EP#437: Embarrassingly Parallel Computing - Steve Furber - Circuit Break - A MacroFab Podcast

Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Parker Dillmann (00:10):
Welcome to circuit break from MacroFab. A
weekly show about all thingsengineering, DIY projects,
manufacturing, industry news andARM Processors. We're your
hosts, electrical engineers,Parker Dillmann and Stephen
Kraig. This is episode 437.

Stephen Kraig (00:28):
This week, our guest is Steve Ferber. Steve is
professor emeritus of computerengineering at the University of
Manchester in the UK. He is bestknown for his work on the BBC
microcomputer and the first ARMprocessor at Acorn Computers in
the 19 eighties, and morerecently for the Spinnaker
project that has delivered amachine with a 1000000 ARM

(00:50):
Processors for brain modelingapplications.

So thank you so much, Steve, for joining us on
the podcast this week.

Steve Furber (00:59):
It's a pleasure.

So before we, like, completely jump into it,
Steve, what what like, how howdid you get started, like, just
I mean, there there's a lot ofstuff there to unpack of of your
history of starting with the BBCmicrocomputer and arm and
beginning with arm processors.Because like nowadays, there's
an ARM Processor in practicallyalmost every single device

(01:25):
nowadays. Let's let's wind theclock all the way back and talk
about, like, your early careerand how did you like get started
with like the BBC micro,computer? Oh, what actually
actually, I bet you most of ourlisteners don't even know what
the BBC microcomputer isactually as well.

So my story really starts at Cambridge University
in the UK. I went to Cambridge,to study mathematics as an
undergraduate and in in in wasthere for 4 years, studying
mathematics. And then because Ihad a long interest in flying, I
got the opportunity to take aPhD in aerodynamics in the

(02:06):
engineering department. So Ibecame a very sort of
theoretical engineer. It wasstill mainly math, so a bit of
experimentation.
And my interest in flying sortof manifest in terms of I did
some model flying in my earlieryears, and I joined the
university glider club inSouthern Airfield every
Wednesday afternoon for a year,amassing a total of about 54

(02:29):
flying minutes over that year.And through all of that, III
decided that that actually mightbe simpler to get involved in in
in, flight simulation. Thatmight be a more productive way
of of doing this flying thing.They're actually trying to get
near real airplanes, which seemto occupy a lot of time and or
money. And around the time I wasthinking about this, I I got to

(02:54):
hear about a bunch of studentswho were starting a new student
society called the CambridgeUniversity Processor Group.
And this was a student who builtcomputers for fun. And we're
we're talking here in lateseventies. And I thought, if I'm
interested in building some kindof flight simulator, then
knowing a bit about computers isan obvious place to start. So I

(03:16):
got involved in the process togroup. I started building
computers, which in those daysinvolved very scary things like
ordering microprocessors fromshops in California using credit
cards, which for imprecuniousstudents were very frightening
things to do.
And I I bought, a 2650. Anybodyremember the Signetix 2650? That

(03:38):
was an 8 bit microprocessor.Ordered some memory chips, and
started putting these thingstogether to form a computer. Now
I such should say that in in theCambridge University Processor
Group, the real men built theircomputers out of TTL, and it's
only the wimps like me who whowent to this newfangled
microprocessor stuff.

So so the p the wimps were using integrated
circuits. I guess, moreintegrated versus TTL. But, no,
I I've never heard of theSyngetix 2650 I'm looking at
right now.

It didn't it didn't stay around long. Anyway,
so I so I got involved in thestudent society and and and
built machines that startedworking and, got on with my
engineering, PhD and and then Ihad a research fellowship
sponsored by Rolls Royce afterthat. And while I was doing

(04:33):
that, I was approached by HermanHouser, who told me he was
thinking of starting aconsultancy in the
microprocessor business and waslooking for a few willing people
to do odd jobs for him. Andthat's how I got drawn into what
later became Acorn Computers.When I got involved, it was very

(04:55):
embryonic, and it was calledCambridge Processor Unit Limited
or CPU Limited.
And I remember the first job wedid was developing, a
microprocessor based controllerfor a fruit machine, which I
think in the US, you call a 1armed bandit. Yeah? Is that the
right terminology? Up until thelate seventies, these were
largely electromechanicaldevices, and they were just

(05:19):
beginning to think aboutswitching them to microprocessor
control. And and and wedeveloped this control system.
Then 1 of the other guys thatwas involved in that started
thinking about designing ahobbyist microprocessor kit. And
in the UK, Clive Sinclair andand and Chris Curry, who was

(05:40):
involved with Herman, used thisNational Semiconductor thing
called the MK 14, which theysold as a hobby's kit. And and
Sophie Wilson looked at this andsaid, I can do better than that.
And and and produce this singleboard. You know, this the kind
of computers that were aroundthen, which were 8 bit micros
with hexadecimal displays andhexadecimal keypads.

