WEBVTT

00:00:00.080 --> 00:00:08.519
as Gamers we're running machines with four eight maybe 16 cores if we're lucky

00:00:05.440 --> 00:00:12.240
so when AMD launched their first 96 core

00:00:08.519 --> 00:00:13.719
epic CPU I figured well we're done here

00:00:12.240 --> 00:00:18.760
that ought to be enough chip for even the most dedicated Potato Head but the

00:00:16.160 --> 00:00:24.920
Spud Masters at AMD said no and last year launched a

00:00:20.920 --> 00:00:28.160
128 cor CPU surely that must be enough

00:00:24.920 --> 00:00:30.560
right no it wasn't and thanks to Super

00:00:28.160 --> 00:00:37.440
Micro who sponsored this video I have the brand new AMD epic 9965 a CPU with

00:00:33.800 --> 00:00:41.280
not 96 not 128 but

00:00:37.440 --> 00:00:44.480
192 cores making it alongside its Turin

00:00:41.280 --> 00:00:48.239
family members the fastest CPU on the

00:00:44.480 --> 00:00:52.280
planet by a wide margin we're talking

00:00:48.239 --> 00:00:56.600
384 threads terabytes of ddr5 ECC memory

00:00:52.280 --> 00:01:01.920
128 Gen 5 PCI lanes and a 500 wat TDP I

00:00:56.600 --> 00:01:01.920
can only think of one word what holy sh

00:01:09.640 --> 00:01:17.000
I never thought I'd find myself saying only 500 watts of power consumption on a

00:01:14.640 --> 00:01:22.680
CPU but when you consider that AMD packed 192 cores into the same space

00:01:20.520 --> 00:01:27.400
they used for half as many just two years ago you got to agree it's pretty

00:01:25.200 --> 00:01:33.320
impressive I mean sure that's more than my entire first gaming PC but but

00:01:30.439 --> 00:01:39.040
compared to the king epic at that time this new chip has double the core count

00:01:36.200 --> 00:01:44.280
and the same maximum boost clock for real world performance that utterly

00:01:41.079 --> 00:01:46.880
obliterates it so how did they do it

00:01:44.280 --> 00:01:52.119
well for starters just like AMD's latest ryzen 9000 desktop processors this

00:01:49.360 --> 00:01:55.600
monster is powered by Zen 5 which comes with a bunch of architectural

00:01:53.560 --> 00:01:59.119
improvements like NextGen Branch prediction which allows these chips to

00:01:57.320 --> 00:02:03.360
predict their next move two steps in advance rather than one not to mention

00:02:01.240 --> 00:02:08.360
with greater accuracy and lower latency all while increasing cash bandwidth

00:02:05.759 --> 00:02:13.840
implementing a wider execution path and adding a full 512-bit data path for AVX

00:02:11.360 --> 00:02:20.560
512 workloads if all that sounds like gobbley that's fine it boils down to a

00:02:16.640 --> 00:02:23.480
claimed 16% average IPC uplift from zen4

00:02:20.560 --> 00:02:27.360
to Zen 5 in PC workloads and potentially even more for high performance compute

00:02:25.840 --> 00:02:33.280
and the Enterprise workloads that typically run on servers let's get it

00:02:29.959 --> 00:02:36.280
installed but wait while 500 W seems

00:02:33.280 --> 00:02:38.400
downright reasonable for 192 cores it

00:02:36.280 --> 00:02:43.480
does mean you can't just plop them into any old chassis and expect it to work

00:02:40.840 --> 00:02:48.319
you need a compatible BIOS sufficient Cooling and sufficient power so keep

00:02:46.159 --> 00:02:52.519
that in mind but it isn't going to be a problem for us today since super micro

00:02:50.280 --> 00:02:59.159
can build practically anything including this as 2126 HSN hypers server that they

00:02:57.120 --> 00:03:05.560
provided for this video This Magnificent machine can handle not one but two of

00:03:02.080 --> 00:03:11.640
these monsters which is 1,000 watts of

00:03:05.560 --> 00:03:13.159
just CPU that's 384 cores 768 threads

00:03:11.640 --> 00:03:17.680
now typically if you were to buy one from Super Micro it's going to come

00:03:15.159 --> 00:03:22.080
preassembled but we went out of our way to take the CPU out so I can put it back

00:03:19.760 --> 00:03:28.080
in and show you how these new Chips use the same socket as last generation so if

