WEBVTT

00:00:00.080 --> 00:00:06.759
Intel's chicken is plucked their net income is down

00:00:03.799 --> 00:00:10.559
114% their margins have shrunk by half and their manufacturing capabilities are

00:00:08.800 --> 00:00:16.440
so out of date that they're paying their competitor tsmc to produce chips for

00:00:13.360 --> 00:00:18.840
them and adding insult to injury their

00:00:16.440 --> 00:00:26.199
new Sapphire Rapids lineup of server CPUs has finally limped to the Finish

00:00:21.640 --> 00:00:28.840
Line two months after AMD's incredibly

00:00:26.199 --> 00:00:34.200
powerful not to mention cheaper epic Genoa counter punch with poor pricing

00:00:31.960 --> 00:00:38.960
and customer hostile Hardware as a service features it's starting to feel

00:00:36.920 --> 00:00:43.360
like we're watching a Goliath Mega Corporation crumbling in slow motion

00:00:41.480 --> 00:00:48.719
under the weight of its own hubris so why then is pulseway

00:00:46.239 --> 00:00:53.239
sponsoring us to tell you about a crappy product that you probably shouldn't buy

00:00:51.039 --> 00:00:57.840
it's because they know that with a new server and workstation processor come

00:00:55.680 --> 00:01:02.480
many of the same old problems and the rmm tools in both pulseway mobile and

00:01:00.440 --> 00:01:06.960
web apps are more than serviceable Solutions and despite the fact that

00:01:04.600 --> 00:01:12.000
these chips are as outdated as they are brand new Intel has their sight set

00:01:09.640 --> 00:01:17.080
clearly on the future an accelerated future and as bad as things look at the

00:01:14.159 --> 00:01:22.640
moment that future might be pretty bright good gravy this thing is thick

00:01:20.360 --> 00:01:26.720
like it's not bigger than last gen but look at the size of it like you could

00:01:24.600 --> 00:01:31.720
hurt a man with it maybe even kill him if you get the angle just right you know

00:01:29.479 --> 00:01:36.240
Sapphire fire Rapids the code name for Intel's fourth generation of Zeon

00:01:33.799 --> 00:01:40.040
scalable CPUs is based on the same golden Cove core architecture that can

00:01:38.119 --> 00:01:47.439
be found in the performance cores of their 12th gen desktop processors Yes

00:01:43.799 --> 00:01:50.439
you heard me last year's consumer chips

00:01:47.439 --> 00:01:53.119
that's a yikes but it does still mean a

00:01:50.439 --> 00:01:59.320
15% IPC uplift over their previous generation Zeon so that's good and

00:01:56.600 --> 00:02:04.079
there's a lot more that's different AMD was the first to bring chiplet CPUs to

00:02:01.719 --> 00:02:09.360
the world processors that use multiple dieses with a high-speed interconnector

00:02:06.439 --> 00:02:15.360
glue as Intel derisively put it at the time how the turntables have turn taed

00:02:13.160 --> 00:02:19.920
this generation xeons then will glue together up to four chiplets to reach

00:02:17.360 --> 00:02:25.720
core counts as high as 60 but the devil's in the details and Intel's glue

00:02:22.720 --> 00:02:28.000
differs from amds in some key ways while

00:02:25.720 --> 00:02:33.200
AMD sticks their compute and their caches together and then lose these core

00:02:31.280 --> 00:02:38.640
complexes to a separate chiplet that contains the memory controller and the

00:02:35.120 --> 00:02:41.319
io Intel has opted to keep all memory

00:02:38.640 --> 00:02:47.159
and IO functions on die with the logic and cache then they take these near

00:02:44.519 --> 00:02:52.519
complete CPUs and wire them to each other using a fabric interconnect which

00:02:49.599 --> 00:02:57.560
means that in some ways it really is more like just gluing multiple CPUs

00:02:54.879 --> 00:03:01.040
together which is hilarious but hey at least they can glue them together in a

00:02:59.200 --> 00:03:07.120
lot of different ways Intel has announced and supposedly released 52 SKS

00:03:04.799 --> 00:03:12.040
of fourth generation Zeon scalable with more apparently coming so it seems like