(06:02):
And, usually, you bought theseas a kit, and you'd assembled
them yourselves and put themtogether. And and and that thing
that Sophie designed was theAcorns became the Acorn System
1. So Acorn was originally atrading name, not rather than
the company name. Only later didthe company rename itself to
follow that trading name. Sothat's how I got drawn in

(06:24):
through through this fruitmachine controller development
and and and building stuff forfun.
I was still primarily working atthe university in my research
fellowship. But I'd I'd I'dstarted building computers
firstly as a hobby, and then Iwas using these computers in my
aerodynamics research to tomonitor these these fluid flow

(06:46):
systems experiments that I wasbuilding. And it was a kind of
deal with Acorn that I designedstuff for them. They give me
bits to play with. I'd use thosebits to design more stuff and
and and to give the designs tothem.
It was a very informalrelationship between us. Then
Acorn moved on from the system1, and it built these rack based

(07:10):
systems that were really fairlyprofessional microprocessor
controllers, with floppy diskdrives and things, and they they
were very expensive. And ChrisCurry decided they wanted
something cheaper to get it backinto the hobbyist market, and
they developed the Acorn Atom.And there are lots of stories I
could tell you about the Atom. Idid not have a hand in

(07:30):
developing the Atom.
But, it was sold as a kit, andwe had all sorts of rejects sent
back by people who built thesekits. 1 said, you know, I
understand that solder is hotand heat is bad for integrated
circuits. So I very clear if heglued all these chips into
circuit board, and it stilldidn't work. That was the the

(07:51):
last product Acorn sold as akid. The the Atom, became fairly
well known in the UK.
And, in fact, it was responsiblefor, some leading technology. 1
thing that came out of theCambridge Computer Lab was a
local area network technologythat was built into the atom
called Ethernet, so you couldnetwork classrooms of of atoms

(08:13):
together. But then we heardabout the BBC's interest in in
developing, in having a computeraround which they could base a
series of TV programs. And theyopened up the bidding for this.
And and Acorn put in a proposal,which was basically based on the
machine I'd built the yearbefore at home for monitoring my

(08:36):
fluid dynamics experiments.
And we turned this into acircuit and and and into what
also it be became the BBC Microbecause Acorn got the contract.
And there was a famous week whenthe beginning of the week, we
had nothing. And at the end ofthe week, the BBC were coming
going to come and see what wehad, and we put the prototype
together. And and and by theFriday, when the BBC arrived,

(09:00):
indeed, the hardware was justworking, and the software was
running on it. And, they weresuitably impressed by Acorn's
ability to move things forwardvery quickly that Acorn got the
contract.

Wait. Wait. You're saying you developed a
computer within a week, or Ishould say built a computer
within a week?

I as I say, I we have the concept at the
beginning of the week. I builtthis machine at home the
previous year, which embodiedquite a lot of the principles in
this concept. But we haven'tstarted building the actual
machine, which was, if you like,based on the concept, but
running twice as fast. And andso we did we did the prototype
development from scratch, youknow, wire wrapping the circuit

(09:42):
board. The first the first stagewas actually drawing the circuit
diagram on a piece of paper,which I did.
And then wire wrapping theprototype, and then debugging
it, and then getting thesoftware up and running. Yeah.
That took a week.

So this is oh, no. Let's back up a little bit.
So a fruit machine, by the way,is is we would call that a slot
machine in the United States soit's a it's a gambling machine
you said 1 armed bandit and I'mlike I think that's a slot
machine

yeah okay

and I want to talk about so you built a
computer at home this is theyear prior to run your fluid
dynamics experiments. You gottatalk about that.

Yeah. To say I built it to run those
experiments will probably beslightly misleading, and I built
it for fun, because I likedbuilding digital electronic
circuits. And having built it, Ithen used it, to basically do
data logging and and simplethings, linked to my fluid

(10:45):
dynamics experiments. And in theengineering department, we had
we had computers. I think Ithink the small 1 we had was a
computer automation LSI 4.
If that means anything toanybody, I'm not sure these
days. But it was a fairly bigand heavy and immobile thing. So
having, you know, home builtrack of microprocessor based

(11:06):
boards that that I've developedwas just a more convenient way
of interfacing into thesesensors in the experiments.
These were doing things likemeasuring pressure at different
points in in an axial flowcompressor stage.

You know, actually so I've I've I've 1
question that's gonna rewind theclock even a little bit more.
You had mentioned your yourschooling. You you originally
went for to school for maths andthen got your PhD in
AeroAeronautical, was it? Yeah.Aeronautical.
What what I'm curious about ishad was was electronics involved

(11:44):
at all in that? Or did you getall of your electronics training
from the processor group atCambridge?

I I pretty much got all the training from the
processor group. There was therewas a little bit of computing in
the, in the maths degree that Itook. I think that was that was
using another computer nobody'sever heard of called a modular 1
computer. I think running alanguage called Focal. Does that
sound right?
I don't know. I do remember thatit it used Hewlett Packard

(12:13):
storage tubes. Right? So thecomputer were right on the
display. And then when youwanted, you had to raise the
whole display and start again.
So they were basically a astorage oscilloscope technology.
But yeah. So so all all mydigital electronics, I learned
through building stuff in theprocessor group. And, you know,

(12:36):
friends in the processor groupwho helped debug debug things.
There were people there who knewa lot more than I did about
putting chips together to buildsystems.

Oh, that's fascinating.