00:03:24.560 --> 00:03:29.519
you have one you in theory could use it

00:03:28.080 --> 00:03:33.560
but you are going to want something that can handle more power not every

00:03:31.519 --> 00:03:37.519
motherboard is going to be rid to handle a 500 W CPU right out of the box you

00:03:35.920 --> 00:03:43.599
should check your motherboard manual this one purpose buil for this platform

00:03:40.200 --> 00:03:47.159
so we should have no problems at all M

00:03:43.599 --> 00:03:49.239
PTM 7950 LTT Store.com it's a thermal

00:03:47.159 --> 00:03:54.239
pad and actually perfect for this application because it's super durable

00:03:51.760 --> 00:03:58.159
these things can handle like hundreds or thousands of heat Cycles while barely

00:03:56.079 --> 00:04:01.680
even losing any of their original performance which is awesome on a CPU

00:04:00.439 --> 00:04:08.280
this big it's important to make sure everything is tightened down properly so that the CPU actually makes contact with

00:04:05.640 --> 00:04:13.720
all of those thousands of pins because unlike most desktop CPUs these days the

00:04:11.280 --> 00:04:18.079
CPU is actually held in by the force of the cooler not by like a little bracket

00:04:16.359 --> 00:04:20.840
on top I mean there is a bracket but that mostly just kind of holds it in

00:04:19.680 --> 00:04:26.280
place while you're tightening everything down order 1 2 3 4 5 6 fully tightened

00:04:24.360 --> 00:04:30.160
one at a time yeah not a little bit at a time not a little bit at a time which is

00:04:28.479 --> 00:04:33.960
the conventional wisdom you know no I tried to do that on one of the CPUs and

00:04:31.919 --> 00:04:39.160
it it actually didn't post look how scary that is gorgeous now while we

00:04:37.479 --> 00:04:44.680
could have spec this machine with up to 9 terabytes of 6,000 megat transfer per

00:04:42.199 --> 00:04:52.360
second registered ECC memory across its 24 dim slots that would have cost um

00:04:49.320 --> 00:04:55.000
probably as much as my car I think maybe

00:04:52.360 --> 00:04:59.199
more than both of my cars so I think the 1 and 1 12 terabytes that super micro

00:04:56.840 --> 00:05:03.880
sent over is going to be plenty for us this is 6400 I heard that AMD is doing

00:05:01.800 --> 00:05:08.759
like special models with certain providers that support 6400 that's

00:05:06.880 --> 00:05:14.080
incredible I guess this is one of them that's a big Improvement hold on that

00:05:10.560 --> 00:05:16.720
means standard server memory is faster

00:05:14.080 --> 00:05:24.840
than standard desktop memory what hasn't changed from last gen is PCIe expansion

00:05:19.560 --> 00:05:26.600
we still get 128 or 160 Gen 5 Lanes in

00:05:24.840 --> 00:05:30.720
single and dual socket configs respectively and this particular chassis

00:05:28.639 --> 00:05:35.720
can be speced for either eight eight Lane expansion slots or four 16 Lane

00:05:33.880 --> 00:05:41.600
expansion slots if we have dual width cards they also crammed another 16 Lanes

00:05:38.360 --> 00:05:44.319
into an ocp 3.0 mezan slot here all of

00:05:41.600 --> 00:05:49.479
which well most of which we aren't using for now but you definitely could while

00:05:47.080 --> 00:05:53.400
our hypers server is fairly well-rounded a solid setup for virtualization which

00:05:51.360 --> 00:05:57.520
we'll show you later or Cloud workloads where you need a lot of cores and not a

00:05:55.560 --> 00:06:01.600
ton else aside from memory this chassis is also perfect for high performance

00:05:59.280 --> 00:06:05.280
computer cute and even AI I mean you've already got the maximum amount of CPU so

00:06:03.960 --> 00:06:09.120
with the rear slot stuffed with high-speed networking gpus accelerator

00:06:07.280 --> 00:06:14.000
cards maybe even all three at the same time there's not a lot that a machine

00:06:11.080 --> 00:06:18.800
like this can't do don't wait today is the last day of our Black Friday and

00:06:16.039 --> 00:06:22.479
cyber monday deals at LT store grab some of the best deals of the year on

00:06:20.400 --> 00:06:27.680
screwdrivers magnetic cable management and more at lg. gblack Friday we're

00:06:26.080 --> 00:06:33.080
going to have it linked down below what we will be installing now is a few of

