as gamers we're running machines with four, eight, maybe 16 cores if we're lucky, so when AMD launched their first 96-core EPYC CPU, I figured well, we're done here, that ought to be enough chip for even the most dedicated Potato Head, but the

Spud Masters at AMD said no, and last year launched a 128-core CPU, surely that must be enough, right? No, it wasn't, and thanks to Super Micro who sponsored this video, I have the brand new AMD EPYC 9965, a CPU with not 96, not 128, but

192 cores, making it, alongside its Turin family members, the fastest CPU on the planet by a wide margin. We're talking 384 threads, terabytes of DDR5 ECC memory, 128 Gen 5 PCI lanes, and a 500 watt TDP. I can only think of one word: what the...

[Music] I never thought I'd find myself saying only 500 watts of power consumption on a CPU, but when you consider that AMD packed 192 cores into the same space they used for half as many just two years ago, you gotta agree it's pretty

impressive. I mean, sure, that's more than my entire first gaming PC, but compared to the king EPYC at that time, this new chip has double the core count and the same maximum boost clock for real-world performance that utterly obliterates it. So, how did they do it?

Well, for starters, just like AMD's latest Ryzen 9000 desktop processors, this monster is powered by Zen 5, which comes with a bunch of architectural improvements like next-gen branch prediction, which allows these chips to predict their next move two steps in

advance rather than one, not to mention with greater accuracy and lower latency, all while increasing cache bandwidth, implementing a wider execution path, and adding a full 512-bit data path for AVX-512 workloads. If all that sounds like gobbledygook, that's fine. It boils down to a

claimed 16% average IPC uplift from Zen 4 to Zen 5 in PC workloads and potentially even more for high-performance compute and the enterprise workloads that typically run on servers. Let's get it installed, but wait. While 500W seems downright reasonable for 192 cores, it

does mean you can't just plop them into any old chassis and expect it to work. You need a compatible BIOS, sufficient cooling, and sufficient power, so keep that in mind. But it isn't going to be a problem for us today since Supermicro can build practically anything, including

this as a 2U 2126 HSN hyperserver that they provided for this video. This magnificent machine can handle not one, but two of these monsters, which is 1,000 watts of just CPU, that's 384 cores, 768 threads. Now typically, if you were to buy one

from Supermicro, it's going to come pre-assembled, but we went out of our way to take the CPU out so I can put it back in and show you how these new chips use the same socket as last generation. So, if you have one, you in theory could use it,

but you are going to want something that can handle more power. Not every motherboard is going to be rated to handle a 500W CPU right out of the box. You should check your motherboard manual. This one is purpose-built for this platform, so we should have no problems at all. M

PTM 7950 LTT store.com, it's a thermal pad and actually perfect for this application because it's super durable. These things can handle like hundreds or thousands of heat cycles while barely even losing any of their original performance, which is awesome on a CPU.

This big, it's important to make sure everything is tightened down properly so that the CPU actually makes contact with all of those thousands of pins because unlike most desktop CPUs these days, the CPU is actually held in by the force of the cooler, not by like a little bracket.

On top, I mean there is a bracket but that mostly just kind of holds it in place while you're tightening everything down. Order 1, 2, 3, 4, 5, 6 fully tightened one at a time. Yeah, not a little bit at a

time, not a little bit at a time, which is the conventional wisdom. You know, no, I tried to do that on one of the CPUs and it it actually didn't post. Look how scary that is. Gorgeous! Now, while we could have spec'd this machine with up to

9 terabytes of 6,000 megat transfer per second registered ECC memory across its 24 DIMM slots, that would have cost, um, probably as much as my car. I think maybe more than both of my cars. So I think the 1 and 1/2 terabytes that Supermicro

sent over is going to be plenty for us. This is 6400. I heard that AMD is doing like special models with certain providers that support 6400. That's incredible. I guess this is one of them. That's a big improvement. Hold on to that.

