WEBVTT

00:00:00.480 --> 00:00:04.880
the year was 2018

00:00:03.760 --> 00:00:10.559
which actually now that i think about it wasn't really that long ago but things

00:00:08.000 --> 00:00:16.640
are moving really fast and it's already time for an update so earlier this year

00:00:13.679 --> 00:00:23.039
we did a video about the first ever open source CPU architecture risk 5 in

00:00:19.520 --> 00:00:25.519
collaboration with sci-5 the makers of

00:00:23.039 --> 00:00:31.679
the first ever commercial risk 5 hardware and today today we get to go

00:00:28.400 --> 00:00:33.920
deeper so sci-fi sponsored our trip down

00:00:31.679 --> 00:00:38.960
to their headquarters in san francisco to see some early concepts of real

00:00:36.880 --> 00:00:44.559
hardware products that are being built using their ip all the way from a normal

00:00:42.239 --> 00:00:50.480
SSD like what you'd install in your pc or laptop to a functioning media server

00:00:47.760 --> 00:00:55.680
and it might not look like much but buried under this mess of wires as

00:00:54.000 --> 00:01:03.680
there so often is darby treasure

00:01:06.799 --> 00:01:13.520
let's start with risk fives benefits in a nutshell one it's brand new which

00:01:11.439 --> 00:01:17.119
means that it sheds a lot of the legacy bloat that accompanies traditional

00:01:15.280 --> 00:01:23.600
processor architectures making it both extremely scalable and extremely power

00:01:20.000 --> 00:01:26.240
efficient and two it's open source which

00:01:23.600 --> 00:01:29.040
means no expensive licensing fees for the companies that use it in their

00:01:27.680 --> 00:01:32.079
products it always comes down to money doesn't it

00:01:31.280 --> 00:01:38.640
but it has some problems even if you had an

00:01:35.520 --> 00:01:41.360
open source CPU architecture and you

00:01:38.640 --> 00:01:45.439
knew how to build a CPU unless you're already vc backed out the butt and you

00:01:43.759 --> 00:01:50.640
can scrape together a minimum order quantity on the order of hundreds of

00:01:47.759 --> 00:01:55.600
thousands of chips a foundry like tsmc or global foundries is very unlikely to

00:01:53.600 --> 00:02:00.000
even pick up the phone that is where sci-fi's real plan for the

00:01:58.640 --> 00:02:04.000
future comes in so today if you

00:02:02.000 --> 00:02:07.439
yeah you like you right there need to build a custom chip for some reason

00:02:05.920 --> 00:02:12.160
let's say uh you want to build a microcontroller for a car or a smart

00:02:09.679 --> 00:02:17.040
thermostat there's a good chance that you would need to build an entire team

00:02:14.319 --> 00:02:22.239
of people that specialize in chip design and manufacturing looks expensive but

00:02:19.920 --> 00:02:26.560
check this out in the same way that large-scale computing has largely moved

00:02:24.879 --> 00:02:31.680
from servers in a closet under the stairs to the cloud where processing

00:02:29.200 --> 00:02:37.040
power storage and network speed can be ordered a la carte

00:02:33.440 --> 00:02:39.440
sci-5 has created the pizza ordering app

00:02:37.040 --> 00:02:43.360
of custom chip development so you jump onto their website select

00:02:42.160 --> 00:02:50.640
things like performance memory size the type of ports and

00:02:47.760 --> 00:02:55.680
interfaces that you want and then as you go it generates a block diagram for you

00:02:54.000 --> 00:03:00.640
in real time then you click build and it goes to a

00:02:58.319 --> 00:03:05.040
cloud instance that chugs away generating and verifying the processor

00:03:03.360 --> 00:03:10.800
that you defined then the next day you can download the

00:03:07.440 --> 00:03:12.879
verilog rtl and fpga images that you can

00:03:10.800 --> 00:03:16.080
then program onto a board and you're ready to rock

00:03:14.159 --> 00:03:21.920
it's basically self-serve this is in stark contrast to

