WEBVTT

00:00:00.839 --> 00:00:07.120
64 cores

00:00:04.600 --> 00:00:15.519
256 threads running at 1.3 to 1.5 gigahertz with 32

00:00:12.559 --> 00:00:19.359
megabytes of cash meet

00:00:16.480 --> 00:00:24.000
the knights landing zeon phi one of the most insane cpus in the world

00:00:22.640 --> 00:00:29.279
how insane well how does a theoretical two and a

00:00:26.320 --> 00:00:33.840
half teraflop sound got no context for that

00:00:31.279 --> 00:00:37.280
don't stress it's like it's like a lot like

00:00:34.719 --> 00:00:41.040
like like look how big this thing is i can take over the world with this sort

00:00:38.800 --> 00:00:44.000
of power but first i need to pay the bills which

00:00:42.640 --> 00:00:48.800
is why today's video is brought to you by tunnelbear tunnelware makes it easy

00:00:46.480 --> 00:00:52.079
to privately and securely browse a more open internet try telebear for free at

00:00:50.640 --> 00:01:02.160
tunnelbear.com LTT

00:01:02.160 --> 00:01:08.640
so some background information is in order here as this is no ordinary CPU

00:01:07.200 --> 00:01:14.080
knights landing is the current generation of xeon phi Intel's version

00:01:11.600 --> 00:01:19.600
of what's referred to as a mini core processor which is exactly what it

00:01:16.400 --> 00:01:22.640
sounds like it's built around the idea

00:01:19.600 --> 00:01:25.520
of trading off per thread performance

00:01:22.640 --> 00:01:31.040
for an obscene number of threads and maximum throughput

00:01:28.080 --> 00:01:37.119
with the theory being that a simpler scaled down core design can dramatically

00:01:34.240 --> 00:01:40.640
scale up in number while still utilizing existing tools

00:01:38.880 --> 00:01:47.759
so we can actually trace knight's landings lineage back to project larabee

00:01:44.000 --> 00:01:50.640
first unveiled in mid 2008 as a GPU

00:01:47.759 --> 00:01:55.520
designed around the x86 architecture Intel even announced their intent to

00:01:53.040 --> 00:01:58.479
release a consumer graphics card version of it back in 2010

00:01:57.680 --> 00:02:04.880
but since nobody is talking about dedicated team

00:02:02.320 --> 00:02:10.000
blue graphics cards today you should probably realize that that

00:02:06.880 --> 00:02:12.080
never launched so xeon phi is what rose

00:02:10.000 --> 00:02:18.319
from the ashes of that project first appearing in 2012 as a pci express

00:02:15.280 --> 00:02:20.959
add-in card codename knight's corner

00:02:18.319 --> 00:02:26.400
and it actually did end up getting used in what was the world's fastest

00:02:23.360 --> 00:02:28.640
supercomputer until june 2016. the

00:02:26.400 --> 00:02:33.519
newest revision though knight's landing looks a little different it launched

00:02:30.560 --> 00:02:41.040
last year and unlike knight's corner it works exclusively in the lga 3647 socket

00:02:37.840 --> 00:02:45.599
that high-end xeon pearly cpus also slot

00:02:41.040 --> 00:02:47.680
into so adding support for avx-512 and

00:02:45.599 --> 00:02:52.400
its many extensions gives knight's landing a lot of flexibility in its

00:02:50.400 --> 00:02:56.640
target market so then who's the target market i'm glad you

00:02:54.239 --> 00:03:01.360
asked uses for this include protein folding simulation

00:02:58.319 --> 00:03:03.599
weather prediction ai and neural network

00:03:01.360 --> 00:03:09.280
research and development and molecular simulation with libraries like

00:03:05.920 --> 00:03:12.319
tensorflow and the uiuc's namd being

00:03:09.280 --> 00:03:15.519
able to take advantage of it as well as

00:03:12.319 --> 00:03:19.840
GPU acceleration okay then Linus

00:03:15.519 --> 00:03:19.840
so why do you have one

00:03:20.080 --> 00:03:27.840
um i guess because super micro pulled a rookie mistake and

00:03:25.120 --> 00:03:31.519
sent it to me i mean obviously all i was gonna do was game on it and i'm actually

00:03:30.159 --> 00:03:35.120
kidding about that at 1.3 gigahertz i can tell you guys

00:03:33.519 --> 00:03:38.799
right now without firing up a single benchmark what the gaming performance

00:03:36.799 --> 00:03:43.440
will look like like that

00:03:40.560 --> 00:03:48.400
as for productivity though yeah right out of the gate it falls flat

00:03:45.280 --> 00:03:50.640
on its face coming in at under half of

00:03:48.400 --> 00:03:55.840
core i9s and nearly a third of threadripper's score in 7-zip and

00:03:53.040 --> 00:04:00.959
cinebench CPU mark and real bench not looking much better even blender a

00:03:58.879 --> 00:04:06.319
conventionally multi-core friendly benchmark sees terrible performance

00:04:04.080 --> 00:04:11.599
given the 1900 price tag for the chip alone

00:04:09.840 --> 00:04:15.840
so no these results won't do it all

00:04:13.200 --> 00:04:19.680
xeon phi is clearly unlike anything we've ever tried to benchmark before

00:04:17.359 --> 00:04:24.400
you'd need like a phd or something to properly evaluate this thing

00:04:21.840 --> 00:04:29.600
which is where dr kinghorn our friend over at puget systems comes in as a

00:04:27.280 --> 00:04:34.639
chemist and mathematician dr kinghorn is no stranger to high-level computation

