WEBVTT

00:00:00.080 --> 00:00:06.319
this is not the future of the pc

00:00:03.679 --> 00:00:12.719
this is if the enthusiast in you screams that no you can't just replace gpus and

00:00:09.440 --> 00:00:16.560
cpus with integrated socs i am here to

00:00:12.719 --> 00:00:19.119
tell you that apple silicon gober apple

00:00:16.560 --> 00:00:24.320
obviously knows it and make no mistake Intel AMD and NVIDIA all know it too the

00:00:22.640 --> 00:00:29.679
writing is on the wall and now it's just a matter of time pcs need to change

00:00:27.599 --> 00:00:34.000
and i'm going to tell you how and why right after i tell you about our

00:00:31.359 --> 00:00:38.320
sponsor glasswire glasswire lets you see past and present network activity detect

00:00:36.160 --> 00:00:41.920
malware on your pc or Android device and block its connections to prevent things

00:00:39.840 --> 00:00:45.840
from getting worse use offer code Linus for 25 off at the link below

00:00:53.440 --> 00:01:00.079
a lot of you really liked our how to build a pc the last guide you'll ever

00:00:57.120 --> 00:01:05.119
need and i'm really glad that you do because it took so much effort to make

00:01:02.559 --> 00:01:08.960
but look at that run time and it could have been even longer

00:01:06.880 --> 00:01:14.320
think about how many components go into a typical pc build and then look at this

00:01:11.280 --> 00:01:16.560
unassuming little box the mac studio in

00:01:14.320 --> 00:01:21.600
our testing the m1 ultra variant pulled a grand total of

00:01:18.200 --> 00:01:23.600
223.4 watts from the wall under the

00:01:21.600 --> 00:01:28.159
heaviest load that we could give it and the m1 max version didn't even break

00:01:25.600 --> 00:01:34.400
three digits if we ignore the rosetta results they achieved 88 and 61 percent

00:01:31.200 --> 00:01:37.280
respectively of the core i9 12900k and

00:01:34.400 --> 00:01:43.200
rtx 3090 bench's overall performance in our productivity testing at just 34 and

00:01:39.840 --> 00:01:44.560
15 of the total system power at most the

00:01:43.200 --> 00:01:51.040
12 900k the CPU alone can draw more power than

00:01:47.920 --> 00:01:52.880
the whole m1 ultra mac studio and

00:01:51.040 --> 00:01:57.759
developers are still figuring out how to optimize for apple silicon now

00:01:55.680 --> 00:02:02.320
look at that pc and all its components again all the raw materials that go into

00:02:00.000 --> 00:02:07.200
making it what it is it's wasteful right in almost every way except economically

00:02:04.560 --> 00:02:11.920
it's not that x86 is inherently bad but it has limitations that even Intel have

00:02:09.840 --> 00:02:16.480
tried to overcome themselves with projects like titanium while extensions

00:02:14.560 --> 00:02:20.720
have given it more life its complex instruction set design is at its core an

00:02:19.120 --> 00:02:28.160
old one that dates back to the late 1970s literally every x86 based CPU

00:02:25.120 --> 00:02:31.440
boots up in 8086 compatibility mode and

00:02:28.160 --> 00:02:35.280
can in theory run software made for the

00:02:31.440 --> 00:02:36.959
og ibm pc precious die space and

00:02:35.280 --> 00:02:41.519
transistors are used for supporting instructions and modes that might be

00:02:39.040 --> 00:02:46.319
redundant or not even remotely useful today but must be there to maintain

00:02:44.160 --> 00:02:51.360
compatibility with the architecture and now if i'm Intel or AMD looking at my

00:02:48.720 --> 00:02:56.800
options i'm sweating a little and weighing a switch to ARM while neither

00:02:53.760 --> 00:02:59.519
currently sells an armed CPU both hold

00:02:56.800 --> 00:03:04.879
armed licenses and AMD is on record saying that they're willing and able to

00:03:01.680 --> 00:03:07.360
do it it wasn't that long ago even that

00:03:04.879 --> 00:03:12.400
Intel had their own reduced instruction set CPU called the i860 and i-960 and

00:03:10.800 --> 00:03:15.920
you'd be kidding yourself if you thought they weren't prototyping stuff on the

00:03:14.080 --> 00:03:21.200
side like apple was when they were suffering through 14 nanometer plus plus

00:03:18.480 --> 00:03:24.959
plus a change in the pc industry would of course have a teething period like

00:03:22.879 --> 00:03:30.159
apple silicon is currently experiencing but it could eventually be a good thing