00:06:29.240 --> 00:06:37.160
these 15 TB kokia NVMe ssds in the eight

00:06:33.080 --> 00:06:38.599
slots that each support PCIe gen 5x4

00:06:37.160 --> 00:06:42.639
funny thing about this chassis by the way if you're building out high

00:06:40.560 --> 00:06:46.479
performance storage for AI or I don't know something else that needs fast

00:06:43.840 --> 00:06:51.840
storage maybe with software like WKA it can be configured with up to 24 2 and 1

00:06:49.639 --> 00:06:56.240
12 in Drive Bays thanks to the modular design of the front so you could install

00:06:54.000 --> 00:07:03.240
an eight drive back plane like this one in each section with a bunch of 30 TB

00:06:59.759 --> 00:07:06.560
CM7 and you could have over 700 terab of

00:07:03.240 --> 00:07:08.599
Gen 5 storage in a single 2ou machine

00:07:06.560 --> 00:07:16.479
that's a theoretical read performance of 240 gbt per second across those drives

00:07:12.120 --> 00:07:19.000
in a single chassis or more than 330 GB

00:07:16.479 --> 00:07:23.000
per second if you went with 15s you know we get excited by fast ssds in our

00:07:21.039 --> 00:07:27.240
gaming systems but when you stack them up together in a stupid fat server like

00:07:25.080 --> 00:07:32.479
this it's kind of mind-blowing speaking of blowing if you're speacking 500 w

00:07:29.680 --> 00:07:36.960
CPUs like we've got or gpus for AI or rendering you uh might actually

00:07:35.160 --> 00:07:40.560
appreciate the modularity of this chassis and the fact that it doesn't

00:07:38.919 --> 00:07:46.080
necessarily have all those Drive bays in there because these are some flipping

00:07:43.120 --> 00:07:49.360
wild counter rotating fans but they're still going to pull air better if they

00:07:47.759 --> 00:07:52.879
don't have a whole bunch of restriction in front of them so the bottom line is

00:07:51.319 --> 00:07:56.840
whatever your needs are super micro's hypers systems and the entire rest of

00:07:54.879 --> 00:08:01.599
the lineup are designed with flexibility and modularity in mind you ready to put

00:07:59.879 --> 00:08:08.280
the top on did you look at the power supplies yet 2,000 Watts 80 plus

00:08:05.520 --> 00:08:13.639
titanium two of them naturally what are the odds this boot's first

00:08:10.120 --> 00:08:16.440
time but those giant sockets even if you

00:08:13.639 --> 00:08:22.199
do it properly like I don't know this thing same 85% 90 you did it properly

00:08:20.520 --> 00:08:26.560
yeah but I wasn't here when you put in the other one

00:08:23.840 --> 00:08:31.319
so waa she's going to go she's going to go these things take forever to post the

00:08:29.199 --> 00:08:36.039
first time that actually makes sense think how long your desktop ryzen takes

00:08:33.000 --> 00:08:38.880
to do memory training the first time

00:08:36.039 --> 00:08:43.919
well is it up F yeah oh look oh that's great and you doubted well I didn't

00:08:41.200 --> 00:08:51.560
doubt much I said 85 90% all right can I reboot it uh-oh oh boy I think it's set

00:08:48.399 --> 00:08:55.000
to Max yeah I think so we should just

00:08:51.560 --> 00:08:56.760
leave it like that right yeah maybe not

00:08:55.000 --> 00:09:02.839
optimal speed that sounds more manageable oh thank God oh my God that's

00:08:59.560 --> 00:09:04.160
oh oh that's downright Pleasant wow now

00:09:02.839 --> 00:09:09.640
that we know it boots we should probably put a GPU in there right and what better

00:09:06.640 --> 00:09:12.160
GPU to go with the oops Top Dog AMD epic

00:09:09.640 --> 00:09:18.279
Enterprise CPU then a top dog not Enterprise but still awesome Radeon RX

00:09:14.920 --> 00:09:20.040
7900 XTX now despite this chassis being

00:09:18.279 --> 00:09:24.279
well equipped for gpus I mean the power connectors are like right there so they

00:09:21.560 --> 00:09:28.839
just plug right in and they've got a 2600 wat power supply option that you

00:09:26.440 --> 00:09:34.560
can upgrade to these slots are meant for full height fulllength workstation and