00:04:31.680 --> 00:04:39.199
like this and agreed to remote into our night's landing workstation and do some

00:04:36.320 --> 00:04:45.040
of the testing for us while also running his own 14 core xeon 2690 v4 workstation

00:04:43.199 --> 00:04:50.160
through the same suite to give us a point of comparison the results

00:04:48.000 --> 00:04:55.120
actually pretty surprising in linpack the higher core clock of the

00:04:52.880 --> 00:05:01.520
traditional xeon gives it an edge in smaller problem sizes but beyond 5000

00:04:58.720 --> 00:05:06.560
the tables turn dramatically with the molecular dynamics library and

00:05:03.520 --> 00:05:08.479
AMD things seem to scale predictably

00:05:06.560 --> 00:05:12.800
with the number of atoms being simulated and all the tests wind up being far less

00:05:10.800 --> 00:05:16.000
favorable to the xeon phi the greatest gains in this benchmark

00:05:14.479 --> 00:05:19.039
actually come from dedicated compute gpus

00:05:18.240 --> 00:05:23.120
though that is fine for the phi these days

00:05:21.280 --> 00:05:28.080
since it sits in a socket now instead of taking up valuable PCIe slots

00:05:25.919 --> 00:05:32.240
tensorflow neural networks bring things back though into the xeon phi's

00:05:30.000 --> 00:05:36.560
wheelhouse especially in batch sizes larger than 64.

00:05:34.400 --> 00:05:40.560
so xeon phi takes a significant lead over its traditional xeon cousin

00:05:38.800 --> 00:05:45.919
but it should be noted that both of them get creamed by a GPU

00:05:44.160 --> 00:05:50.240
further highlighting the importance of large banks of compute gpus for tasks

00:05:48.880 --> 00:05:55.039
like these though that doesn't mean that xeon phi

00:05:52.639 --> 00:05:58.479
is useless its dramatically better performance than a standard xeon makes

00:05:57.280 --> 00:06:03.280
it great for supplementing GPU compute in workloads

00:06:00.800 --> 00:06:07.759
like these which brings us neatly to the conclusion then

00:06:04.800 --> 00:06:12.080
not quite as good as a GPU for tasks that can be accelerated

00:06:10.160 --> 00:06:16.880
far better than a traditional xeon though in most

00:06:14.319 --> 00:06:20.800
and finally utterly worthless for just about anything on the Windows desktop i

00:06:19.039 --> 00:06:23.840
was actually surprised to see Windows even install we ended up needing a

00:06:22.400 --> 00:06:29.199
server version of Windows to see all the cores but that is sort of irrelevant

00:06:26.560 --> 00:06:32.639
anyway because this is intended for use in super computers where like

00:06:31.520 --> 00:06:37.440
massive nodes of these things can all work

00:06:34.880 --> 00:06:42.479
together to solve complex problems along with the banks of gpus that accompany

00:06:39.919 --> 00:06:45.680
them i mean you're almost certainly not going to be buying anything like this

00:06:44.000 --> 00:06:51.600
personally anytime soon unless you're like a you like to science for funsies

00:06:48.960 --> 00:06:56.319
with that said though you yes you are very likely to benefit from the work

00:06:53.520 --> 00:07:00.319
being done using hardware like this and we're glad we got the opportunity to

00:06:57.840 --> 00:07:04.479
play with it it was fun speaking of things that are fun

00:07:02.560 --> 00:07:09.680
it's fun to tell you guys about squarespace and thank them for

00:07:06.240 --> 00:07:12.319
sponsoring this video squarespace has 24

00:07:09.680 --> 00:07:17.440
7 support via live chat so anyone can create a beautiful

00:07:13.840 --> 00:07:20.400
functional website in just a few minutes

00:07:17.440 --> 00:07:24.400
if all you want is like uh a one-page online presence that's their cover pages

00:07:22.400 --> 00:07:28.000
feature and they've got tons of other great features as well including their

00:07:25.840 --> 00:07:32.960
logo designer the ability to publish in the apple news format and not to mention

00:07:31.440 --> 00:07:36.639
that the whole creation tool is cloud-based so it's always up you can

00:07:35.199 --> 00:07:40.400
access it from anywhere and updating your site is as simple as like dragging

00:07:38.479 --> 00:07:44.479
over some pictures clickity-clacking some text and you are ready to rock

00:07:42.400 --> 00:07:47.919
squarespace starts at just 12 bucks a month and you can start a trial with no

00:07:46.080 --> 00:07:52.400
credit card required by heading to the link in the video description

00:07:49.440 --> 00:07:56.319
squarespace.com forward slash LTT and when you decide to sign up for

00:07:53.919 --> 00:08:01.360
squarespace forever you can use offer code LTT to get 10 off

00:07:58.960 --> 00:08:03.759
on your first purchase so thanks for watching guys if you

00:08:02.479 --> 00:08:07.680
dislike this video you can hit that button but if you liked it hit like get

00:08:05.599 --> 00:08:11.199
subscribed maybe consider checking out where to buy the stuff we featured

00:08:10.080 --> 00:08:15.039
if you like work at a university or something

00:08:13.360 --> 00:08:20.720
it's like i'll know that's what happened if like you know 2 000 xeon 5's show up

00:08:18.960 --> 00:08:24.080
on like our amazon report or whatever anyway the point is also in the

00:08:22.240 --> 00:08:28.160
description is our merch store as well as our community forum which you should

00:08:25.440 --> 00:08:28.160
totally join