The the technology we used then was was was a
wiring pen. So you basically gotthese preformed circuit boards,
and you put chip sockets intothe circuit boards and, you
know, soldered them at thecorners to hold them in place.
But then you made theconnections with a wiring pen
and wrapping wires around combsand then around pins. And when
you soldered them, theinsulation on the wire melted

(13:08):
and you had a connection. Right.
Right.

Which which makes for a for a good a good circuit
connection, but it must takeforever to build a a sizable
computer.

Yeah. I mean, you're you're you're making
connections 1 at a time. I mean,it's it's like wire wrapping.

Right.

But quite a lot cheaper to implement, and and
you don't end up with a boardthat's got, you know, 2 inch
long pins sticking out the back.But it's a it's a more compact
format. But it's debuggable andrepairable, and you can take
wires off and Right.

It's it's rapidly prototypable.

Yeah. Yeah. It's a rapid prototyping system,
basically.

So what were what were the compared to, like,
nowadays, what were thechallenges back then when this
kind of technology was juststarting to kick off? What was
different now then than now?

I think 1 1 major difference was that the the
speeds that were being used inthose boards were in the
megahertz region. And and thethere's some manual wiring and
running lots of wires throughthe same combs, so very close to
each other. That all worked atthe megahertz range, and it
clearly would not work with thekind of gigahertz frequencies

(14:24):
that we see today. Today'stechnology is much more
demanding in terms of signalintegrity running around the
system. In those days, you youreally could I mean, you could
wire chips together in patchboards and expect them to work
at the speeds that were beingused then.
You wouldn't stand a chancethese days of of doing that kind

(14:45):
of thing. What was challengingthen was was coping with some of
the higher frequency stuff. Theclock I used in my system, I
think, was 16 megahertz. Andgetting that crystal oscillator
to work reliably required a fairamount of help from my friends,
because we didn't have thetools. We, you know, we didn't
have oscilloscopes and thingsthat you need to to really get

(15:06):
crystal oscillators workingproperly.

Yeah. You just kinda had to bank on people who
have done it before and had thatexperience, because you couldn't
see what you were doing really.

That's right. Yeah. Yeah. I mean, the may the
major debugging tool in thosedays was was a logic probe. I
don't know if you've come acrossthose, but a logic probe, you
basically have a couple of clipsthat you clip across the 5 volt
rail, because everything was 5volts.
And then the thing had a pin atthe front and a green LED and a

(15:41):
red LED, and the LED would lightup to tell you if the signal was
high or low. And and, you know,so you tended to build circuits
that you could run and thenstop. When you stop them, you
could then go measure all thesignal levels and see if they
were right.

Yeah. Imagine trying to see a LED blink at 16
megahertz. Probably wasn'tworking.

Yeah. As I say, you you needed to get the thing
into a static condition in orderto sensibly use the logic probe.

0III can't imagine trying to because, like,
nowadays, you use with, like, adigital logic analyzer, which
but can record all that overtime, and then you can go back
and look at the data. Becauseyou're you're talking about,
basically, that, but you'd haveto run per clock cycle, and then
read all your data bus, andwrite that all down and then,

(16:33):
you know, iterate the next cycleand then measure all that again.

Yeah. I I remember 1 occasion where III coded a
sort of simple screen operatingsystem and and put it in a
apron, And and there was a bug,and it was running. You entered
the first command, press return,and then something went wrong.
Debugging that was excitingbecause after you'd entered the
command and push return, whatyour brain system did was it

(17:02):
scrolled the screen. Okay.
And scrolling the screen isbasically copying every byte
representing a pixel on thescreen up 1 location. And then
you had to sit there with asingle step button, pushing this
single step button. I think ittook me about 45 minutes just
pushing this button as fast as Icould to get through the screen
scroll before it got to thepoint that I was interested in

(17:24):
where the bug came up.

That is fun.

So after the, you know, working on the BBC
micro, you started working onArm. So how did that transition
happen?

The the story there has has been told a number
of times. I mean, the the BBCMicro was a huge success for
Acorn. And in the earlynegotiations with the with the
BBC predicted, you know, salesof about 12, 000 BBC Micros. And
and that turned from 12, 000into 1 and a half 1000000, in in

(18:02):
2 years. So everybody's estimateof of of the degree of interest
in home computing was completelywrong.

So, actually, so from 12, 000 so, like, BBC
is, like, we're gonna build 12,000 of these over 2 years to you
said 1 point how many million?

1, 100, 000.

How was trying to get get those, like, I can't
imagine, like, the logisticaljump to make that work Cause
you're you're planning 12, like,yo, you're gonna get 12, 000
microcontrollers. Oh, now weactually need 1 point, you know,
2, 000, 000.