00:03:19.440 --> 00:03:26.480
working with a traditional ip provider where you might have to sign an nda and

00:03:24.000 --> 00:03:31.519
hand over some fat stacks before getting anywhere close to actually testing your

00:03:28.879 --> 00:03:36.640
software on your custom chip now right now sci-fi core designer only

00:03:34.400 --> 00:03:39.920
works for the CPU but in the future they'll integrate third-party

00:03:38.000 --> 00:03:44.720
intellectual property like graphics controllers and allow customers to build

00:03:42.319 --> 00:03:49.120
an entire soc through their web interface and then if they want to take

00:03:46.799 --> 00:03:54.959
it a step further they can even have the chips fabbed and delivered through

00:03:51.519 --> 00:03:57.360
scifi's partnership with tsmc

00:03:54.959 --> 00:04:02.239
so the demo room then finally it starts with the sci-fi

00:03:59.720 --> 00:04:06.319
fu540 the same computer that was previously running quake in our office

00:04:04.720 --> 00:04:11.040
currently it's actually doing something a little different it's playing a

00:04:07.760 --> 00:04:14.319
youtube video here which might not seem

00:04:11.040 --> 00:04:16.880
that impressive but this is more of a

00:04:14.319 --> 00:04:22.479
software compatibility demo so the operating system that's running here is

00:04:19.040 --> 00:04:24.960
debian Linux with no risk 5 special

00:04:22.479 --> 00:04:29.919
expertise required in order to use it now it's the daily update stream so you

00:04:26.800 --> 00:04:32.479
can expect frequent updates but

00:04:29.919 --> 00:04:37.759
if you wanted to install some random application let's say a firefox you just

00:04:35.199 --> 00:04:41.040
app get install to be clear

00:04:38.880 --> 00:04:46.720
no one at sci-5 is trying to convince average consumers to run out and buy one

00:04:43.600 --> 00:04:48.880
of these boards and run it at home

00:04:46.720 --> 00:04:52.160
outside of software from the package manager

00:04:50.080 --> 00:04:57.199
very little will run at this time don't expect to download steam and start

00:04:54.639 --> 00:05:02.880
gaming or whatever but the message here is that compatibility is improving about

00:05:00.320 --> 00:05:08.080
94 of the packages in the debian repository support risk 5 and other

00:05:05.919 --> 00:05:13.840
flavors of Linux are working as well including fedora open wrt and open

00:05:11.199 --> 00:05:19.039
embedded and performance is improving too so web browsing

00:05:16.080 --> 00:05:21.260
yeah it's actually super slow right now like let's go ahead and go to our

00:05:20.320 --> 00:05:23.360
website

00:05:23.360 --> 00:05:30.000
oh boy but the problem here

00:05:26.880 --> 00:05:32.240
is that the engine behind the browser

00:05:30.000 --> 00:05:36.880
doesn't have a javascript just in time compiler so it's kind of like having a

00:05:33.759 --> 00:05:39.280
10 year old engine on a brand new car

00:05:36.880 --> 00:05:44.400
with more optimization that should get as much as 10 times faster about

00:05:42.160 --> 00:05:50.560
equivalent to an entry-level quad-core army 53 which as some of you probably

00:05:47.039 --> 00:05:52.160
realize still isn't an overabundance of

00:05:50.560 --> 00:05:57.919
performance if only there was some way to add

00:05:54.320 --> 00:06:00.240
co-processing capability to it

00:05:57.919 --> 00:06:06.080
oh wait there is so this demo right here comes courtesy

00:06:02.720 --> 00:06:08.479
of microsemi a microchip company these

00:06:06.080 --> 00:06:15.280
guys build field programmable gate arrays and fpgas are expensive but these

00:06:12.639 --> 00:06:20.639
things are really cool so basically they're hardware chips that you can

00:06:17.600 --> 00:06:23.199
program to offload certain workloads to

00:06:20.639 --> 00:06:28.240
hardware rather than software allowing your device to perform a specific task