00:03:28.000 --> 00:03:35.519
and not just for efficiency think about it Intel and AMD are alone in the CPU

00:03:33.200 --> 00:03:41.280
race right now just like AMD and NVIDIA are currently alone in the GPU race and

00:03:38.799 --> 00:03:46.000
if Intel had their way in the 90s or in 2015 there'd be only one

00:03:44.400 --> 00:03:50.400
thanks to Intel's control over performance and efficiency enhancing

00:03:47.599 --> 00:03:55.760
extensions used on modern cpus like sse and avx and AMD's control over the

00:03:52.640 --> 00:03:58.159
64-bit extensions that every desktop CPU

00:03:55.760 --> 00:04:02.400
since core 2 has used we don't have any other desktop CPU manufacturers anyone

00:04:00.400 --> 00:04:06.799
entering the x86 market would have to reverse engineer those extensions to

00:04:04.720 --> 00:04:10.879
compete because AMD and Intel have a technology sharing agreement that

00:04:08.159 --> 00:04:16.400
forbids sub licensing and expires if either company changes hands by contrast

00:04:13.760 --> 00:04:20.239
ARM doesn't sell anything directly they develop the base specifications and then

00:04:18.400 --> 00:04:23.520
they license that to other companies who then have the option of tweaking and

00:04:21.919 --> 00:04:28.320
adding to those designs to tailor them to suit their needs this is how apple

00:04:25.919 --> 00:04:33.520
silicon came to be and if ARM took off on the desktop we could see not only

00:04:30.720 --> 00:04:37.919
mobile players like qualcomm samsung and mediatek enter the ring but maybe google

00:04:35.759 --> 00:04:43.759
or even hp they used to make their own cpus once too and we can't forget NVIDIA

00:04:40.800 --> 00:04:48.800
whose last consumer ARM soc uses GPU cores from 2014 and lives on in the

00:04:46.160 --> 00:04:53.199
nintendo switch and shield tv plus actually it's kind of bizarre that

00:04:51.120 --> 00:04:57.840
NVIDIA hasn't continued to produce consumer socs like perhaps contractual

00:04:56.080 --> 00:05:01.759
obligations within and outside of NVIDIA's control had something to do

00:04:59.759 --> 00:05:05.840
with it because not only were they close to straight up buying ARM a sign that

00:05:04.160 --> 00:05:11.199
they knew times were changing but they're also still actively developing

00:05:08.320 --> 00:05:15.280
ARM socs just not for consumers their grace data center soc for ali

00:05:13.120 --> 00:05:19.680
acceleration has a surprisingly similar level of integration to apple silicon on

00:05:17.120 --> 00:05:24.400
paper and is just begging to be scaled down so that consumers can use it just

00:05:21.680 --> 00:05:31.759
like the tesla's and not quadro gpus do with g-force but let's back up a sec

00:05:27.600 --> 00:05:35.120
duopoly aside if x86 is so bad and ARM

00:05:31.759 --> 00:05:38.720
is so good why is x86 still the king of

00:05:35.120 --> 00:05:42.000
the pc well prior to apple silicon

00:05:38.720 --> 00:05:44.240
you'd be forgiven for thinking that ARM

00:05:42.000 --> 00:05:50.800
chips simply can't perform like Intel and AMD qualcomm has a

00:05:47.759 --> 00:05:53.039
virtual monopoly over non-apple ARM pcs

00:05:50.800 --> 00:05:57.039
and a literal monopoly in the Windows ecosystem thanks to an exclusivity

00:05:55.039 --> 00:06:01.280
agreement with microsoft when that deal was penned it probably did make some

00:05:58.800 --> 00:06:06.080
sense qualcomm was the third biggest chip maker even as of mid-2020 and

00:06:04.240 --> 00:06:11.039
there's still a giant who's recently been eyeing a slice of the ARM pie for

00:06:08.400 --> 00:06:15.440
themselves so assuming qualcomm would become the biggest supplier of desktop

00:06:13.120 --> 00:06:19.919
ARM chips it made sense for microsoft to want to optimize for that hardware

00:06:17.440 --> 00:06:24.639
unfortunately for microsoft and for us there is no polite way to put this even

00:06:22.240 --> 00:06:28.319
qualcomm's latest chips suck which recently prompted google to follow

00:06:26.479 --> 00:06:32.000
samsung and apple and going their own way with tensor and may eventually

00:06:30.160 --> 00:06:36.960
prompt microsoft to do the same qualcomm's utter failure to compete in

00:06:34.160 --> 00:06:41.360
the pc space so far has given Intel AMD and especially NVIDIA precious breathing