means standard server memory is faster than standard desktop memory what hasn't changed from last gen is PCIe expansion we still get 128 or 160 Gen 5 Lanes in single and dual socket configs respectively and this particular chassis can be spec'd for either eight eight

Lane expansion slots or four 16 Lane expansion slots if we have dual width cards they also crammed another 16 Lanes into an OCP 3.0 mezzanine slot here all of which well most of which we aren't using for now but you definitely could while

our hyper server is fairly well-rounded a solid setup for virtualization which we'll show you later or cloud workloads where you need a lot of cores and not a ton else aside from memory this chassis is also perfect for high performance computing and even AI I mean you've

already got the maximum amount of CPU so with the rear slot stuffed with high-speed networking GPUs accelerator cards maybe even all three at the same time there's not a lot that a machine like this can't do don't wait today is the last day of our Black Friday and

Cyber Monday deals at LT store grab some of the best deals of the year on screwdrivers magnetic cable management and more at lg.gblack Friday we're going to have it linked down below what we will be installing now is a few of

these 15 TB Kioxia NVMe SSDs in the eight slots that each support PCIe Gen 5x4 funny thing about this chassis by the way if you're building out high performance storage for AI or I don't know something else that needs fast

storage maybe with software like WKA, it can be configured with up to 24 2.5-inch and 12 3.5-inch drive bays thanks to the modular design of the front. So you could install an eight drive back plane like this one in each section with a bunch of 30 TB

CM7 and you could have over 700 terabytes of Gen 5 storage in a single 2U machine. That's a theoretical read performance of 240 GB/s across those drives in a single chassis, or more than 330 GB/s if you went with 15Ks, you know.

We get excited by fast SSDs in our gaming systems, but when you stack them up together in a stupid fat server like this, it's kind of mind-blowing. Speaking of blowing, if you're speaking 500W CPUs like we've got, or GPUs for AI or

rendering, you uh might actually appreciate the modularity of this chassis and the fact that it doesn't necessarily have all those drive bays in there because these are some flipping wild counter-rotating fans, but they're still going to pull air better if they

don't have a whole bunch of restriction in front of them. So the bottom line is whatever your needs are, Supermicro's Hyper systems and the entire rest of the lineup are designed with flexibility and modularity in mind. You ready to put the top on? Did you look at the power

supplies yet? 2,000 watts 80 plus titanium, two of them naturally. What are the odds this boot's first time, but those giant sockets even if you do it properly like I don't know this thing. Same 85%, 90%. You did it properly.

Yeah, but I wasn't here when you put in the other one, so, uh, she's going to go, she's going to go. These things take forever to post the first time. That actually makes sense. Think how long your desktop Ryzen takes to do memory training the first time.

Well, is it up? F yeah! Oh, look, oh that's great. And you doubted? Well, I didn't doubt much. I said 85-90%. All right, can I reboot it? Uh-oh, oh boy. I think it's set to Max. Yeah, I think so. We should just

leave it like that, right? Yeah, maybe not optimal speed. That sounds more manageable. Oh, thank God. Oh my God, that's oh, oh, that's downright pleasant. Wow. Now that we know it boots, we should probably put a GPU in there, right? And what better GPU to go with the oops, Top Dog AMD epic

Enterprise CPU, then a top dog, not Enterprise, but still awesome Radeon RX 7900 XTX. Now, despite this chassis being well equipped for GPUs, I mean the power connectors are like right there, so they just plug right in, and they've got a 2600 watt power supply option that you

can upgrade to. These slots are meant for full-height, full-length workstation and Enterprise GPUs, not weird shaped gaming GPUs. So our plan is to put our GPU outside the server with a riser and a jumpered ATX power supply. Nice. Flawless. Oh my God, it

scrolls. How much stuff can I do and still maintain 0% utilization? Could I launch a browser? Oh, I'm sure you can play a YouTube video. It's only 1%. Oh, no, it's it's at zero again. Once it settles in and it's doing its thing, it's nothing.