00:03:09.680 --> 00:03:17.319
part of their strategy then was to find every possible segment of the market and

00:03:14.239 --> 00:03:19.360
build a just right option for it inside

00:03:17.319 --> 00:03:25.000
it are two well one for now but we'll put the second one in in a minute two

00:03:21.360 --> 00:03:28.200
Intel Zeon Platinum 8468 processors each

00:03:25.000 --> 00:03:30.519
with 48 cores 96 threads and a max turbo

00:03:28.200 --> 00:03:36.640
speed of 3.8 GHz these chips are equipped with 105 megab

00:03:32.760 --> 00:03:40.480
of cache have a 350 wat TDP and come in

00:03:36.640 --> 00:03:40.480
at an eye watering list price of

00:03:59.799 --> 00:04:05.680
in the socket and putting the heat sink on that you actually install the CPU to

00:04:03.799 --> 00:04:12.280
the heat sink and put the whole thing on the socket wild right okay so um

00:04:13.319 --> 00:04:20.799
paste you need so much of it we should call

00:04:17.400 --> 00:04:22.479
this oh a bit got in the socket that's

00:04:20.799 --> 00:04:26.919
probably fine you need so much of this they should call this CPU the paste

00:04:24.600 --> 00:04:34.479
eater also because it's been so successful so far okay sorry uh anyo um

00:04:32.160 --> 00:04:40.320
this thing is this is keyed so it only goes in one way oh my God this is

00:04:37.120 --> 00:04:43.479
ridiculous seriously this is what holds

00:04:40.320 --> 00:04:45.680
this in this ARM yeah thanks I hate it

00:04:43.479 --> 00:04:52.120
no no no no no no no no no this is what helps you get it out so you put it oh

00:04:48.320 --> 00:04:56.720
this levers it out okay so you just

00:04:52.120 --> 00:04:58.680
stick this boy in here oh my god

00:04:56.720 --> 00:05:02.160
seriously Intel you have how many billions of dollars to spend on

00:05:00.000 --> 00:05:06.199
engineering and this is the solution this is exactly the kind of precision

00:05:04.240 --> 00:05:10.600
engineered Jank that I would expect from Intel and then when it's time to get it

00:05:08.000 --> 00:05:13.960
out you use this little lever to pop it up okay we're not going to do that right

00:05:12.000 --> 00:05:18.960
now ridiculous okay we got a little bit of

00:05:16.039 --> 00:05:22.840
thermal compound in the socket but um it looks like it mostly fell between the

00:05:20.720 --> 00:05:29.080
pins so hopefully this is going to work for RAM we have 512 gigs of 4800 megat

00:05:26.360 --> 00:05:33.639
transfer per second ddr5 which sounds like a lot but that is a fraction of the

00:05:32.080 --> 00:05:39.240
maximum that this server can handle thanks to Intel's support for up to

00:05:35.400 --> 00:05:42.280
eight channels of ddr5 per CPU and

00:05:39.240 --> 00:05:44.919
unlike AMD's competitor up to two dims

00:05:42.280 --> 00:05:53.160
per channel that puts the total memory support for this system with two CPUs in

00:05:47.199 --> 00:05:56.160
it at 12 terabytes the fck each CPU also

00:05:53.160 --> 00:05:58.479
brings 80 Lanes of PCIe Gen 5 expansion

00:05:56.160 --> 00:06:02.840
to the table up to five integrated Hardware accelerators assuming you pay

00:06:00.680 --> 00:06:08.240
Intel to unlock them we'll definitely get to that later and on their soon to

00:06:05.360 --> 00:06:14.240
be released zeeon Max series of CPUs they're gluing up to 64 GB of high

00:06:11.440 --> 00:06:20.240
bandwidth memory 2E directly on the package this could end up giving AMD's

00:06:16.919 --> 00:06:22.960
3D vcash technology a run for its

00:06:20.240 --> 00:06:27.120
money coming back to glue for a moment This Server may only have two sockets in

00:06:25.080 --> 00:06:31.479
it but the highest tier of sapphire Rapids chips can glue as many as eight

00:06:29.280 --> 00:06:37.080
of these chips together in a single server meaning that you could have up to