00:06:25.680 --> 00:06:33.039
in this case computer vision really really quickly

00:06:30.560 --> 00:06:38.639
so their plan is actually to take this entire thing here and turn it into a

00:06:35.840 --> 00:06:44.560
single soc that they're calling polar fire marrying risk fives real-time Linux

00:06:41.840 --> 00:06:49.039
capabilities to their programmability with full cache and memory coherency i'm

00:06:47.360 --> 00:06:52.960
going to challenge this thing let's try a chair

00:06:50.639 --> 00:06:56.639
what do you think can you do a chair oh airplane

00:06:54.800 --> 00:07:02.160
wait ah there it is not had it for a second chair

00:06:59.520 --> 00:07:08.240
it's still early still early days but hey person scott person confirmed here

00:07:05.199 --> 00:07:11.199
first Linus is a real person

00:07:08.240 --> 00:07:15.520
and it gets even more modular the risk five foundation includes some really

00:07:13.520 --> 00:07:22.240
influential members these days including NVIDIA who has one of their nvdla deep learning

00:07:19.280 --> 00:07:28.240
accelerators running the yolo you only look once algorithm for object detection

00:07:26.000 --> 00:07:33.360
so in a similar fashion to the last demo we saw it pulls images from the webcam

00:07:30.960 --> 00:07:37.280
here pushes them into the nvdla's buffers where the object is detected

00:07:35.599 --> 00:07:43.440
then it displays the results on the monitor it's just a more powerful

00:07:40.000 --> 00:07:45.759
example of fundamentally the same idea

00:07:43.440 --> 00:07:52.240
let's see if it picks up the phone yeah look at that and the scalability goes

00:07:48.319 --> 00:07:54.560
down as well so this is sci-fi's fe310

00:07:52.240 --> 00:07:59.360
on a high five one board it's an embedded board that is physically

00:07:56.560 --> 00:08:03.360
compatible with arduino but with about 10 times the performance

00:08:01.199 --> 00:08:07.759
so the demo we're looking at here doesn't look like much but

00:08:05.759 --> 00:08:12.720
what we're seeing is that it can work on a computational task in the background

00:08:10.720 --> 00:08:17.120
and a real-time one which is the blinking of these leds right here so

00:08:15.039 --> 00:08:22.400
there is a performance penalty to our led going off exactly on time

00:08:20.400 --> 00:08:27.680
interrupting whatever else is going on but for some applications like medical

00:08:25.120 --> 00:08:31.120
for example key tasks need to be performed

00:08:28.720 --> 00:08:35.839
right now and actually this design across the table from upbeat is

00:08:33.680 --> 00:08:41.200
targeting the chinese fitness wearable industry and is expected to show up in

00:08:38.240 --> 00:08:47.600
future devices from huami it integrates a similar e3 series core but with extra

00:08:44.880 --> 00:08:52.000
ip including a cnn or convolutional neural network and a graphics

00:08:49.680 --> 00:08:54.560
accelerator it looks really big but this is just development stuff it's right

00:08:53.279 --> 00:08:59.839
there cool huh now this next station is a

00:08:57.040 --> 00:09:04.320
little more relatable for pc enthusiasts we talked at considerable length

00:09:01.360 --> 00:09:09.839
recently about the complexity of pushing SSD performance up when nand flash

00:09:07.360 --> 00:09:15.680
performance has gone down in recent years it requires very high speed

00:09:12.720 --> 00:09:21.040
controller chips and scifi's partner fadu is actually working on what they

00:09:18.080 --> 00:09:28.080
hope will be the fastest consumer SSD on the market with a planned ship date of

00:09:23.120 --> 00:09:32.080
q3 2019 so this here is a fadoo asic

00:09:28.080 --> 00:09:34.480
with three sci-5 e51 core ips so that's

00:09:32.080 --> 00:09:39.040
their 64-bit high performance embedded core and those are driving the SSD

00:09:36.880 --> 00:09:44.720
controller algorithms that do all of the page mapping and whatnot and fadu claims