00:09:30.839 --> 00:09:37.680
Enterprise gpus not weird shaped gaming

00:09:34.560 --> 00:09:41.399
gpus so our plan is to put our GPU

00:09:37.680 --> 00:09:44.720
outside the server with a riser and a

00:09:41.399 --> 00:09:44.720
jumpered ATX power

00:09:44.920 --> 00:09:54.040
supply nice Flawless oh my God it

00:09:51.079 --> 00:09:58.399
Scrolls how much stuff can I do and still maintain 0% utilization could I

00:09:56.839 --> 00:10:03.800
launch a browser oh I'm sure you can play a YouTube video it's only 1% oh no

00:10:01.279 --> 00:10:09.000
it's it's at zero again once it settles in and it's doing its thing it's nothing

00:10:06.079 --> 00:10:15.720
it going to load One Core it's 384 cores it could load two cores and still be

00:10:12.519 --> 00:10:17.240
below 1% duplicate this tab uh well hold

00:10:15.720 --> 00:10:22.200
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

00:10:20.560 --> 00:10:28.839
yet remember this is this is happening on the GPU CPU Works a little harder it works a

00:10:26.279 --> 00:10:33.040
little harder CPU is at 10% so I think most of the this is GPU 4K here we go

00:10:31.320 --> 00:10:40.160
there's like some cores there's one core I see at 9% the rest of them are vibing thriving

00:10:37.560 --> 00:10:44.760
one might even say okay the GPU is at 77% according to task manager how is

00:10:43.120 --> 00:10:49.240
that even possible I think it's just cuz it's video decode okay let's disable

00:10:47.600 --> 00:10:53.639
Hardware acceleration I think this is going to hurt a little bit no 2% yeah

00:10:51.839 --> 00:11:00.639
it's going to hurt it's going to hurt so much our memory is 0% too it actually

00:10:57.480 --> 00:11:03.600
just says 0.0 but I guess to get to 0.1

00:11:00.639 --> 00:11:06.760
we're talking 100 gigs I need like a script to to open these over and over

00:11:05.480 --> 00:11:11.200
again I want to see how maybe we should just revisit that video on this machine

00:11:09.680 --> 00:11:16.639
how many Chrome tabs we get some more RAM we revisit how many Chrome tabs but

00:11:13.600 --> 00:11:19.200
we like automate it this time but let's

00:11:16.639 --> 00:11:25.040
move on let's move on for now uh cin bench do you have R23 so we can see all

00:11:21.320 --> 00:11:27.560
the little things uh I do but it wasn't

00:11:25.040 --> 00:11:31.079
launching oh wow it just does nothing no I know doesn't even try to I looked in

00:11:29.720 --> 00:11:36.399
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

00:11:33.880 --> 00:11:42.720
threats and 2024 is going to be a supreme disappointment what it's thread

00:11:39.120 --> 00:11:44.320
limited to 256 for some reason I don't

00:11:42.720 --> 00:11:48.880
actually know why that's it you didn't even say what the score is if that score

00:11:46.320 --> 00:11:54.000
was my child I wouldn't even speak to it oh what a flipping disappointment get

00:11:51.800 --> 00:11:58.360
out of my sight you could try blender ready CPU rendering let's go I feel like

00:11:56.959 --> 00:12:03.279
this is going to take longer just to prepare the job than to actually render

00:12:00.440 --> 00:12:09.880
it that's a lot of blue it's a blue screen of

00:12:05.040 --> 00:12:13.040
success damn 12 seconds she doesn't get

00:12:09.880 --> 00:12:15.720
much clock speed though does she yeah

00:12:13.040 --> 00:12:19.639
this is also running with smt on like we're kind of at the point where you

00:12:16.920 --> 00:12:24.800
have so many cores that I feel like a lot of applications are better off

00:12:21.600 --> 00:12:27.800
without multi- threading or some might

00:12:24.800 --> 00:12:31.120
be anyways these CPUs between the two of

00:12:27.800 --> 00:12:34.440
them have over twice as much cash as my

00:12:31.120 --> 00:12:37.519
first gaming PC had system memory like

00:12:34.440 --> 00:12:41.240
the first one I bought for myself yeah

00:12:37.519 --> 00:12:43.680
okay cash yeah and these don't even have

00:12:41.240 --> 00:12:48.480
that much cash no seven zip can't perform that great there's no way it'll

00:12:45.639 --> 00:12:54.480
go across this many threads CPU threads oh what limited again let's go

00:12:52.880 --> 00:13:04.360
compare this to Intel's new desktop chips real quick so 7 700 giips