Yeah. It was it was an interesting challenge to
Acorn's manufacturing people. Imean, all the manufacturers
subcontracted. Acorn itselfdidn't didn't build things. So
what it had to do was just bringin more subcontractors to to
have more assembly linesproducing the product.
But the BBC Micro was a bigsuccess in in that sense. And
that sort of launched Acorn intoquite a prominent position,

(19:03):
particularly in the UK and someother regions. So they're quite
successful in in Australia, NewZealand, Netherlands, Germany.
The attempt to sell the BBCMicro in in the US, was not a
big success. In fact, it wasrather an expensive failure.
That's a different story. As faras the the development team was

(19:24):
concerned, at that point, theworld was moving from 8 bit
microprocessors to 16 bitmicroprocessors. And and we at
Acorn were looking at all the 16bit microprocessors built in
mechanism built in mechanism tosupport a second processor. So

(19:46):
it's very easy to hang, any anymicroprocessor you like to build
a little system with themicroprocessor and some memory
hanging into the 2nd processorport and and then see what it
would do. And we did that with alot of microprocessors, most of
the ones that were around at thetime.
And we didn't like any of them.And there were 2 main reasons we

(20:09):
didn't like them. 1 was we usedextensive real time, performance
in the BBC micro. 6502 has quitegood real time response, And all
of these 16 bit microprocessorshad much worse real time
response. I mean, the the figureI remember is the National
Semiconductor 32016 had a memorymemory divide instruction, which

(20:33):
took 360 clock cycles tocomplete, and and the clock was
6 megahertz.
So you're talking 60microseconds. And during that
time, it was not interruptible.So if you were trying to handle
a double density floppy diskstream data stream, which
generates a byte every 32microseconds, you couldn't

(20:54):
without paying for morehardware. The real time response
was the first thing. And thesecond thing is that we
determined by then that the mainthing that that affected how
fast a processor went was thethe the processor's ability to
use memory bandwidth.
And the memory was the expensivecomponent in any, small

(21:17):
microprocessor system. If youpaid for the memory, what you
really wanted was a processorthat would use all the memory
bandwidth. And none of theprocessors that were available
from the big semiconductormanufacturers at the time would
do this. This was partlybecause, a lot of them had
architectures inspired by the 19seventies mini computers. So

(21:39):
they had very complexinstruction sets and they just
couldn't keep up with thememory.
So we were sort of frustrated bythese 2 factors when our boss at
ACORN, Hermann Hauser, starteddropping papers on our desk
about the, the risk work atBerkeley and Stanford, and how a

(21:59):
postgraduate class had designeda microprocessor that was very
competitive in in a year. Andand that got us thinking about,
you know, maybe if you could doit with a postgraduate class, we
could we could design our ownmicroprocessor. So that's where
the Arm idea came from. Wevisited National Semiconductor's
design facility in Haifa inIsrael, where they were

(22:23):
developing the first 2 0 16, andthey were on revg orh. I can't
remember which.
And they still haven't got thebugs out. And, again, this was a
result of the complexity of theinstruction set. So this this
new idea from Berkeley of ofreduced instruction set
computing, really simplifyingdown the instruction set to its
basics, and using the siliconresource for stuff that gave you

(22:47):
more performance than complexinstructions, such as pipelining
the processor. That kind ofappealed to us. And, Sophie
Wilson started playing with setarchitecture, and and and then I
took that and turned that into amicroarchitecture.
We gave it to our very smallVLSI design team to put

(23:09):
together. And 18 months later,from from the start, much to our
surprise, we had a processorthat worked really well. None of
us had any experience inprocessor design before that. We
were just you know, we beforethe armed, we all thought
processor design had a kind ofmystique to it, and only these
big semiconductor houses coulddo it. At the end of the armed

(23:30):
development, we realized it'sit's just another piece of logic
like everything else.
And and and if you go about itin a sensible way, you can do it
with a with a very smallresource. You could then. It it
was a particular sort of peak ofsimplicity, I think. The ARM
used fewer transistors than someof the 8 bit microprocessors

(23:51):
around at the time, and wepushed very hard to keep it
simple.

Did I hear that right? You said the first
processor that came backfunctioned and and worked well?
Yeah. Wow. Were there any bugsat all?

There was 1 very minor bug, in an obscure corner
of the barrel shifter operation.So 1 of the barrel shifter
functions didn't work asdesigned. A piece of metal there
had ended up in the wrong place.But apart from that, the whole
process had worked exactly as wehad expected, in in including

(24:28):
performance. You know?
It it what I said earlier aboutmemory bandwidth, that was the
reason why Acorn went straightfrom 8 to 32 bits. Okay? Because
the 32 bit bus gives you twicethe bandwidth of the 16 bit bus
if you can use it. So all was 32bits. And but it was very much
inspired by the risk work.

(24:49):
You know, we didn't implementthe big register windows
function that on the Berkeleyrisk machines because we thought
it was too expensive for ourapplication domain. But we did
take most of the other riskideas, the load store
architecture, you know, the regthe regular, largest register
file, separation of of loadstore instructions from data

(25:12):
processing instructions, all ofthat we, we we took from the the
the risk philosophy. With a fewoptimizations because, you know,
we kinda thought the riskexamples were academic
prototypes. We were a commercialcompany, so we wanted things a
bit tighter, a bit less slack inthe code density. And Sophie

(25:34):
Wilson, who who did the ISAwork, of course, had written all
Acorns BBC BASIC interpreters,so she had a very good
understanding of what wasrequired to support a high level
language.
So the arm had some somefeatures which which you might
not think of as very risk, suchas the load store multiple

(25:54):
instructions, which would loador or store the any subset of
the register set in the singleinstruction. And and and, you
know, that was clearly usefulfor procedure entry and exit and
stuff like that.

So what was the first high level language that
was was for the compiler forthis microcontroller? What was
it c, or is it was it beforethen?