00:06:34.120 --> 00:06:42.720
960 threads of computation on a single board absolutely dwarfing AMD epic which

00:06:40.520 --> 00:06:49.160
is expected to cap out at just two sockets and 512 threads per system even

00:06:46.360 --> 00:06:52.800
on their upcoming Bergamo CPUs that's a lot of cores to do out to your virtual

00:06:50.800 --> 00:06:56.440
machines thankfully you can easily manage and monitor all that stuff

00:06:54.440 --> 00:07:00.160
offsite with pulseway mobile app it helps you track performance and you can

00:06:58.360 --> 00:07:04.000
get notified of a VM is acting out of turn and snipe it from the comfort of

00:07:02.240 --> 00:07:07.759
wherever your phone is not only does pulseway give you complete control over

00:07:05.840 --> 00:07:12.080
your servers and network computers from any mobile phone or tablet but their web

00:07:09.960 --> 00:07:16.440
app is also great for it management if you need to run updates or have concerns

00:07:14.080 --> 00:07:21.039
about CPU temperatures pulseway has intuitive automation tools for Windows

00:07:18.520 --> 00:07:24.599
Mac and Linux this means that you can approach monitoring across your entire

00:07:22.919 --> 00:07:29.120
environment without the need to micromanage every operation now let's

00:07:27.720 --> 00:07:35.080
get this thing up and running shall we maybe give cine bench a go to get an idea of all core

00:07:32.080 --> 00:07:37.919
performance I hope it actually turns on

00:07:35.080 --> 00:07:44.360
we did not anticipate this uh dang it Adam give her one of these

00:07:41.639 --> 00:07:49.000
yeah it's not there yeah okay well hold on let's put the shrouds in before they

00:07:45.800 --> 00:07:50.319
start trying to blow away the computer

00:07:49.000 --> 00:07:55.440
you have to have the touch uh it's not turning on we got blinking lights yeah

00:07:53.479 --> 00:08:01.310
we're g we're blinking up here it's still not turning on holy

00:07:59.039 --> 00:08:02.919
that's it that's the noise we

00:08:02.919 --> 00:08:08.479
want well let's see if it actually boots

00:08:06.360 --> 00:08:13.400
we did get it to turn on but there's a memory training error which means it's

00:08:10.479 --> 00:08:17.919
not seated properly so all we can really do until we find out where our torque

00:08:15.199 --> 00:08:22.759
wrenches because engineering has it is kind of try to get these the

00:08:20.199 --> 00:08:29.159
same very clever packaging this is apparently our torque screwdriver do you

00:08:25.199 --> 00:08:30.280
happen to have a T30 bit okay how do I I

00:08:29.159 --> 00:08:36.039
don't know oh you've never used it

00:08:32.919 --> 00:08:37.360
no damn it ttin well sure yeah I mean

00:08:36.039 --> 00:08:43.560
you've probably used more Torque screwdrivers than I have manual why

00:08:39.959 --> 00:08:49.040
would I use a manual I've got people for

00:08:43.560 --> 00:08:53.680
that okay inch oh my god wow that's fun

00:08:49.040 --> 00:08:56.080
okay so how many newton meters it's uh

00:08:53.680 --> 00:09:01.800
0.904 this is scary this seems like too much okay then go to last the good news

00:08:59.800 --> 00:09:04.680
is if we Breck this thing we've already benchmarked everything so we'll just

00:09:03.240 --> 00:09:08.399
continue the video and pretend this didn't happen well one thing we know for

00:09:06.480 --> 00:09:12.320
sure is that it's a bit of a finicky socket to install in that's good for the

00:09:10.720 --> 00:09:16.680
dozens of you who will deal with this to know the good news as I mentioned before

00:09:15.000 --> 00:09:21.600
is that we already tested all this stuff ahead of time so wow what a fast

00:09:19.120 --> 00:09:26.760
cinebench render that we would have been looking at on this monitor but wait a

00:09:24.440 --> 00:09:30.959
challenger from team red approaches that's right we got another server and

00:09:29.120 --> 00:09:36.680
this one is equipped with a single Flagship AMD epic