00:13:00.680 --> 00:13:07.160
which is 700,000 nips so it's like 3 and

00:13:04.360 --> 00:13:11.079
half 9950 X's oh no no no wait wait wait that's B that's kilobytes per second

00:13:09.199 --> 00:13:14.639
yeah oh I see yeah so we're double the performance but

00:13:12.560 --> 00:13:19.880
like okay well now we should play a game on it right sure I mean we have to this

00:13:16.920 --> 00:13:27.199
is like the worst possible CPU to play video games oh yeah we're at 4K High

00:13:23.880 --> 00:13:29.440
settings oh man that was cheesy that was

00:13:27.199 --> 00:13:33.600
cheesy not going to lie okay clearly this is not a gaming

00:13:31.079 --> 00:13:41.160
CPU I mean it's working for me what can I say no you could probably run if you

00:13:37.720 --> 00:13:43.480
had enough gpus like 60 instances of

00:13:41.160 --> 00:13:48.440
Counter Strike at this point dude no though should we do it it's been so long

00:13:47.079 --> 00:13:55.839
how many years has it been since we've done many gamers one CPU this time we

00:13:51.120 --> 00:13:59.160
could actually do one CPU with 192 cores

00:13:55.839 --> 00:14:01.560
dude that would be crazy yeah we should

00:13:59.160 --> 00:14:06.240
do it I want to do it here did you know alt f4 makes the game run faster did you

00:14:03.720 --> 00:14:10.040
know that alt f4 doesn't do jack in this particular no I'm winning I'm winning

00:14:07.600 --> 00:14:13.480
I'm going to win I'm going to win Jake Jake I'm going to win okay what's next

00:14:12.320 --> 00:14:19.720
Doom maternal is a good one cuz Doom maternal is actually like fairly multi threaded supposedly here we go horde

00:14:16.959 --> 00:14:25.839
mode 4K everything Ultra nightmare motion blur off raate tracing off looks

00:14:22.120 --> 00:14:28.519
pretty good it's 200 FPS uh

00:14:25.839 --> 00:14:34.040
yeah wow it actually does use some cores ehy yeah but like I I mean relatively

00:14:31.560 --> 00:14:39.839
speaking no but that's crazy though it's actually

00:14:37.240 --> 00:14:45.839
using like 16 cors I feel like but we're running at 200 FPS that means we could

00:14:43.639 --> 00:14:52.279
run multiple instances of Doom maternal on this thing wasn't the CPU usage at 1%

00:14:49.360 --> 00:14:57.880
yeah when I say multiple I mean a big multiple all right this actually feels

00:14:55.279 --> 00:15:02.199
great yeah yeah click things he's almost dead kill him glory kill use more fuel

00:15:00.000 --> 00:15:06.000
buddy oh my God you're going to die J oh no oh no oh no leave me alone he called

00:15:03.720 --> 00:15:12.360
you an arrogant fool quick Glory kill I did it yeah again yeah let's go yeah

00:15:09.759 --> 00:15:15.839
this is great it's running lovely okay let's do something not stupid on it pass

00:15:14.600 --> 00:15:20.800
mark would be a good one cuz it'll show us like comparison to CPUs that are

00:15:18.519 --> 00:15:26.240
relevant I bet you in a lot of cases the numbers that we get running across both

00:15:23.959 --> 00:15:32.639
CPUs will be the same as running across one or maybe even worse and that's just

00:15:29.639 --> 00:15:35.199
software limitations but you would never

00:15:32.639 --> 00:15:40.360
configure a system like this for a single workstation that's running one t

00:15:38.120 --> 00:15:44.120
a single workload yeah or at least if you threaded it across all of them you

00:15:42.519 --> 00:15:47.680
would have them be well aware of the fact that there's two CPUs and the

00:15:46.000 --> 00:15:52.240
performance degradation you get when you try to cross them yeah and we we're

00:15:49.839 --> 00:15:55.720
literally off the chart yeah like a lot so nobody else has submitted the CPU yet

00:15:54.360 --> 00:15:59.880
so let's run memory Mark and then we can submit it sure and then we can see what

00:15:57.759 --> 00:16:04.199
it's like compared to other CP all right one of the tests is

00:16:01.399 --> 00:16:09.440
just how much memory you have I think we're going to do okay on that one yeah

00:16:06.240 --> 00:16:13.399
maybe okay oh

00:16:09.440 --> 00:16:15.399
63rd mid AF the read and the right are

00:16:13.399 --> 00:16:19.560
going to be strong because we just have 12 memory channels running at 6,400