00:09:42.240 --> 00:09:50.720
that the sci-5 cores were one-third of the power and area of competing designs

00:09:48.320 --> 00:09:56.560
now we couldn't plug it in to verify any of this it's still very early stages but

00:09:53.600 --> 00:10:01.839
here's something we were able to plug in in years past

00:09:58.320 --> 00:10:05.200
this home media server or nas device

00:10:01.839 --> 00:10:07.279
from wd would have had an soc based on

00:10:05.200 --> 00:10:10.800
licensed ARM intellectual property hooked up to its shingled magnetic

00:10:08.880 --> 00:10:14.720
recording hard drive and then handling streaming media over your network to a

00:10:12.640 --> 00:10:18.720
device like this laptop well

00:10:16.720 --> 00:10:23.600
not today now they're really far away from talking

00:10:20.880 --> 00:10:28.240
about performance at this point but the demo that you're seeing is running on

00:10:25.519 --> 00:10:33.279
real risk five silicon with the cost savings that come with it and they're

00:10:30.720 --> 00:10:37.680
hopeful that on top of a cost savings thanks to a greater degree of control

00:10:35.519 --> 00:10:42.320
over the hardware they could create custom instructions that improve the

00:10:39.360 --> 00:10:47.519
data path increasing performance dang leaving us with just a couple of

00:10:44.720 --> 00:10:52.880
housekeeping items here so one sci-five had not one but three debugging tool

00:10:50.399 --> 00:10:58.959
partners demoing their wares including iar sager and lauterbach and had a

00:10:56.560 --> 00:11:03.920
couple really cool security demos this secure boot demo checks for a properly

00:11:01.680 --> 00:11:08.160
signed Linux image and if everything's fine it boots normally but if

00:11:06.240 --> 00:11:11.279
something's amiss this light goes off

00:11:11.760 --> 00:11:15.320
wait for it

00:11:17.360 --> 00:11:23.680
there it is bad evil Linux

00:11:21.200 --> 00:11:28.240
cannot boot Linux authentication failed now obviously this is

00:11:26.399 --> 00:11:32.720
not how it'll actually work in the real world but it's it's very cute

00:11:30.880 --> 00:11:39.600
and then over here we have the hex 5 multi-zone demo

00:11:36.320 --> 00:11:41.040
so the sci-5 processor is running this

00:11:39.600 --> 00:11:48.320
motor control a console and a real-time program that's

00:11:44.880 --> 00:11:50.320
making this led blink and each of them

00:11:48.320 --> 00:11:56.399
is in its own bucket so the idea here is that if the

00:11:53.600 --> 00:12:02.800
led blinker were to get attacked it can't turn around and in turn attack

00:11:59.440 --> 00:12:02.800
your motor control

00:12:03.519 --> 00:12:08.959
which basically concludes our portion of

00:12:06.800 --> 00:12:13.120
today's exercise but you guys still have a homework assignment

00:12:10.399 --> 00:12:18.079
this is pretty cool if you've ever even thought that chip design is kind of cool

00:12:15.440 --> 00:12:23.040
go to sci-fi site and fire up their core designer just give it a try i think you

00:12:20.639 --> 00:12:27.120
guys might be impressed at how cool it is even just as

00:12:24.639 --> 00:12:31.600
like kind of a fun learning tool to look at what components there are to a CPU if

00:12:30.160 --> 00:12:36.560
you've never really given it any thought anymore we're going to have that linked in the video description

00:12:34.720 --> 00:12:40.160
so thanks to sci-fi for sponsoring this video thanks to you guys for watching it

00:12:38.639 --> 00:12:43.440
if you guys disliked it you can hit that button but if you liked it hit like get

00:12:42.000 --> 00:12:48.240
subscribed or maybe consider checking out where to buy the stuff we featured at the link below also down there is our

00:12:46.800 --> 00:12:53.839
merch store which has cool shirts like this one and our community forum which

00:12:50.399 --> 00:12:53.839
you should totally join