00:06:39.360 --> 00:06:45.600
room and we're not saying that the pc as we know it is dead today it definitely

00:06:43.520 --> 00:06:50.080
still has some life left and they'll first probably try to integrate x86 socs

00:06:48.560 --> 00:06:53.919
more tightly before they consider switching over Intel for example could

00:06:52.160 --> 00:06:57.919
very well be working on an soc that combines alder lake's efficiency cores

00:06:56.000 --> 00:07:01.360
with a larger number of xc graphics cores and ai engine blocks taking the

00:06:59.680 --> 00:07:05.759
dye space that would otherwise have been allocated to the bigger performance

00:07:02.960 --> 00:07:10.000
cores even so reduced instruction set architectures proved their worth in the

00:07:07.680 --> 00:07:15.360
late 90s with early ARM licenses and with power pc which was famously used in

00:07:12.479 --> 00:07:19.199
the first imax clock for clock power pc tended to outperform x86 by a wide

00:07:17.440 --> 00:07:22.639
margin though this did depend on the task and eventually powerpc hit a

00:07:21.199 --> 00:07:27.599
performance and efficiency wall just like x86 is today Intel's titanium for

00:07:25.680 --> 00:07:33.120
its part was supposed to be a sort of successor to risk as a concept by using

00:07:30.319 --> 00:07:38.639
one hp termed explicitly parallel instruction computing or epic yes

00:07:36.560 --> 00:07:43.680
naming has always been bad unlike our swag it from ltdstore.com it describes

00:07:41.280 --> 00:07:49.120
exactly what it is a sweater and a jacket as we know today titanium didn't

00:07:46.560 --> 00:07:53.599
work out so well still even Intel saw that the x86's days were numbered their

00:07:51.360 --> 00:07:57.520
plan for pentium 4 was more or less to just clock it as high as it would go and

00:07:55.599 --> 00:08:01.680
hope that the numbers kept going up and now that we're back to this phase again

00:07:59.520 --> 00:08:05.840
where we can't physically fit more cores on a die because new process nodes are

00:08:03.919 --> 00:08:09.840
providing less additional density something's got to give

00:08:07.440 --> 00:08:15.440
the current solution stacking dies will only get you so far when you're already

00:08:11.840 --> 00:08:17.680
pushing 241 watts by comparison a lot of

00:08:15.440 --> 00:08:20.960
ARM's advantages today come not strictly from its reduced instruction set but

00:08:19.360 --> 00:08:24.560
from how it's typically bundled into a system-on-chip configuration where

00:08:22.639 --> 00:08:28.879
virtually the entire computer is super close to the CPU cores why does this

00:08:26.639 --> 00:08:33.680
matter when we already have apus it's all about density and power draw simply

00:08:31.440 --> 00:08:37.919
expanding the size of the die isn't normally a good idea from power and

00:08:35.599 --> 00:08:43.200
thermal perspective but ARM cores are typically much smaller than x86 and sip

00:08:40.560 --> 00:08:47.279
far less power so not only can you fit more of them into the same space

00:08:45.360 --> 00:08:53.600
you're also not as limited in terms of die size in general m1 ultra's die area

00:08:50.320 --> 00:08:54.880
is bigger than the rtx 3090 ti let's do

00:08:53.600 --> 00:09:01.120
a case study AMD's 8 core ryzen 7 5700g desktop apus

00:08:58.959 --> 00:09:05.839
are somewhere between an m1 and m1 pro in terms of die size the apu dedicates

00:09:03.839 --> 00:09:11.040
just over a quarter of the utilized die space to CPU cores and cache alone and

00:09:08.800 --> 00:09:15.600
just 10 to the GPU for thermal management there's also a fair amount of

00:09:13.120 --> 00:09:20.240
unused white space even within the cores themselves meanwhile even m1's big

00:09:18.080 --> 00:09:25.680
performance CPU cores are tiny by comparison with all cores taking just

00:09:22.640 --> 00:09:28.000
over 20 percent of that smaller die area

00:09:25.680 --> 00:09:33.519
the GPU takes up double the relative die space as the apu which means there's way

00:09:30.560 --> 00:09:37.600
more GPU horsepower on tap and still room for other functions like the

00:09:35.360 --> 00:09:41.680
massive media engine neural engine and storage controller there's very little

00:09:39.600 --> 00:09:46.080
white space by comparison not only are those ARM cores efficient enough to

00:09:43.279 --> 00:09:49.839
handle such a compact layout but when so much of the system is contained within