00:16:18.040 --> 00:16:22.160
megat transfers per second click on memory

00:16:23.000 --> 00:16:30.160
threaded oh wait what is that so that's megabytes a second that's how far I have

00:16:28.000 --> 00:16:33.880
to zoom in to find the highend available RAM let's go yeah we're definitely

00:16:31.800 --> 00:16:39.680
winning on that one let's go I think maybe disc Mark we have to do too these

00:16:36.120 --> 00:16:41.600
kyok ssds are nuts could we be a little

00:16:39.680 --> 00:16:45.120
bit limited by our low thread speed though oh 100% dude do you remember how

00:16:43.519 --> 00:16:48.959
many drives we used to have to put together to see numbers like this yeah

00:16:47.360 --> 00:16:54.040
the crazy thing is like our storage server probably does 20 gigaby a second

00:16:51.800 --> 00:16:57.519
this is one drive all those drives but this is one drive let's have a look at

00:16:55.880 --> 00:17:04.679
our compare here so this is last gen this is the one

00:17:02.319 --> 00:17:10.919
we just tested that's a lot of generation overgeneration uplift yeah

00:17:07.839 --> 00:17:12.160
multi-threaded basically 30% on the Dot

00:17:10.919 --> 00:17:17.480
from what I've read like other applications can be

00:17:14.600 --> 00:17:22.520
45% or more well that's what a huge increase in clock speed and core count

00:17:20.400 --> 00:17:26.880
is going to do for you Zen 5 might have been a bit of a wet fart on the desktop

00:17:24.559 --> 00:17:31.200
but on the Enterprise side it is anything but this is the result did last

00:17:29.400 --> 00:17:36.559
night which is finally posted okay 189,000 and this is the master list of

00:17:33.400 --> 00:17:39.720
all dual CPU submissions which naturally

00:17:36.559 --> 00:17:42.360
a lot of AMD in there dude we were so

00:17:39.720 --> 00:17:48.240
close to the top of the chart we're very close it's so close kind of infuriating

00:17:45.080 --> 00:17:50.840
actually okay what is this though that's

00:17:48.240 --> 00:17:59.120
the 64 core frequency optimized one that's only 64 cores per C okay but up

00:17:54.440 --> 00:18:01.240
to 5 GHz the first last gen epic setup

00:17:59.120 --> 00:18:08.960
is all the way down in fourth place flipping 35,000 points behind us and

00:18:04.880 --> 00:18:12.159
those are the 3dv cash ones which are

00:18:08.960 --> 00:18:15.880
very expensive so that's our last gen

00:18:12.159 --> 00:18:18.520
competition right so we're 60,000 points

00:18:15.880 --> 00:18:24.799
above that absolutely huge generational uplift 30% as cool as it is to see this

00:18:21.600 --> 00:18:27.320
mini course to pack this many in AMD had

00:18:24.799 --> 00:18:32.559
to design a smaller denser version of their Zen 5 Core called Zen 5c

00:18:30.280 --> 00:18:37.200
presumably standing for compact because they actually squished the same

00:18:34.200 --> 00:18:39.640
functional units into a smaller space so

00:18:37.200 --> 00:18:44.039
unlike efficiency cores you might see in other CPUs they use the same

00:18:41.559 --> 00:18:48.640
architecture and actually a more advanced process node compared to the

00:18:45.880 --> 00:18:54.679
full-size cores which pays off in a big way in power efficiency but these

00:18:51.919 --> 00:18:58.400
compact cores do have drawbacks AMD actually has a handy little visual to

00:18:56.240 --> 00:19:01.640
make this easier to understand smaller size

00:18:59.280 --> 00:19:06.600
but the squishiness costs you in cash and it limits your maximum clock speed

00:19:04.520 --> 00:19:09.480
so for our many gamers project we're actually more interested in checking out

00:19:08.360 --> 00:19:15.200
the Epic 9755 sure it's got 64 fewer course but

00:19:12.520 --> 00:19:18.559
they're the bigger faster cashier ones which is pretty important when you're

00:19:16.240 --> 00:19:25.120
running that many instances of cod or whatever although with that said Jake

00:19:20.720 --> 00:19:26.600
the 9575 F sure only 64 cores but if we

00:19:25.120 --> 00:19:30.240
had two of them running at that kind of frequency

00:19:28.919 --> 00:19:36.720
um it may be pretty interesting depending on how many gpus we can get working anyway all of this conversation

