mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Hardware (https://www.mersenneforum.org/forumdisplay.php?f=9)
-   -   1000 core chip that runs on AA batteries! (https://www.mersenneforum.org/showthread.php?t=21390)

tServo 2016-06-21 16:19

1000 core chip that runs on AA batteries!
 
They claim it consumes 7 tenths of a watt executing 115 billion instructions/sec.
Interesting, not from a practical standpoint but to shows what's possible and maybe on the horizon. It sure would be the bomb, tho, to put several of these to work on our stuff.
[URL="https://www.ucdavis.edu/news/worlds-first-1000-processor-chip"]https://www.ucdavis.edu/news/worlds-first-1000-processor-chip[/URL]

It looks like they cut corners on the design all over the place ( cache, instruction set,
FP thruput ( single precision only, of course), but still ......

science_man_88 2016-06-21 18:24

okay that's about 6.086956 picojoules per instruction by my math so at full instruction speed in theory I get about 10.834782 joules used per second ( aka watts) but other than running multiple tests LL isn't parallelizable so it's only use might be TF or some other thing done in parallel.

airsquirrels 2016-06-21 19:32

FFT multiplication is (sort of) paralizable. Thus the good performance of clLucas/CUDALucas, etc.

xilman 2016-06-21 20:01

[QUOTE=science_man_88;436643]okay that's about 6.086956 picojoules per instruction by my math so at full instruction speed in theory I get about 10.834782 joules used per second ( aka watts) but other than running multiple tests LL isn't parallelizable so it's only use might be TF or some other thing done in parallel.[/QUOTE]
Do you [b]really[/b] believe that figure?

The raw data is 0.7W and 115e9 instructions per second. All we can take away from that is that the power draw lies between 0.75 and 0.65W, with the performance lying between 114.5e9 and 115.5e9 ips.

Thus all we can conclude is that the energy per instruction lies somewhere between 0.65/115.5e9 and 0.75/114.5 pJ, or between 5.6pJ and 6.6pJ.

IOW, there's a good chance that not even the first digit you quote is correct. Why, then, do you state nine digits? That amount of precision is total BS.

science_man_88 2016-06-21 21:48

[QUOTE=xilman;436652]Do you [b]really[/b] believe that figure?

The raw data is 0.7W and 115e9 instructions per second. All we can take away from that is that the power draw lies between 0.75 and 0.65W, with the performance lying between 114.5e9 and 115.5e9 ips.

Thus all we can conclude is that the energy per instruction lies somewhere between 0.65/115.5e9 and 0.75/114.5 pJ, or between 5.6pJ and 6.6pJ.

IOW, there's a good chance that not even the first digit you quote is correct. Why, then, do you state nine digits? That amount of precision is total BS.[/QUOTE]

partly because I was typing it from a calculator. also I think you'll find that based on your own calculations, that 6 shows as the first digit in your range about 7 out of 11 times.

chalsall 2016-06-21 21:56

[QUOTE=science_man_88;436664]partly because I was typing it from a calculator. also I think you'll find that based on your own calculations, that 6 shows as the first digit in your range about 7 out of 11 times.[/QUOTE]

You are clearly too young, or perhaps simply too inexperienced, to understand experimental error.

Rather than giving six decimal points of meaningless precision, give a +/- on the measure.

science_man_88 2016-06-21 21:59

[QUOTE=chalsall;436666]You are clearly too young, or perhaps simply too inexperienced, to understand experimental error.

Rather than giving six decimal points of meaningless precision, give a +/- on the measure.[/QUOTE]

I was just doing straight math from the given numbers. doing it from xilman's range of 5.6-6.5 pJ it gives 9.97 watts to 11.7 watts at full power rounded to 3 places.

chalsall 2016-06-21 22:01

[QUOTE=science_man_88;436667]I was just doing straight math from the given numbers.[/QUOTE]

And that was your mistake.

bgbeuning 2016-06-22 02:04

Just think about the RAM bottleneck ...

science_man_88 2016-06-22 10:55

[QUOTE=bgbeuning;436684]Just think about the RAM bottleneck ...[/QUOTE]

[QUOTE]and they transfer data directly to each other rather than using a pooled memory area that can become a bottleneck for data.[/QUOTE]

so they don't rely on ram.

Antonio 2016-06-22 16:55

[QUOTE=science_man_88;436702]so they don't rely on ram.[/QUOTE]

So what magic do they use to transfer instructions and data into/out of the processors? :smile:


All times are UTC. The time now is 02:38.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.