I think the first the very first high ish level
language was BASIC.

Okay. It was BASIC?

I mean, the BASIC interpreter was kinda written
before we got the silicon,because we did we did have
emulators. So we built softwareemulators so we could develop
code and debug it before we hadsilicon. I think the first
compiled language probably wasc. There were various people got

(26:46):
very interested in importingtheir favorite language to the
ARM when we we had the earlyprototype systems available.

What was the first commercially available arm
after you've done thisprototype?

It depends what you mean by commercially
available. Right? We did sellthe old development system based
on that very first Arm silicon,but that was sold in, I guess,
tens or or may maybe the orderof 100 units. That was the
second process. This is the BBCMicro.
So you you got a box that youcould plug into the BBC Micro,

(27:22):
and that box had a longprocessor and and 1, or I think
more typically 4 megabytes ofmemory in it. So the first 1
that was sold in in significantvolume was was arm 2, which
process shrunk. So arm 1 was 3micron CMOS. Arm 2 was 2 micron
CMOS. And arm 2 was theprocessor around which the first

(27:47):
Acorn Archimedes products weredeveloped.
And the Archimedes sold in innot huge, but significant 50,
000 units a year or somethinglike that. So so it's it's not
huge numbers, but but that thosewere the first volume sales. And
the arm 2 was also licensed toVLSI Technology, who who were

(28:09):
the silicon manufacturer, andthey made it available to to
third parties. And, you know,Radius, I think, built a
graphics accelerator for theApple 2 that was based around
the arm 2.

Yeah. The other in my computer architecture
class, we actually built amicrocontroller that the
peripherals for amicrocontroller with an arm 2
core. Yeah. So we did all, like,the VLSI design around that arm
2 core. So that's kinda cool.

Yeah. I mean, was the arm 2 core a hardcore? Or or
or or or or was that also Ithink it was synthesized on to
FPGA?

It was synthesized.

Sorry. It was?

It it was synthesized.

Synthesized. Right. Okay. Yeah.

It was all on a on a Spartan FPGA, I think, if I
recall. Okay. Yeah. So arm ARMstand used to stand for, Acorn
risk machine. When when did theychange it to, advanced risk
machines?
When was when did that happen?

So that happened in 1990 when this this was
actually must have been veryshortly, a week or 2 after I
left Acorn to take up myposition at Manchester. Apple
came knocking on Acorn's doorsaying that, they quite like to

(29:37):
use the arm in the Newtonproducts that they were
developing, but they'd be morecomfortable if it wasn't owned
by a competitor. So how aboutsetting up the Arm activity as a
joint venture? And they werepushing on an open door because
Acorn was quite keen to shed thecost of supporting ARM

(29:57):
development. Its business wasn'tgrowing fast enough to really
fund that.
And Advanced Risk MachinesLimited was set up as a joint
venture between Acorn, Apple,and VLSI Technology had, I
think, had a minor role in that,but but but they were the 3rd
founding company. And it was atthat point when Arm was changed

(30:19):
from Acorn Risk Machine toAdvanced Risk Machine.

How'd you feel when you heard that?

III thought that was fine. I think I think they,
know, they obviously couldn'tkeep the the Acorn name if they
were gonna move us into aseparate company.

I guess it's it's different if it was called,
like, the Steve risk computer orsomething like that.

Do you

think why do you think Arm is everywhere now
nowadays?

So People ask about the success of Arm. And,
you know, as with all of thesethings, there's quite a lot of
serendipity comes into thestory. I think Arm was fortunate
in that, it was established as aseparate company at the
beginning of the nineties, justwhen people were beginning to

(31:09):
talk about systems on chip. Sothrough the eighties, a
microprocessor was a chip. Okay?
And then in the nineties, thenumber of transistors that
Moore's Law was delivering wasbeginning to get to the point
where you could put amicroprocessor and, quite a lot
of the rest of the system on thesame chip. And because ARM was

(31:30):
small and simple, it occupiedless of the real estate of an
early SoC than did itscompetitors. So there's more
room for the other stuff. And Ithink that made a difference.
Also, Arm turned out remarkablylow power.
Again, that can be attributed toits simplicity. But if you're

(31:53):
building an SoC, then you want aprocessor that doesn't burn too
much of the power budget,because, again, you've got more
power budget for everythingelse. And and and so I think the
the the the those technicalaspects were significant in the
very early days, and thatenabled ARM to get the Nokia

(32:13):
business. Now Nokia in thenineties was the biggest mobile
phone handset manufacturer inthe world, and they were
basically setting the pace forGSM, which is the early digital
mobile phones. And the fact thatArm got into Nokia meant they
got into nearly everything elseas well, in the mobile phone

(32:35):
area.
And and that was huge. But also,a factor in Arm's success has
been the business model thatRobin Saxby developed when he
came in as CEO of Arm, when itwas set up as Advanced Risk
Machines. And and and that isthis IP licensing model. And and

(32:58):
and that basically means thatthat that ARM is available to
anybody, you know, at at atcost. But if if you base that on
a royalty only basis, thenroyalties are are quite tough
for cash flow, because the moneycomes little and late with