00:19:34.799 --> 00:19:40.480
leads us to a really important point with the diverse use cases that you see

00:19:38.679 --> 00:19:45.919
in the data center space there's no longer a one-size fits all chip or even

00:19:44.000 --> 00:19:49.880
core design that can be the master of every workload anyway let's change gears

00:19:48.360 --> 00:19:55.200
and fire up some more appropriate workloads in Linux we're in Ubuntu the

00:19:52.159 --> 00:19:56.840
latest one 2410 so very new kernel this

00:19:55.200 --> 00:20:02.679
is one of those tests that I have a lot of context for that's actually not that

00:19:59.559 --> 00:20:05.080
good of a score dang it 60 seconds but

00:20:02.679 --> 00:20:10.960
here watch this let's set it to only run on our first CPU so we need to just copy

00:20:08.880 --> 00:20:15.840
that's the cores those are the extra threads copy those and Linux is so easy

00:20:13.880 --> 00:20:21.320
to do this uh the only tricky thing is we have to tell y cruncher to be smart

00:20:18.280 --> 00:20:24.840
about it too you can see it's only on

00:20:21.320 --> 00:20:29.000
num Noe zero yep only using memory on

00:20:24.840 --> 00:20:31.400
num Noe zero which is important

00:20:29.000 --> 00:20:37.760
and only using 384 threads which is 1 CPU 47 seconds what were we at 60 before

00:20:35.120 --> 00:20:42.799
so we picked up like 13 seconds by turning off one of our entire CPU

00:20:40.400 --> 00:20:49.559
sockets even sillier let's do just the cores no threads okay

00:20:46.120 --> 00:20:53.200
damn 42.3 seconds so what we're trying

00:20:49.559 --> 00:20:55.320
to say is more not necessarily better

00:20:53.200 --> 00:21:01.240
especially on an application like this which is super memory bound you do not

00:20:58.679 --> 00:21:10.799
want to be Crossing CPUs unless you absolutely have to okay this is last gen

00:21:05.039 --> 00:21:12.159
yeah bergo so we go from 60 seconds

00:21:10.799 --> 00:21:16.440
actually see they tried it with multi-threading and without so we went

00:21:13.799 --> 00:21:20.600
from 60 seconds which was our 2p time down to 42 let's do something practical

00:21:18.799 --> 00:21:24.880
well it might not be practical it is like a relevant use case this is proxmox

00:21:23.440 --> 00:21:28.919
the hypervisor and we're going to make some virtual machines well that seems

00:21:26.840 --> 00:21:32.880
very practical yeah just not in the way that I'm going to do it we have all

00:21:30.960 --> 00:21:40.000
these cores right well I guess we're going to need a few VMS you see what

00:21:35.320 --> 00:21:41.640
that's doing oh you're just copy pting

00:21:40.000 --> 00:21:47.600
your VMS I'm a little worried it's going to get upset but I'm how many are you

00:21:43.919 --> 00:21:50.159
doing like few that's not actually that

00:21:47.600 --> 00:21:53.279
silly I mean massive parallelization of light compute workloads that that's

00:21:51.799 --> 00:21:57.520
actually right in this thing wheelhouse yeah but it's like Windows 11 desktop

00:21:55.400 --> 00:22:02.159
and and I'm talking like 192 okay yeah no normally it would be

00:21:59.799 --> 00:22:05.720
like AWS instances or something like like little Linux instances yeah yeah

00:22:04.039 --> 00:22:11.559
I'm talking like running web services full desktop window okay that's really

00:22:07.600 --> 00:22:14.240
stupid but will it work we have 384

00:22:11.559 --> 00:22:19.039
cores that gives us two cores of em that's not bad it's not great I put the

00:22:17.120 --> 00:22:23.520
one image on a raid zero of two of those kyok ssds and they're linked images so

00:22:21.360 --> 00:22:28.279
if you look at the actual storage usage it should be literally exactly the same

00:22:26.080 --> 00:22:32.200
as one that's pretty cool usually when you bulk start vmss and proxmox it does

00:22:30.279 --> 00:22:36.720
them one at a time but that takes too long so I made a script to just do them

00:22:33.799 --> 00:22:42.279
all at once how's our CPU usage oh yeah it's 15 15% look at the RAM usage climb

00:22:40.480 --> 00:22:46.080
oh my goodness I think I gave each of them eight gigs or like seven and a half