00:09:33.320 --> 00:09:39.720
9654 but that's not fair lonus you might

00:09:36.680 --> 00:09:45.480
say you can't put two CPUs up against

00:09:39.720 --> 00:09:47.320
one I can and I will two Intel 8468

00:09:45.480 --> 00:09:54.760
costs $14,400 one Epic 9654 costs around

00:09:51.760 --> 00:09:57.839
$112,000 same number of cores better

00:09:54.760 --> 00:09:59.880
price and on paper they even hit similar

00:09:57.839 --> 00:10:07.519
all core frequencies shall we drag Ras them well spoiler we

00:10:03.040 --> 00:10:10.920
did it already so yes AMD is faster how

00:10:07.519 --> 00:10:14.800
about a blender render haha yes we did

00:10:10.920 --> 00:10:16.680
this already and AMD is faster again but

00:10:14.800 --> 00:10:20.040
as interesting as those tests are we should really be looking at workloads

00:10:18.399 --> 00:10:24.320
that are closer to what these server chips were actually built for like

00:10:22.160 --> 00:10:29.399
managing databases or virtualization unfortunately for Intel

00:10:26.880 --> 00:10:35.000
the story doesn't really get any Ros you're here postgres sees Intel get

00:10:32.040 --> 00:10:39.959
absolutely trounced by AMD I mean losing by that much in one of the most popular

00:10:37.279 --> 00:10:45.880
types of databases that's a really bad sign and what about grx a molecular

00:10:42.800 --> 00:10:47.839
Dynamic simulator well here Intel

00:10:45.880 --> 00:10:52.920
actually manages to pull out a win so their high performance compute dreams

00:10:49.639 --> 00:10:55.120
might not be totally crushed until you

00:10:52.920 --> 00:10:59.760
consider their margin of Victory here compared to their margin of defeat in

00:10:57.399 --> 00:11:04.200
pretty much any other General compute task epic is just looking like the

00:11:02.360 --> 00:11:09.279
obvious choice here unless you know for a fact that you're going to run one

00:11:06.760 --> 00:11:14.120
application and one only that happens to benefit from Intel's chip design and for

00:11:12.399 --> 00:11:19.519
all the Ruckus that Intel has made about efficiency improvements they lose in

00:11:16.680 --> 00:11:23.880
every single one of our tests and we understand that a dual socket system

00:11:21.440 --> 00:11:28.399
will use more power than a single socket system so this isn't a onetoone

00:11:25.480 --> 00:11:33.760
comparison but here's the thing if AMD only needs a single socket for 96 cores

00:11:31.839 --> 00:11:38.360
well what am I supposed to do hamstring it by saying no you must have more

00:11:35.560 --> 00:11:42.560
sockets of course not and neither would any decent procurement officer who's

00:11:40.160 --> 00:11:48.360
evaluating this hardware for their needs if you want a many core server Intel

00:11:45.279 --> 00:11:50.200
Sapphire Rapids costs more to buy and

00:11:48.360 --> 00:11:57.040
costs more to run and they have no one to blame but themselves

00:11:52.880 --> 00:11:59.360
today but as we said Intel is laser

00:11:57.040 --> 00:12:03.560
focused on the future and that future is AI so they put extra effort into

00:12:01.560 --> 00:12:08.720
implementing accelerators for AI inferencing and other common workloads

00:12:05.920 --> 00:12:12.800
directly onto the CPU now if you're a consumer the concept of a hardware

00:12:10.480 --> 00:12:17.519
accelerator might already be familiar unlike the general purpose cores on your

00:12:14.839 --> 00:12:22.399
chip accelerators are hyper efficient for specific workloads but then

00:12:20.000 --> 00:12:26.279
essentially useless for anything else Hardware video decoders for example are

00:12:24.480 --> 00:12:30.480
what makes it possible to watch YouTube for hours on end without your phone

00:12:27.880 --> 00:12:35.160
turning into a hot potato but that is all they do and they take up valuable

00:12:32.680 --> 00:12:39.480
die space Intel clearly thought that was worthwhile and the first big one they

00:12:36.880 --> 00:12:44.560
built in is AMX which stands for advanced Matrix

00:12:42.399 --> 00:12:49.480
extensions sorry sorry you can think of this one as kind