(33:19):
royalties.
So Saxe, we developed this modelwhere you pay an upfront entry
fee to get your license, andthen you pay downstream
royalties as well. And, ofcourse, the upfront license fee
is very good for cash flow. It'sit's big, and it's early. And
that kept on going and buildingwithout any external financial

(33:40):
input, right through their earlyyears. And so the so there's the
technology in the at least inthe first instance, there's the
business model.
And then, there's what Arm did,particularly in the nineties, by
way of road mapping, which wasnot all about making bigger and

(34:01):
faster processors. It was aboutunderstanding that when you put
the processor into a system onchip, debugging becomes a whole
new ballgame. Right? The model,through to the end of the
eighties, was the in circuitemulator. You know, you pull
your processor out of thecircuit board, plug it in in
circuit emulator, and then youcan see what's happening and

(34:21):
debug your software.
Once the processor is embeddedin an SoC, that model breaks,
because you need to build a newin circuit emulator for every
chip, and that's just not gonnawork. So they they developed the
on chip debug technology. Sothey gave you effectively ICE
functionality around theprocessor on the chip. And they

(34:44):
really focused on making the armeasy to use. We we had this sort
of view that, you know, thatclever designers like to build
stuff to impress their cleverdesigner friends.
But if you want to address a bigmarket, actually, you want to
make stuff simple. And you haveyou have to worry about the

(35:06):
large teams of designers outthere who perhaps aren't quite
as clever as your friends, andbuild something that that, you
know, the engineer on the jobcan can you can work with. And
and Arm, I think, did a verygood job of focusing on that
aspect of the business, ofmaking the Arm easy to use in
very complex SOCs.

So looking back, did you ever or did you
ever anticipate, or were you,like, blown away by how popular
the BBC Micro and then whatbecame Arm?

We were certainly blown away by the pop popularity
of the BBC Micro, you know,which exceeded all predictions,
as did, of course, a lot ofother computers. This isn't
unique to the BBC Micro. In theUK, Sinclair was building the
zx8081, and those marketsexploded too. So it wasn't just
the BBC Micro. Everybody wassurprised by interest in home

(36:01):
computing and the way peoplecould find applications.
You know? So BBC Micros wereused in large numbers just as
dumb terminals attached tomainframe computers, for
instance, because they werecheaper than the standard
commercial terminals. So thereare all sorts of surprises for
people. And certainly with Arm,I don't think anybody in the

(36:23):
early days could have predictedthe success today. I mean, the
scale is just mind boggling.
I think that the last number Iheard is that over 250, 000,
000, 000 ARM powered chips arebeing shipped by ARM's very
large partner network. So you'retalking about 40 chips for every

(36:44):
human on the planet is armpowered. Nobody anticipated
that, but the growth is just hasbeen sort of rapid and
exponential since ARM limitedwas formed in 1990.

So let's, talk about something more recent, the
Spinnaker project, which wetouched on a little bit earlier.
So can you explain what theSpinnaker project is?

Yeah. Spinnaker is a machine that was conceived
slightly over 20 years ago as amassively parallel computer
optimized for brain modelingapplications. Brains are are
based on networks of biologicalneurons, and the way these
neurons are assembled in in verylarge numbers makes for an

(37:31):
embarrassingly parallel model ifyou try and compute it. And so
Spinnaker was was intended toexploit that embarrassingly
massive parallelism by buildingan embarrassingly massively
parallel computer. And we setourselves the goal early on of
putting a 1000000 ARM processorsinto this machine.
It was clear even from theoutset that with a 1000000 ARM

(37:53):
processors, you don't get closeto the scale of the human brain,
but you are in the region of ofmouse brain scale, potentially.
It's it's based on, in manyways, a relatively conventional,
highly parallel computer. What'sunconventional is firstly that
the processors are all low endenergy efficient embedded

(38:14):
processors, rather than high endmath processors that you find in
typical HPC applications. Andthe second is the way we
implemented the connectivity.The brain is hugely connected.
Each neuron connects to manythousands of others. And so to
implement a model on a machine,you need to replicate that

(38:35):
degree of connectivity. And thatrequired that we moved away from
the typical point to pointcommunication you find in a in a
normal, parallel computertowards a multicast
communication. So when themessage is sent by 1 processor,
it can be delivered to 100 or1000 of others, and and and and

(38:56):
but effectively, thecommunication path is a tree
across the machine. And if youlike, the key innovation in
Spinnaker is how we implementthe routing for those trees,
because they're they'recompletely flexible.
So you can define any treestructure, and therefore, you
can model any part of the brain.Different parts of the brain

(39:17):
have quite differentconnectivity diagrams. So we we
virtualize the connectivitythrough this innovation on

So were these cores actual, like, silicon
cores, or was this allsynthesized on a lot of
different FPGAs? Like, what'sthe actual architecture like?

It it it it's based on an ASIC. So, yes, the
the the the ARM cores aresynthesized in the standard ARM
synthesis root way using ARM'spreferred tool flow. But then
they're they're implemented with18 cores on a chip, on spinnaker
1, with this interconnecttechnology joining them together

(39:56):
on the chip and and and formingthe links to other chips. So the
the core of Spinnaker is is acentimeter square ASIC with 18
arm cores on each ASIC.

Okay.