It's going to load one core. It's 384 cores, it could load two cores and still be below 1%. Duplicate this tab, uh, well hold on, hold on first I want to I want to try different, I want to try different resolutions cuz we're not even at 1080p

yet. Remember this is this is happening on the GPU. CPU works a little harder, it works a little harder. CPU is at 10%, so I think most of this is GPU. 4K here we go. There's like some cores, there's one core

I see at 9%. The rest of them are vibing, thriving, one might even say. Okay, the GPU is at 77% according to task manager. How is that even possible? I think it's just cuz it's video decode. Okay, let's disable hardware acceleration. I think this is

going to hurt a little bit. No, 2%. Yeah, it's going to hurt, it's going to hurt so much. Our memory is 0% too. It actually just says 0.0, but I guess to get to 0.1 we're talking 100 gigs. I need like a

script to to open these over and over again. I want to see how maybe we should just revisit that video on this machine. How many Chrome tabs we get, some more RAM, we revisit how many Chrome tabs, but we like automate it this time. But let's move on, let's move on for now. Uh, cin

bench, do you have R23 so we can see all the little things? Uh, I do, but it wasn't launching. Oh wow, it just does nothing. No, I know, doesn't even try to. I looked in task manager and you see it for a sec and then it goes away. I think it just

doesn't know what to do with this many threads and 2024 is going to be a supreme disappointment. What is thread limited to 256 for some reason I don't actually know why that's it. You didn't even say what the score is, if that score

was my child, I wouldn't even speak to it. Oh what a flipping disappointment, get out of my sight. You could try Blender ready CPU rendering, let's go. I feel like this is going to take longer just to prepare the job than to actually render it. That's a lot of blue, it's a blue

screen of success. Damn, 12 seconds. She doesn't get much clock speed though, does she? Yeah, this is also running with SMT on. Like we're kind of at the point where you have so many cores that I feel like a lot of applications are better off

without multi-threading or some might be anyways. These CPUs between the two of them have over twice as much cache as my first gaming PC had, system memory, like the first one I bought for myself. Yeah, okay, cache, yeah. And these don't even have that much cache. No, 7-Zip can't

perform that great. There's no way it'll go across this many threads, CPU threads. Oh, what, limited again? Let's go compare this to Intel's new desktop chips real quick. So 7700 GiB/s, which is 700,000 MiB/s, so it's like 3 and

a half 9950X's. Oh no no no wait wait wait, that's B, that's kilobytes per second. Yeah, oh I see. Yeah, so we're double the performance but like, okay, well now we should play a game on it, right? Sure, I mean we have to. This

It's like the worst possible CPU to play video games. Oh yeah, we're at 4K high settings. Oh man, that was cheesy. That was cheesy. Not going to lie. Okay, clearly this is not a gaming CPU. I mean, it's working for me. What can

I say? No, you could probably run, if you had enough GPUs, like 60 instances of Counter-Strike at this point, dude. No, though, should we do it? It's been so long. How many years has it been since we've done many gamers, one CPU? This time we

could actually do one CPU with 192 cores, dude. That would be crazy. Yeah, we should do it. I want to do it. Here, did you know alt+F4 makes the game run faster? Did you know that? Alt+F4 doesn't do jack in this

particular. No, I'm winning. I'm winning. I'm going to win. I'm going to win. Jake, Jake, I'm going to win. Okay, what's next? Doom Eternal is a good one, cuz Doom Eternal is actually like fairly multi-threaded supposedly. Here we go, horde

mode, 4K, everything Ultra nightmare, motion blur off, ray tracing off. Looks pretty good. It's 200 FPS. Uh yeah, wow, it actually does use some cores. Eh, yeah, but like I, I mean, relatively speaking, no, but that's crazy though. It's actually

using like 16 cores, I feel like, but we're running at 200 FPS. That means we could run multiple instances of Doom Eternal on this thing. Wasn't the CPU usage at 1%? Yeah. When I say multiple, I mean a big multiple. All right, this actually feels