00:09:47.839 --> 00:09:53.920
it significantly less energy has to be used for signaling which also happens to

00:09:51.920 --> 00:09:59.120
be much faster if we tried to push an x86 processor to these densities in

00:09:56.240 --> 00:10:02.560
proximity with this many GPU cores it would quickly become a thermal

00:10:00.640 --> 00:10:07.920
meltdown event sort of like the opposite of how alex had to save a frozen core i9

00:10:05.200 --> 00:10:11.040
12900 ks literally frozen go check it out on flow plane along with more

00:10:09.120 --> 00:10:17.360
exclusive when you're done here and it's scales m1 max sports a larger die than

00:10:14.320 --> 00:10:20.640
the rtx 3070 and m1 ultra is about a

00:10:17.360 --> 00:10:22.959
third again bigger than the rtx 3090 ti

00:10:20.640 --> 00:10:27.680
with the die that large the computy bits of the 3090 ti would take up roughly

00:10:25.040 --> 00:10:31.440
half the m1 ultra's die area while apple currently dedicates just over a third

00:10:29.839 --> 00:10:35.839
that means that if apple built a dedicated big chip design and weren't

00:10:33.680 --> 00:10:39.440
using their own metal API separate from the rest of the industry they

00:10:36.959 --> 00:10:44.160
legitimately could reach high-end GPU levels of real world performance a shame

00:10:42.000 --> 00:10:47.519
that apple had to go their own way but imagine if NVIDIA were to build such

00:10:46.640 --> 00:10:52.880
a thing they could corner the ARM desktop or notebook space very quickly if

00:10:51.040 --> 00:10:56.959
developers were on board especially it's not as simple as gluing them together

00:10:54.959 --> 00:11:01.200
obviously and that GeForce power consumption is still a problem but it's

00:10:59.120 --> 00:11:05.839
also not an impossible hurdle and with the success of apple silicon despite its

00:11:03.680 --> 00:11:09.600
growing pains it's impossible for the industry to ignore the advantages to

00:11:07.519 --> 00:11:13.200
this approach you can have excellent performance coupled with incredibly low

00:11:11.519 --> 00:11:17.680
power draw on a much more open architecture that allows for bigger or

00:11:15.839 --> 00:11:21.519
more aspirational system integrators to tweak things as they see fit this

00:11:19.600 --> 00:11:25.279
potential future will suck for enthusiasts who like getting their hands

00:11:23.200 --> 00:11:29.279
dirty but the alternative is laid bare before us power targets shooting up at

00:11:27.680 --> 00:11:33.360
record pace to the point where we're looking at new connectors to help

00:11:30.720 --> 00:11:38.720
deliver all the juice this generation of hardware now is already incredibly power

00:11:36.399 --> 00:11:41.920
hungry with Intel and AMD squeezing higher and higher frequencies out of

00:11:40.240 --> 00:11:46.959
their cores and NVIDIA rumored to require up to 600 watts for their

00:11:44.320 --> 00:11:51.760
next-gen top-end GPU there's really no backing down we have to ask ourselves

00:11:49.600 --> 00:11:54.880
is it worth it and at what point does it stop being worth it

00:11:53.519 --> 00:11:59.120
at least it's always worth listening to the segways the sponsors like backblaze

00:11:57.440 --> 00:12:03.040
starting at just seven dollars a month backblaze offers an affordable and easy

00:12:01.120 --> 00:12:06.639
to use cloud backup solution they make it simple by allowing you to backup

00:12:04.640 --> 00:12:10.639
almost anything from your mac or pc and access it from anywhere in the world via

00:12:08.399 --> 00:12:13.600
their web or mobile apps backblaze even lets you restore your data by mail a

00:12:12.320 --> 00:12:18.560
hard drive with your data will be shipped right to your door and once you're done you can return the hard

00:12:16.480 --> 00:12:23.279
drive within 30 days for a full refund with over 55 billion files restored and

00:12:21.680 --> 00:12:27.519
two exabytes of data under their management backblaze has got you covered

00:12:25.600 --> 00:12:31.519
so don't be that person who forgets to back up their important files sign up

00:12:29.440 --> 00:12:35.680
and get a 15 day free trial with no credit card required today at

00:12:33.040 --> 00:12:39.040
backblaze.com LTT thanks for watching guys this one's a

00:12:37.040 --> 00:12:42.480
bit more philosophical than usual so maybe go check out something more down

00:12:40.560 --> 00:12:48.320
and dirty for a palette cleanser like our 6.9 gigahertz GPU released on six

00:12:45.040 --> 00:12:48.320
nine it's nice