How does the user define these network trees in,
the project?

So the the the goal with with Spinnaker is is
to enable regular users to usethe machine without
understanding any of thisinterconnect technology. And and
users can describe the networksthey want to simulate in a
language called Pine, whichsimply stands for Python Neural
Network. And that language wasdeveloped by collaborators at

(40:41):
CNRS in Paris. And and you canrun a Pine model on your laptop
using a simulation engine suchas Brian or 1 of several others.
When you've developed your Pinemodel, if you want it to run
faster, you can then simply portthat model to Spinnaker, And the
software stack that we developedat Manchester will take that
Pyne model, allocate, processorresources to modeling the

(41:05):
various neuron populations inthe model, and it will compute
the routing tables to make thenecessary connections.
For a standard user who's mainlyinterested in, if you like, the
neuroscience and the networklevel, They don't need to know
anything about what's happeningat the low level in Spinnaker.

(41:25):
The enthusiastic user who wantsto develop, for example, their
own neural model or their ownlearning rule, both of which are
implemented in code, has to diga bit deeper. But again, they
can do that within anenvironment where the software
stack handles all the routing.So I think very few of our users

(41:47):
actually get directly involvedin understanding how the routing
works, and in driving itthemselves. That's all handled
by the software stack, and it'sbasically it's a problem of
mapping 1 graph, which describesthe neural network you want to
simulate onto another graph,which describes the machine.

(42:08):
Okay. And and the goal in thecommunications architecture is
to is to virtualizecommunication, so that that
mapping can happen asefficiently as possible for as
many different problems aspossible.

Right. So the user's abstracted from having to
go that deep.

Yes. As I say, I suspect very few Spinnaker users
have done anything hands on withthe routing.

So given the current rise in, like, neural
nets and AI, what's your thoughton, like, the current state of
that kind of technology?

It's been a very interesting parallel development
because we've been working onSpinnaker for over 20 years now.
And over those same 20 years,neural networks as used in
connectionist AI, have come fromnowhere to being absolutely the
dominant technology. And thereare similarities and differences
between the networks that we runon Spinnaker and the networks

(43:08):
that are used in AI. People tendto refer to AI networks as as
second generation, and Spinnakerspiking networks as 3rd
generation. So there aresimilarities and and quite
fundamental differences.
I think what's interesting isthat going forward, people are

(43:29):
beginning to see conversions.And some of the problems
currently facing AI, such as thecompletely unsustainable energy
demands of of current leadingedge large language models might
be addressed by using more brainlike approaches of the sort,
supported by platforms such asSpinnaker. If you look at those

(43:53):
big AI networks, they're dense.So the matrices that describe
the connections are densematrices, and that's why they
run very nicely on GPU's and andsimilar. If you look at the
brain, nothing in the brain isis dense.
Okay? All the all theconnectivity in the brain is
sparse, and the activity issparse. And if you could capture

(44:14):
that sparseness in an effectivelarge language model, then you
might see orders of magnitudereduction in the energy
requirements to run that model.And that would be very
important. I mean, we it hasn'tbeen compellingly demonstrated
yet, but I think we're not faroff.

So, just out of curiosity, outside of your
groundbreaking technology, I'mI'm curious what your hobbies
and other interests are.

Yeah. So so I was while I was going through
and building up these notes,Steve, I saw a picture on your
Wikipedia page of you playing abass guitar. So earlier you were
talking about flying. How didyou go from fly you went from
flying or thinking about flying?Did you act oh, you spent you
said an hour in a glider?

I I spent a cumulative 54 minutes flying
time over 1 year of standing onan airfield every Wednesday
afternoon at the CambridgeUniversity Glider Club and
ending up in the air verylittle, which I decided in my
3rd year when I had my finalexams was probably not the most
efficient use of my time. Sothat was my 1 year of of real

(45:28):
flying as it were.

That's where you got into processing. But
how'd you go into the bassguitar?

Well, I've I've been playing guitar since my
teens. Okay? And and III playedin a in the Christian music
group at at at university.There's what in the sideline is
where I met my wife, so it hadlots of benefits. They they also
used to say in this music group,it wasn't a great music group,
but it was a great marriagebureau.
So I and, you know, III don'tconsider myself as sort of, you

(45:58):
know, the top tier musician,that I'm a sort of I can
technically hack my way throughthings on both 6 string and bass
guitar. Mainly, I've played bassfor the last 30 years, because
that the the the church musicgroups I've been in, there's
been more need for a bass than aregular guitar. But just
recently, I've been playing abit more 6 string because, for

(46:19):
various reasons, we haven't hada keyboard player. And just
playing bass doesn't really workfor leading singing. You you
need a bit more middle thanthat.
Yeah. So, you know, me middlerange music has has has has been
a long term interest of mine andor or activity. And and and and

(46:40):
I enjoy that because it's acompletely different mental
process from from what what youregularly do. I mean, if you if
you play music in in a in a bandor church group, you've got to
sort of listen to the others,and and and fix in, and adapt,
and work out what to do whensomebody gets something wrong.
Right?
Which happens quite a lot at thelevel I play at. Just just patch

(47:01):
things over, and it's a it's avery different mental process.
Otherwise, in terms of hobbies,yeah, I guess I'm I'm not really
done that much on the flyingfront for a long time now, not
even with the models. I likewalking. I go out for a bit of a
walk every day to try and keepmy art working for a bit longer.
You know? It's a nice fall ofhazards when you get to my age.

