Parallella: Supercomputer for Embedded Systems

Slide 1

Slide 1 text

Chicago ACM Chicago, IL, USA December 9, 2015 Parallella: Supercomputer for Embedded Systems

Slide 2

Slide 2 text

Moore’s Law: 2x every 18 months

Slide 3

Slide 3 text

Moore’s Law: 2x every 18 months Moore’s Law 1993 - 2013 http://www.washingtonpost.com/blogs/innovations/wp/2015/04/14/10-images-that-explain-the- incredible-power-of-moores-law/ https://www.parallella.org/board/

Slide 4

Slide 4 text

No content

Slide 5

Slide 5 text

RayHightower.com

Slide 6

Slide 6 text

IBM Blue Gene https://commons.wikimedia.org Cray X-MP

Slide 7

Slide 7 text

RJ-45 Power μUSB μHDMI μSD

Slide 8

Slide 8 text

RISC ARM 18 cores: 2 ARM + 16 RISC

Slide 9

Slide 9 text

No content

Slide 10

Slide 10 text

No content

Slide 11

Slide 11 text

No content

Slide 12

Slide 12 text

Watts => Dollars.

Slide 13

Slide 13 text

No content

Slide 14

Slide 14 text

17.8 million watts $17.8 million per year

Slide 15

Slide 15 text

5 watts for Parallella?

Slide 16

Slide 16 text

No content

Slide 17

Slide 17 text

http://rayhightower.com/blog/2014/09/09/solar-powered-parallella/

Slide 18

Slide 18 text

http://rayhightower.com/blog/2014/09/09/solar-powered-parallella/

Slide 19

Slide 19 text

http://rayhightower.com/blog/2014/09/09/solar-powered-parallella/ 5 volts 1 amp 5 watts Solar!

Slide 20

Slide 20 text

If one ox could not do the job they did not try to grow a bigger ox, but used two oxen. When we need greater computer power, the answer is not to get a bigger computer, but to build systems of computers and operate them in parallel. -Grace Hopper

Slide 21

Slide 21 text

Concurrency is not Parallelism. https://www.youtube.com/watch?v=cN_DpYBzKso&list=PLOnWKC1gI_OPU8SDIBnCLHsgzNLSbnPJQ&index=3 -Rob Pike, Go

Slide 22

Slide 22 text

Reduced Instruction Set Computer RISC

Slide 23

Slide 23 text

ARM A9

Slide 24

Slide 24 text

RISC ARM 18 cores: 2 ARM + 16 RISC

Slide 25

Slide 25 text

No content

Slide 26

Slide 26 text

Hello Epiphany

Slide 27

Slide 27 text

host.c epiphany.c

Slide 28

Slide 28 text

No content

Slide 29

Slide 29 text

Find all primes up to 16,000,000. Serial on Parallella.

Slide 30

Slide 30 text

1 #include 2 #include 3 #include 4 #include 5 6 #define DEFAULT_MAX_TESTS 16000000 7 8 inline int isprime(unsigned long number) 9 { 10 unsigned long i; 11 unsigned long s = sqrt(number); 12 for(i=3;i<=s;i+=2) 13 { 14 if(number % i == 0) 15 return 0; 16 } 17 return 1; 18 } /* Copyright (c) Adapteva, contributed by M. Thompson with modifications by T. Malthouse. */

Slide 31

Slide 31 text

8 inline int isprime(unsigned long number) 9 { 10 unsigned long i; 11 unsigned long s = sqrt(number); 12 for(i=3;i<=s;i+=2) 13 { 14 if(number % i == 0) 15 return 0; 16 } 17 return 1; 18 } /* Copyright (c) Adapteva, contributed by M. Thompson with modifications by T. Malthouse. */

Slide 32

Slide 32 text

No content

Slide 33

Slide 33 text

237.1 sec

Slide 34

Slide 34 text

Find all primes up to 16,000,000. Serial on Mac OS X.

Slide 35

Slide 35 text

Same serial code, written in C. Build it on OS X.

Slide 36

Slide 36 text

No content

Slide 37

Slide 37 text

14.4 sec

Slide 38

Slide 38 text

Find all primes up to 16,000,000. Parallel on Parallella.

Slide 39

Slide 39 text

27 #include 28 29 // Default max number of prime tests per core 30 // Used if a limit it not provided in argv[1] 31 #define DEFAULT_MAX_TESTS 500000 32 33 int main(int argc, char *argv[]) 34 { 35 unsigned row, col, coreid, i, j; 36 e_platform_t platform; 37 e_epiphany_t dev; 38 /* Copyright (c) Adapteva, contributed by M. Thompson with modifications by T. Malthouse.*/

Slide 40

Slide 40 text

No content

Slide 41

Slide 41 text

18.6 sec

Slide 42

Slide 42 text

Summary: Finding Primes 0" 50" 100" 150" 200" 250" Parallel Parallella Serial Mac Serial Parallella 18.6 sec 14.4 sec 237.1 sec ($2,000.00 Apple MacBook Pro) ($150.00 Parallella)

Slide 43

Slide 43 text

Embarrassingly parallel problem.

Slide 44

Slide 44 text

Mandelbrot Set

Slide 45

Slide 45 text

No content

Slide 46

Slide 46 text

Why?

Slide 47

Slide 47 text

Finite Element Analysis http://www.ce.ncsu.edu/news/article/21550/making-bridges-more-robust-to-earthquakes/

Slide 48

Slide 48 text

break? wear out? work? Finite Element Analysis Will it {

Slide 49

Slide 49 text

Free Body Diagram F F applied friction F gravity F normal

Slide 50

Slide 50 text

Finite Element Analysis http://www.ce.ncsu.edu/news/article/21550/making-bridges-more-robust-to-earthquakes/

Slide 51

Slide 51 text

https://www.e-education.psu.edu/worldofweather/node/2029 Grid spacing inﬂuences accuracy. Weather Prediction

Slide 52

Slide 52 text

Barriers to Entry

Slide 53

Slide 53 text

Barriers to Entry Removing

Slide 54

Slide 54 text

$ ssh $ Xvfb rayhightower.com/blog/2015/12/08/how-to-run-headless-parallella/

Slide 55

Slide 55 text

Parallella vs GPU

Slide 56

Slide 56 text

384 cores MacBook Pro Video

Slide 57

Slide 57 text

Specialized vs. General

Slide 58

Slide 58 text

Embedded Systems

Slide 59

Slide 59 text

No content

Slide 60

Slide 60 text

No content

Slide 61

Slide 61 text

https://www.parallella.org/2015/06/01/the-open-camera-project-1000-bounty-for-open-ﬁrmwaredrivers-for-raspberry-pi-camera-module/