got room for fpga in this chassis in the front or other PCA.
Peripherals, not gpus, cuz gpus, don’t make sense, could probably do raid controllers here. The broadcom cards would probably work well, but yeah, fpga, plus GPU, plus networking plus I mean it’s all uh pcie storage they’re using plx. So you get the full Lane connectivity between gpus.
So you can do ENT, GPU memory sharing in this in this chassis. If that’s your thing, that’s that’s! That’S actually really hard to find chassis that have this kind of a configuration directly to the CPU with a full uh pcie connectivity. I managed to find kokia at Intel. Innovation and they’ve got their flash products which of course, are ubiquitous in the data center, but also they’ve get the new cxl module flash as well in an edsf form factor, I’m wh. This is level one. This has been Intel. Innovation, 2023 going to be sprinkling content from this event throughout a whole. Bunch of videos be sure to check that out be sure to check out my video on the Intel, secure Enclave as well, because it’s like SSL certificates, but for programs, and you never actually give somebody a decrypted program. I think this is actually a really cool thing.
It’S a lot better alternative than something like denuvo, but as a developer, you can also protect your applications cryptographically in a way that is way less honorous on the end users than something like denuvo. But we’ll talk about that in another video. We’Ll talk about cxl. There’S a lot of things that we need to talk about, but Emerald Rapids is coming in hot December 14th.
I can’t wait. I’M wi this level, one I’m signing out. You find me in the level one forums .