Great yeah yeah click things he's almost dead kill him glory kill use more fuel buddy oh my God you're going to die J oh no oh no oh no leave me alone he called you an arrogant fool quick glory kill I did it yeah again yeah let's go yeah

This is great it's running lovely okay let's do something not stupid on it pass mark would be a good one cuz it'll show us like comparison to CPUs that are relevant I bet you in a lot of cases the numbers that we get running across both

CPUs will be the same as running across one or maybe even worse and that's just software limitations but you would never configure a system like this for a single workstation that's running one t a single workload yeah or at least if you threaded it across all of them you

would have them be well aware of the fact that there's two CPUs and the performance degradation you get when you try to cross them yeah and we we're literally off the chart yeah like a lot so nobody else has submitted the CPU yet so let's run memory Mark and then we can

submit it sure and then we can see what it's like compared to other CP all right one of the tests is just how much memory you have I think we're going to do okay on that one yeah maybe okay oh 63rd mid AF the read and the write are

going to be strong because we just have 12 memory channels running at 6,400 megatransfers per second click on memory threaded oh wait what is that so that's megabytes a second that's how far I have to zoom in to find the high-end available

Ram, let's go! Yeah, we're definitely winning on that one, let's go! I think maybe disc Mark we have to do too, these Kyok SSDs are nuts. Could we be a little bit limited by our low thread speed though? Oh, 100% dude. Do you remember how

many drives we used to have to put together to see numbers like this? Yeah, the crazy thing is like our storage server probably does 20 gigabytes a second, this is one drive. All those drives, but this is one drive. Let's have a look at

our compare here, so this is last gen, this is the one we just tested. That's a lot of generation over-generation uplift. Yeah, multi-threaded basically 30% on the dot. From what I've read, like other applications can be 45% or more. Well, that's what a huge

increase in clock speed and core count is going to do for you. Zen 5 might have been a bit of a wet fart on the desktop, but on the Enterprise side it is anything but. This is the result did last night, which is finally posted. Okay, 189,000 and this is the master list of

all dual CPU submissions, which naturally a lot of AMD in there. Dude, we were so close to the top of the chart. We're very close, it's so close, kind of infuriating actually. Okay, what is this though? That's the 64 core frequency optimized one,

that's only 64 cores per CPU, okay, but up to 5 GHz. The first last gen Epic setup is all the way down in fourth place, flipping 35,000 points behind us, and those are the 3D V-cache ones, which are very expensive. So that's our last gen

competition right so we're 60,000 points above that absolutely huge generational uplift 30% as cool as it is to see this mini course to pack this many in AMD had to design a smaller denser version of their Zen 5 Core called Zen 5c

presumably standing for compact because they actually squished the same functional units into a smaller space so unlike efficiency cores you might see in other CPUs they use the same architecture and actually a more advanced process node compared to the full-size cores which pays off in a big

way in power efficiency but these compact cores do have drawbacks AMD actually has a handy little visual to make this easier to understand smaller size but the squishiness costs you in cache and it limits your maximum clock speed so for our many gamers project we're

actually more interested in checking out the Epic 9755 sure it's got 64 fewer cores but they're the bigger faster cashier ones which is pretty important when you're running that many instances of cod or whatever although with that said Jake the 9575 F sure only 64 cores but if we

had two of them running at that kind of frequency um it may be pretty interesting depending on how many GPUs we can get working anyway all of this conversation leads us to a really important point with the diverse use cases that you see in the data center space there's no

longer a one-size fits all chip or even core design that can be the master of every workload anyway let's change gears and fire up some more appropriate workloads in Linux we're in Ubuntu the latest one 24.10 so very new kernel this

is one of those tests that I have a lot of context for that's actually not that good of a score. Dang it, 60 seconds. But here, watch this. Let's set it to only run on our first CPU, so we need to just copy that's the cores, those are the extra

threads, copy those. And Linux is so easy to do this. Uh, the only tricky thing is we have to tell Y-Cruncher to be smart about it too. You can see it's only on NUMA node zero. Yep, only using memory on NUMA node zero, which is important.

and only using 384 threads, which is one CPU. 47 seconds. What were we at? 60 before? So we picked up like 13 seconds by turning off one of our entire CPU sockets. Even sillier, let's do just the cores, no threads. Okay.