So so I I like it. It it goes math,
aeronautical engineering, fluiddynamics, digital processor,
bass guitar. I'm

III like it. That's that's a good sequence.
Yeah. I feel like I've I've madeto my hobby, my my day jobs 2 or
3 times in my career. I've sortof shifted focus.
And then, you know, I I have noregrets about choosing maths for
my first degree because it's avery adaptable subject. Right?

(47:56):
You can use it almost anywhere.By the way, of course, in in the
UK, we do maths in the plural. Iknow in the US, you tend to just
do 1 math.

Right. Right. Yeah. So actually, the shoehorn
that do 1 other question we hadwas, advice for aspiring
engineers. You said, you hadturned your hobby into your
career multiple times.
So do you have any advice forfuture, engineers and computer
scientists?

Well, I I think the main advice I give to young
people who are at the point ofsort of deciding which degree
subject to do or what career tochoose Unless you have a true
vacation, so you know you wantto be a doctor, in which case,
you know, you better do amedical degree. Right? That's
that's pretty easy. But ifyou're not sure what you want to
do, and most young people aren'tat that stage, then think about

(48:47):
what keeps the most doors open.You know, so doing a sensible
core STEM subject, such as mathsor physics or computer science,
you know, you can turn your handto most things with that with
that kind of background.
So don't specialize too soon isis is is probably my advice to
people at that early careerstage. And and and then you

(49:08):
should choose a degree whichinterests you. Right? Because
degrees are are hard work,they're supposed to be. You have
to you have to find themotivation to keep going some of
the time, and and that's easierif you're interested in what
you're doing.

Let's say computer science didn't work
out. What would have been thebackup career for you?

I I think I could have stuck with my fluid
dynamics. I could, I could havegot somewhere as a, as an
academic researcher in the fluiddynamics space, I think. The
decision point I made there wasthe end of my research
fellowship, was do I continuewith the fluid dynamics, or do I
go join Acorn? Was basically,the decision there. And Acorn's

(49:54):
situation, having secured theBBC Micro contract, was so
interesting that it was a fairlyeasy decision to to to go to
Acorn.
But I think had that not been,available, I would probably have
continued further with fluiddynamics.

And if you had, if you could go back in time to
that period and tell yourself apiece of advice, what would it
be?

I think I I don't know. I think I think, you know,
just find stuff that interestsyou, and and and and try and,
you know, try and earn yourliving from stuff you find
really interesting. And then it,you know, it it it feels less
like work. If it's stuff you'dlike to do. Would you would you
still do this if nobody paidyou?

(50:39):
Okay. You need somebody to payyou to pay the bills, but if you
didn't need that, would youstill do this? And and, you
know, with with with what I'mdoing in computing, that's still
true. You know, I'm now retired.I don't need anybody to pay me
to do anything, but I'm stilldoing stuff because it's
interesting.

III view it as, would you do your day job if it
all it gave you was lunch?

Yeah. I like that.

That's about right. And and I mean, there
probably aren't that many peoplewho can answer yes to that. But
I I think if you can answer yesto that, you're very fortunate.
And I always have been able toanswer yes to that.

You certainly made the right decision back
back then.

You know, life is unpredictable. You can only
optimize locally. Right? It's Do

you have anything else, Steven?

No. I think I think that's great. We we really
appreciate you coming on andspending an hour with us, Steve.

Yeah. It's been good fun.

Steve, is there if anyone's got any questions
for you, how can they reach outand talk to you?

People seem to find me on LinkedIn with that
too, a stroll.

Put the link, to your LinkedIn in our show
notes.

Okay. Yeah. So people can find me on LinkedIn
and contact me through that. Idon't read the communication on
LinkedIn every day, but I dotend to notice if somebody sent
a message eventually.

Well, thank you so much, Steve. Okay.

Yeah. I'll do the outro. Okay. Thank you for
listening to circuit break fromMacrofab. We were your host,
Steve and Craig

and Parker Dillman. Take it easy. Later,
everyone. Alright, Steve. I'mgonna just have a little outro
here I I need to

say. Sorry?

I just have a little outro to say. Thank you
Yes You Breaker for downloadingour podcast. Tell your friends
and coworkers about circuitbreak podcast from Macrofab. If
you have a cool idea, project,or topic you want us to discuss,
let Steven and I and thecommunity of Breakers know. Our
community where you can findpersonal projects, discussions
about the podcast, andengineering topics and news is

(52:44):
located atform.macfab.com.

All Episodes

EP#437: Embarrassingly Parallel Computing - Steve Furber

Episode Transcript

Popular Podcasts

On Purpose with Jay Shetty

Stuff You Should Know

Dateline NBC

.css-15opob5{left:0;position:absolute;top:0.8rem;} All Episodes

.css-14f5ked{margin:0;word-break:break-word;display:-webkit-box;-webkit-box-orient:vertical;box-orient:vertical;-webkit-line-clamp:2;overflow:hidden;}EP#437: Embarrassingly Parallel Computing - Steve Furber