00:12:46.959 --> 00:12:54.800
of an evolution of AVX or Advanced Vector extensions it is ideally going to

00:12:52.560 --> 00:12:59.320
provide a massive increase in the CPU's ability to perform AI inference with

00:12:57.040 --> 00:13:04.199
Intel sighting up to a 10 times improvement over ABX 512 I mean to be

00:13:02.120 --> 00:13:10.199
clear this is not going to replace dedicated accelerators like NVIDIA's T4

00:13:06.480 --> 00:13:12.959
cards but it's a lot better than nothing

00:13:10.199 --> 00:13:17.720
so do they win out in a light AI workload then that's a yes with a

00:13:15.920 --> 00:13:23.160
capital h y AMX can perform AI tasks with

00:13:20.880 --> 00:13:27.440
double or triple the performance of AVX 512 that's fast almost as fast as you'll

00:13:25.920 --> 00:13:32.480
be in our new LTT tracksuit in general the lead over AMD

00:13:30.279 --> 00:13:37.440
is Stark even if this doesn't translate to gains in every AI workload and it

00:13:35.199 --> 00:13:42.800
could prove to be more and more useful into the future as we see small AI tasks

00:13:40.560 --> 00:13:47.800
sneak their way into software that might run on a system like this as opposed to

00:13:45.360 --> 00:13:50.920
a dedicated machine learning box that's stuffed full of

00:13:49.079 --> 00:13:55.000
gpus what do the rest of these accelerators do well first their data

00:13:53.320 --> 00:13:59.160
streaming accelerator should improve performance in tasks where data is

00:13:56.680 --> 00:14:03.560
constantly generated by hundreds if not thousands of sources like logging the

00:14:01.360 --> 00:14:06.120
Telemetry data of many different sensors that you might be monitoring after a

00:14:04.800 --> 00:14:10.839
scheduled remote patch that you initiated with pulseway the inmemory

00:14:08.480 --> 00:14:15.639
analytics accelerator moves important analytical calculations of databases

00:14:13.000 --> 00:14:19.240
away from the CPU course then Intel's Dynamic load balance are supposed to

00:14:17.440 --> 00:14:24.600
help manage and distribute work across the many cores in your CPU finally quick

00:14:22.320 --> 00:14:29.120
Assist Technology accelerates encryption and compression the idea here is pretty

00:14:27.320 --> 00:14:32.079
sound and you can think of it kind of like how pulseway works I know I keep

00:14:30.680 --> 00:14:36.800
bringing pulseway into it they sponsor the video what do you want so with

00:14:33.759 --> 00:14:39.600
pulseway you could do everything

00:14:36.800 --> 00:14:43.440
manually right but it's slower and it's less efficient so you can think of

00:14:41.560 --> 00:14:48.320
pulseway as kind of like a hardware accelerator for your workpace you spend

00:14:46.560 --> 00:14:52.480
some of your time using their intuitive automation workflow then your repetitive

00:14:50.519 --> 00:14:57.160
tasks run on autopilot allowing you to focus your brain power on important

00:14:54.240 --> 00:15:02.040
things which improves productivity or saves power uh whichever it appeals to

00:14:59.440 --> 00:15:05.639
you more the thing is though while we would love to put all of these

00:15:03.880 --> 00:15:10.600
accelerators little timesavers efficiency makers to the test it's still

00:15:08.240 --> 00:15:15.680
early days and software support simply isn't there yet and it might never get

00:15:13.880 --> 00:15:20.360
there it all depends on whether organizations end up making use of these

00:15:18.240 --> 00:15:24.360
accelerators there is one feature though that Intel has ready at launch and it's

00:15:22.279 --> 00:15:29.360
got their shareholders a buzz even if nobody else is impressed it's called

00:15:26.720 --> 00:15:34.800
Intel on demand and frankly calling it a feature seems kind of generous what it

00:15:31.880 --> 00:15:40.839
does is it allows customers to upgrade their CPUs by paying Intel to unlock the

00:15:38.079 --> 00:15:45.360
hardware accelerators that are already in the Silicon on the chip but are

00:15:43.360 --> 00:15:49.759
blocked through software the most charitable possible

00:15:47.959 --> 00:15:54.279
take here is that this will allow Enterprise customers to Save A Buck at