00:22:44.039 --> 00:22:50.799
gigs oh you're running a YouTube video so we

00:22:49.440 --> 00:22:54.400
made a little script that when the Windows instance starts it opens a

00:22:52.640 --> 00:23:00.400
random YouTube video on a list of like 250 that I popped in there and we run

00:22:57.559 --> 00:23:06.480
almost 200 100 instances of Windows and play back almost 200 videos with

00:23:04.240 --> 00:23:10.840
software decoding that is flipping Wild open up task manager on one of

00:23:08.799 --> 00:23:15.480
these let's see if it's actually struggling cuz it might very well not be

00:23:14.000 --> 00:23:20.400
sure they're not clocked that high compared to some of the other chips but

00:23:18.200 --> 00:23:25.880
there's N5 for crying out loud like once it's running oh once it's running it's

00:23:22.640 --> 00:23:29.440
not doing that much so I've only given

00:23:25.880 --> 00:23:31.000
it two real cores allowed them to use

00:23:29.440 --> 00:23:35.200
two real cores and then it has four in here so it's chilling out remember it's

00:23:33.240 --> 00:23:38.919
still like booting up it's like indexing Windows it's it's probably checking for

00:23:37.559 --> 00:23:45.120
Windows updates if these all start Windows updates oh my

00:23:40.799 --> 00:23:48.080
God okay so that's that's not bad no way

00:23:45.120 --> 00:23:52.960
dude I think it worked are they all on no

00:23:49.000 --> 00:23:54.200
way this cuz 192 instances of Linux I

00:23:52.960 --> 00:23:58.799
don't know if it would be that impress I mean this will be impressive but this is

00:23:56.159 --> 00:24:03.360
like Windows 11 home Edition it's not stripped down at all I didn't disable

00:24:00.840 --> 00:24:07.919
anything I literally installed Windows booted it up set the script to load it's

00:24:06.000 --> 00:24:13.760
not optimized in any way shape or form how's our usage that's not bad dude half

00:24:10.640 --> 00:24:17.400
of our CPU yeah and of course the io

00:24:13.760 --> 00:24:20.120
delay 0% shout out kokia when you raid

00:24:17.400 --> 00:24:25.279
zero2 ssds that can do 15 gab a second each pretty fast bottom line then it's

00:24:23.679 --> 00:24:29.640
pretty clear that this new lineup of Epic chips is going to pretty much

00:24:27.159 --> 00:24:33.559
dominate everything else on the market like I've kind of lost track of the last

00:24:31.840 --> 00:24:38.080
time any of their competitors had a Fighting Chance in the data center and I

00:24:35.480 --> 00:24:41.000
think AMD knows it while core for core pricing hasn't changed much in the last

00:24:40.120 --> 00:24:46.080
couple of generations let me say that another way

00:24:43.679 --> 00:24:51.000
core for core pricing hasn't changed much so this 192 core chip will set you

00:24:49.640 --> 00:24:57.840
back nearly $155,000 at MSRP that means that this

00:24:54.360 --> 00:25:00.279
one computer on its own has $30,000

00:24:57.840 --> 00:25:04.440
worth of CPUs for the right customer where ongoing expenses like data center

00:25:02.360 --> 00:25:09.520
rack space Cooling and power are the bigger concerns this could be a very

00:25:06.640 --> 00:25:13.600
compelling value so Props to AMD props to Super Micro who sponsored this video

00:25:11.559 --> 00:25:18.000
you can check out their hyper servers and the rest of their h14 epic 905

00:25:16.399 --> 00:25:21.240
server lineup by checking out the showcases that we did on the super micro

00:25:19.960 --> 00:25:25.399
Channel which we're going to have Linked In the video description they have some

00:25:23.159 --> 00:25:32.200
wild stuff like their new flex twins which stuff eight up to 192 C CPUs in a

00:25:29.200 --> 00:25:36.520
2u like this size guys that's

00:25:32.200 --> 00:25:41.080
1,532 cores or a casual 37 dish,

00:25:36.520 --> 00:25:42.600
thousand in a 48 U mindblowing stuff if

00:25:41.080 --> 00:25:47.120
you guys enjoyed this video why not check out the one we made about our

00:25:44.279 --> 00:25:50.200
super micro Grand twin and some sick AI video processing Tech that was super

00:25:49.120 --> 00:25:54.919
cool and I'm kind of wondering are we ever going to deploy that yes nice