Damn, 42.3 seconds. So what we're trying to say is more is not necessarily better, especially on an application like this, which is super memory-bound. You do not want to be crossing CPUs unless you absolutely have to. Okay. This is last gen, yeah. Berg, so we go from 60 seconds.

Actually, see they tried it with multi-threading and without. So we went from 60 seconds, which was our 2P time, down to 42. Let's do something practical. Well, it might not be practical. It is like a relevant use case. This is Proxmox.

the hypervisor, and we're going to make some virtual machines. Well, that seems very practical. Yeah, just not in the way that I'm going to do it. We have all these cores, right? Well, I guess we're going to need a few VMs. You see what

that's doing oh you're just copying your VMs I'm a little worried it's going to get upset but I'm how many are you doing like a few that's not actually that silly I mean massive parallelization of light compute workloads that that's

actually right in this thing's wheelhouse yeah but it's like Windows 11 desktop and and I'm talking like 192 okay yeah no normally it would be like AWS instances or something like like little Linux instances yeah yeah I'm talking like running web services full desktop windows okay that's really

stupid but will it work we have 384 cores that gives us two cores of RAM that's not bad it's not great I put the one image on a RAID 0 of two of those Kyocera SSDs and they're linked images so if you look at the actual storage usage

it should be literally exactly the same as one that's pretty cool usually when you bulk start VMs and Proxmox it does them one at a time but that takes too long so I made a script to just do them all at once how's our CPU usage oh yeah

it's 15 15% look at the RAM usage climb oh my goodness I think I gave each of them eight gigs or like seven and a half gigs oh you're running a YouTube video so we made a little script that when the Windows instance starts it opens a

random YouTube video on a list of like 250 that I popped in there and we run almost 200 100 instances of Windows and play back almost 200 videos with software decoding that is flipping wild open up task manager on one of these let's see if it's actually

struggling cuz it might very well not be sure they're not clocked that high compared to some of the other chips, but there's N5 for crying out loud. Like, once it's running, oh, once it's running it's not doing that much, so I've only given

it two real cores. Allowed them to use two real cores and then it has four in here, so it's chilling out. Remember, it's still like booting up, it's like indexing Windows, it's it's probably checking for Windows updates. If these all start Windows updates, oh my

God, okay, so that's that's not bad. No way, dude, I think it worked. Are they all on? No way. This cuz 192 instances of Linux, I don't know if it would be that impressive. I mean, this will be impressive, but this is like Windows 11 Home Edition, it's not

stripped down at all. I didn't disable anything. I literally installed Windows, booted it up, set the script to load, it's not optimized in any way, shape, or form. How's our usage? That's not bad, dude. Half of our CPU, yeah, and of course, the IO

delay 0%. Shout out Kokia, when you raid zero 2 SSDs that can do 15 gigabytes a second each, pretty fast. Bottom line, then, it's pretty clear that this new lineup of Epic chips is going to pretty much dominate everything else on the market.

Like, I've kind of lost track of the last time any of their competitors had a fighting chance in the data center, and I think AMD knows it. While core for core pricing hasn't changed much in the last couple of generations, let me say that another way.

core for core pricing hasn't changed much so this 192 core chip will set you back nearly $155,000 at MSRP that means that this one computer on its own has $30,000 worth of CPUs for the right customer where ongoing expenses like data center

rack space, cooling and power are the bigger concerns. This could be a very compelling value, so props to AMD, props to Super Micro who sponsored this video. You can check out their hyper servers and the rest of their H14 Epic 905 server lineup by checking out the

showcases that we did on the Super Micro channel, which we're going to have linked in the video description. They have some wild stuff like their new Flex Twins, which stuff eight up to 192 CPUs in a 2U, like this size guys, that's

1,532 cores, or a casual 37 dish, thousand in a 48U. Mindblowing stuff. If you guys enjoyed this video, why not check out the one we made about our Super Micro Grand Twin and some sick AI video processing tech that was super

cool, and I'm kind of wondering are we ever going to deploy that? Yes, nice.