00:15:51.720 --> 00:15:59.000
the time of purchase and then upgrade their chips down the line to accelerate

00:15:56.079 --> 00:16:02.360
new workloads no need to remove replace place and reconfigure a brand new server

00:16:01.040 --> 00:16:06.240
which can be a hassle especially with how finicky these CPUs are you just

00:16:04.680 --> 00:16:10.880
unlock the features you need and you're on your way massively reducing the time

00:16:08.319 --> 00:16:16.360
to value for customers well that's great right savings on initial purchases is a

00:16:13.079 --> 00:16:18.800
good thing well kind of except that at

00:16:16.360 --> 00:16:24.360
risk of painting their entire 52 skew lineup with a broad brush Sapphire

00:16:20.959 --> 00:16:26.120
Rapids is a pretty bad deal up front and

00:16:24.360 --> 00:16:30.639
that's even before you pay Intel's ransom for unlocking the hardware they

00:16:28.120 --> 00:16:36.560
already sold you you I mean I'm not sure how it happened but Intel managed to

00:16:32.839 --> 00:16:39.160
price their topend chip 50% higher than

00:16:36.560 --> 00:16:43.920
amds which as we've already discussed kind of dunks on it so then how is it

00:16:41.680 --> 00:16:49.399
that Intel still has so much of the Enterprise market and why aren't more

00:16:46.360 --> 00:16:51.720
customers crossing the line over to AMD

00:16:49.399 --> 00:16:57.560
well it's a combination of factors first up is long-term deals many of these

00:16:54.680 --> 00:17:02.199
Sapphire rapid CPUs were purchased long ago and customers just been waiting

00:17:00.399 --> 00:17:06.520
through the delays for them to finally be delivered Intel also has both the

00:17:04.760 --> 00:17:10.240
manufacturing capacity and the engineering staff to support the kinds

00:17:08.720 --> 00:17:15.360
of customers who buy chips in the hundreds or the thousands or even the

00:17:12.559 --> 00:17:19.760
tens of thousands rather than one by one like you or I might do over the next

00:17:18.039 --> 00:17:24.400
couple of years though it's clear that AMD is going to continue to consume

00:17:22.559 --> 00:17:30.240
Intel's valuable data center market share especially in servers that rely on

00:17:26.760 --> 00:17:31.960
general purpose High Core count SE CPUs

00:17:30.240 --> 00:17:35.840
for specific workloads that can take advantage of Intel's accelerators though

00:17:33.880 --> 00:17:41.799
AMD has a lot of work to do and it's going to take some time but that's okay

00:17:38.840 --> 00:17:47.280
they've clearly got it and as bad as Intel's quarterly results and forecast

00:17:44.200 --> 00:17:49.160
look so do they I mean turning a ship as

00:17:47.280 --> 00:17:53.480
big as Intel takes years and they have plenty of cash on hand to navigate this

00:17:50.919 --> 00:17:56.480
rough patch I mean look at AMD many were skeptical that they would ever be able

00:17:55.280 --> 00:18:01.679
to turn it around after their embarrassing bulldozer line of CPUs but

00:17:59.159 --> 00:18:05.280
a CEO Swap and 5 or 6 years later the company was fully revitalized as they

00:18:03.440 --> 00:18:09.679
began launching product after product that made Intel shake in their boots

00:18:07.600 --> 00:18:12.640
like like what's happening now thanks again to pulseway when it comes to

00:18:11.120 --> 00:18:16.159
monitoring and management these folks are great so if you're an MSP or you

00:18:14.799 --> 00:18:22.200
just want to be able to keep tabs on your system while you're out at dinner pulseway has your back they're hooking

00:18:19.640 --> 00:18:25.880
up LTT viewers with a no risk-free trial so check them out at the link in the

00:18:23.640 --> 00:18:29.039
video description thanks to AMD for getting Intel to actually start

00:18:27.039 --> 00:18:32.840
competing again and thanks to pulse for sponsoring this video right now they're

00:18:30.840 --> 00:18:36.000
hooking up our viewers with a free trial so if you want to use pulseway to manage

00:18:34.360 --> 00:18:39.400
your devices like we do check them out in the video description