WEBVTT

00:00:00.080 --> 00:00:05.359
although transistors still keep on shrinking it's getting more and more

00:00:03.520 --> 00:00:09.760
difficult to pack as many of them on a chip as we'd like partial solutions to

00:00:07.680 --> 00:00:13.599
this such as using chiplets to reduce the rate of manufacturing defects and

00:00:11.759 --> 00:00:17.440
stacking transistors on top of each other have been in vogue for a while now

00:00:15.759 --> 00:00:22.320
but it might not be too surprising that some manufacturers have decided to

00:00:19.439 --> 00:00:26.640
simply make the chips themselves bigger when in doubt supersize

00:00:24.800 --> 00:00:30.320
now i'm not saying that your next computer might have a CPU that's so big

00:00:28.480 --> 00:00:33.280
it'll take up half the motherboard but when you get away from personal

00:00:31.599 --> 00:00:37.120
computers and start looking at chips that we might see in data centers in the

00:00:34.960 --> 00:00:40.399
near future you start seeing some pretty eye-watering stuff

00:00:38.960 --> 00:00:45.440
we're talking about designs like the wafer scale engine 2 from cerebros

00:00:42.960 --> 00:00:50.239
currently the largest chip in the world built on a seven nanometer process and

00:00:47.360 --> 00:00:57.039
contains 850 000 cores and is a whopping 21.5

00:00:54.000 --> 00:00:59.920
centimeters or 8.5 inches long that's

00:00:57.039 --> 00:01:05.840
more total area than 25 ryzen desktop cpus perhaps unsurprisingly a chip this

00:01:02.640 --> 00:01:08.720
big and with this many transistors 2.6

00:01:05.840 --> 00:01:14.320
trillion to be exact requires a lot of power the wafer scale engine 2 sucks

00:01:10.880 --> 00:01:16.960
down 15 kilowatts so if you were somehow

00:01:14.320 --> 00:01:22.320
able to drop this into your pc you'd need 15 1000 watt power supplies just to

00:01:20.320 --> 00:01:26.880
keep it fit that's not even counting the rest of the system but despite this the

00:01:24.880 --> 00:01:31.360
new design should actually result in power savings you see data centers and

00:01:29.680 --> 00:01:36.159
supercomputers that do artificial intelligence processing often have to

00:01:33.439 --> 00:01:40.320
use lots of separate chips such as gpus spread across a large facility having

00:01:38.640 --> 00:01:44.960
the same amount of computing power on just one physical chip is far more power

00:01:43.200 --> 00:01:49.840
efficient even if the power consumption rating of that chip is a lot higher than

00:01:47.360 --> 00:01:53.759
a typical GPU but there are other advantages to this approach besides just

00:01:51.840 --> 00:01:57.840
saving energy you might be wondering why we aren't simply just sticking a bunch

00:01:55.600 --> 00:02:02.399
of chiplets onto one package instead to make something like a really big version

00:02:00.000 --> 00:02:06.960
of an AMD epic processor and we'll tell you why right after we think ifixit for

00:02:04.880 --> 00:02:10.399
sponsoring this video ifixit wants to help you fix all of your devices so you

00:02:09.039 --> 00:02:14.879
never have to pay for a costly replacement again from your xbox to your

00:02:12.800 --> 00:02:19.200
toothbrush i fix it as parts and guides for almost any device you can think of

00:02:16.720 --> 00:02:24.239
they have over 70 000 step-by-step guides with photos to make it easy check

00:02:21.680 --> 00:02:27.599
out ifixit.com techwiki to pick up a repair kit and join the right to repair

00:02:26.160 --> 00:02:31.840
movement today so as versatile as chiplets have been

00:02:30.080 --> 00:02:36.319
they still suffer from having more latency than one big monolithic

00:02:34.319 --> 00:02:40.879
processor the little interconnects that move data between chiplets as quick as

00:02:38.160 --> 00:02:44.879
they may be and they are fast are still slower than if you physically

00:02:42.720 --> 00:02:48.879
put computing units directly adjacent to each other to form one big chip

00:02:47.200 --> 00:02:53.120
ultimately this means that huge monolithic chips can process more data

00:02:51.519 --> 00:02:57.440
than a system with the same number of transistors spread out among multiple

00:02:55.360 --> 00:03:01.920
chips and when you consider just how much data has to be processed for ai

00:02:59.680 --> 00:03:06.159
applications and scientific research it makes a difference wafer scale

00:03:04.560 --> 00:03:09.920
technology has already drawn interest from diverse industries including

00:03:07.840 --> 00:03:13.760
national intelligence and healthcare but though it has some obvious advantages

00:03:11.920 --> 00:03:18.480
that doesn't necessarily mean that it's the silver bullet to large-scale compute

00:03:15.920 --> 00:03:21.360
challenges for instance one big issue is the fact that these processors are

00:03:19.920 --> 00:03:26.080
designed to handle lots of data so they also need access to

00:03:24.000 --> 00:03:29.680
a lot of memory and designs with chiplets and larger amounts of memory

00:03:27.760 --> 00:03:33.599
built on the same package may end up being more popular this is similar to

00:03:31.760 --> 00:03:37.519
what tesla has done with their new d1 chip which was developed in part to help

00:03:35.599 --> 00:03:42.720
propel tesla's self-driving ai technology the d1 chip itself is much

00:03:40.640 --> 00:03:48.480
smaller than the wafer scale engine but tesla has included over 11 gigabytes of

00:03:45.519 --> 00:03:53.200
high speed sram in an arrangement of 25 d1 chips connected together to make a

00:03:50.879 --> 00:03:57.519
training tile that's bigger than your head and of course making smaller chips

00:03:55.920 --> 00:04:01.920
reduces the amount of silicon you'll waste due to manufacturing errors as we

00:03:59.680 --> 00:04:05.519
mentioned earlier but regardless of whether a particular company is using

00:04:03.599 --> 00:04:09.599
wafer scale or an arrangement more like teslas putting a large amount of silicon

00:04:07.760 --> 00:04:13.360
on one plane may end up becoming an industry trend because if america has

00:04:11.599 --> 00:04:18.799
taught us anything it's that there's a deep human need to supersize

00:04:16.320 --> 00:04:22.479
and if you feel the need like the video or dislike the video check out our other

00:04:20.720 --> 00:04:26.720
videos and comment below with video suggestions we make videos here get your

00:04:24.960 --> 00:04:29.840
videos here at Techquickie don't forget to subscribe and follow

00:04:30.000 --> 00:04:33.320
see you later