The fastest CPU in the world

Linus Tech Tips is a passionate team of "professionally curious" experts in consumer technology and video production who aim to educate and entertain.


Linus Tech Tips

Discusses the increasing core count of AMD EPYC CPUs, leading to the 192-core EPYC 9965.

Highlights the specifications of the EPYC 9965: 192 cores, 384 threads, large memory capacity, and high PCI lanes.

Emphasizes the CPU's impressive performance and power consumption (500W).

Comments on the power efficiency considering the core count.

Questions how AMD achieved such high core count and performance.

Explains the architectural improvements in Zen 5, including next-gen branch prediction and increased cache bandwidth.

Summarizes the performance uplift claimed by AMD for Zen 5 compared to Zen 4.

Notes the requirements for using the EPYC 9965, including compatible BIOS and cooling.

Mentions the Supermicro 2U 2126 HSN hyperserver used for testing.

Describes the installation process and the compatibility of the CPU with previous generation sockets.

Highlights the importance of a compatible motherboard capable of handling the high power consumption.

Shows the use of thermal pads for CPU installation.

Explains the importance of proper tightening for CPU contact with pins.

Details the system's memory configuration and cost considerations.

Mentions the use of high-speed 6400 MHz memory.

Discusses the PCIe expansion capabilities of the system.

Describes the versatility of the chassis for various workloads.

Details the installation of NVMe SSDs and the chassis's storage capacity.

Highlights the modular design of the chassis and its flexibility for high-performance storage.

Explains the potential storage configuration and read performance.

Discusses the cooling system's efficiency and modularity.

Summarizes the chassis's design philosophy of flexibility and modularity.

Describes the system boot process and power supply configuration.

Shows the system booting successfully and initial CPU usage observations.

Explains the GPU installation process and power supply considerations.

Demonstrates the system's low CPU utilization even under load.

Attempts to run Cinebench R23, encountering limitations due to the high core count.

Runs Blender for CPU rendering, achieving a fast render time.

Tests 7-Zip compression, again encountering thread limitations.

Attempts to play a game, highlighting its unsuitability for gaming.

Considers running multiple game instances.

Tests Doom Eternal, showing surprisingly good performance despite low CPU usage.

Runs PassMark benchmark for comparison to other CPUs.

Runs MemoryMark and DiskMark benchmarks, showcasing high memory and storage performance.

Compares the results to last generation CPUs, showing significant improvement.

Analysis of Benchmark Results and Zen 5c Cores

Analyzes the benchmark results, showing high performance compared to previous generation CPUs.

Discusses the Zen 5c core design and its trade-offs in terms of cache and clock speed.

Considers alternative CPU configurations for specific workloads.

Highlights the importance of choosing the right CPU for diverse data center workloads.

Runs Y-Cruncher benchmark in Linux, showing performance improvement by limiting usage to one CPU.

Further improves performance by disabling threads, highlighting the importance of memory bandwidth.

Compares performance with and without multi-threading.

Tests Proxmox virtualization with numerous virtual machines.

Runs multiple Windows 11 VMs simultaneously, demonstrating the CPU's capability for parallelization.

Shows the system running numerous VMs with low CPU utilization.

Summarizes the findings, concluding that the new EPYC chips are dominant in the market.

Discusses the pricing of the 192-core chip and its value proposition for specific customers.

Promotes Super Micro's hyper servers and other products.

Mentions other Super Micro products and related videos.