Advanced search

Message boards : Multicore CPUs : Parallella

Author Message
Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 33510 - Posted: 15 Oct 2013 | 18:39:30 UTC
Last modified: 15 Oct 2013 | 18:44:24 UTC

It seems that a new concept of computer is comming next month: Parallella
It comes with a Epiphany Multicore chip.
Adapteva is the company behind, it seems they lack of wide experience on putting these chips on the market but they are pushing hard with the help of KickStarter.
Some youtube videos on this subject:
http://www.youtube.com/watch?v=taAG9QNWLNo
http://www.youtube.com/watch?v=jqqxTpz-hJY

What do you think of this? How do you see this compared with a Rasperry Pi cluster?

Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 33946 - Posted: 20 Nov 2013 | 20:04:36 UTC - in response to Message 33510.

No opinnions on this?

I read a few more articles on this and there is still to see what comes out of this proyect.

I was perhaps dreaming that a VM could see every one of these boards as a GPU or a multicore CPU with spetial drivers designed for it.

From my little knowlegde even considering the latency of the ethernet it could make a very nice scalable computer. Some kind of LTSP machine or Beowulf cluster.

What I like of this "dream" is that is very scalable, power efficient and fanless (I hate the noise of the fans).

Maybe someone could ground me to earth.

PD. Even some technical liks on this would be apreciated to search more on this subject.

Thanks!

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 33948 - Posted: 20 Nov 2013 | 23:46:33 UTC - in response to Message 33946.
Last modified: 20 Nov 2013 | 23:48:51 UTC

Could 4 Parallela @ $100 each do as much work as my $400 Q6600-GTX660Ti rig?

Maybe I don't understand the Parallela concept but each board has an ethernet port, HDMI port, 2 USB ports, a serial port plus JTAG port. If I were to buy 10 Parallela I would have all those ports X 10 which to me seems like a needless expense. Gamers like to have multiple monitors so I can justify 3 HDMI for certain users. So on that basis I say the idea is dumb, unless 10 of them could crunch as fast as something comparably priced, for example a GTX 780Ti in a $150 rig. Powerful and inexpensive beats dumb.
____________
BOINC <<--- credit whores, pedants, alien hunters

Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 33950 - Posted: 21 Nov 2013 | 8:10:55 UTC - in response to Message 33948.
Last modified: 21 Nov 2013 | 8:14:55 UTC

I don't think it's intended to reach gammers market at all.

Could 4 Parallela @ $100 each do as much work as my $400 Q6600-GTX660Ti rig?

Nobody knows yet, but the idea behind it that it should, at least perform well for very simple math (It comes with an Epiphany Multicore coprocessor that some complain is too simple).

Maybe I don't understand the Parallela concept but each board has an ethernet port, HDMI port, 2 USB ports, a serial port plus JTAG port. If I were to buy 10 Parallela I would have all those ports X 10 which to me seems like a needless expense.

True, but it's scallable and is fanless that for me is a strong plus.

Profile MJH
Project administrator
Project developer
Project scientist
Send message
Joined: 12 Nov 07
Posts: 696
Credit: 27,266,655
RAC: 0
Level
Val
Scientific publications
watwat
Message 33951 - Posted: 21 Nov 2013 | 11:16:18 UTC - in response to Message 33950.

Right now these things aren't much more than toys, because the absolute performance is very low. They're probably quite educational to play with, but not going to be useful for any heavy-duty compute work (though maybe it'll find some niche vertical).

Matt

Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 33953 - Posted: 21 Nov 2013 | 13:27:35 UTC - in response to Message 33951.

Right now these things aren't much more than toys, because the absolute performance is very low.

Not very sure of that, thats why I am asking.

Acording to Wikipedia a GeForce GTX 770 has the following performance:
Single Prec. 3213 GFLOPS
Double Prec. 134 GFLOPS

A Parallela has 90 GFLOPS (http://www.adapteva.com/parallella-board/)

Maybe we will have to wait and see real performance when they start to come out.


Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 33961 - Posted: 21 Nov 2013 | 18:12:17 UTC - in response to Message 33953.

OK, now I see why each board needs an ethernet port... they're connected by hooking them to a router/switch? I thought they connected through a pin and header combo and kind of an 'extendable bus' sort of thing.

Regarding performance...

The only reason GTX 770 DP is that low is because NVIDIA cripples it on purpose. When not crippled the DP performance is the same as the SP performance. Titan, for example, does not have DP crippled and therefore has excellent DP performance.

Also, you have to look at more than just GLOPS when assessing performance, you must also look at the kinds of operations the device can perform. If it can only add and multiply with DP then it's not as productive at certain kinds of work as a device that can do transcendental, matrix and FFT operations (higher maths) in hardware. Does the Epiphany do higher maths and if so which ones?

____________
BOINC <<--- credit whores, pedants, alien hunters

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2671
Credit: 752,534,974
RAC: 459,487
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 33963 - Posted: 21 Nov 2013 | 21:27:59 UTC

You can take a look at this thread over at Einstein. It's quite long and sometimes off-topic, but you should get a pretty good idea.

The current Parallela is a very intresting proposal, as it provides significantly faster writes than usual many-core interconnects. This may enable to runs special algorithms more efficiently and may lead to downright now ones. However, as always there's a caveat: this also means that you have to get your hands dirty with low level optimizations and coding, with a very good understanding of how the hardware works and how to make the best use of it. It's the opposite of an i7, so to say: the former is a specialized chip needed significant skill and development effort to use efficiently, whereas the latter tries to be as simple as possible for the programmer (given the constrains of the x86 instuction set, of course).

The idea is to have someone do the dirty work and generate optimized libraries for the Parallelas, which you'd then call like a GPU routine to use them as energy-efficient accelerators. This could work, but it's a long way ahead.

To pack so much crunching power into such a tiny space, and to make the chip affordable, Parallela had to make do with very simple individual cores and very small local memory. This provides a further challenge to using them well. Being fanless is just a side effect of the chip being so small. Make any processor small and slow enough and it will run fanless without problems. But then you need many of them to get performance.. and if you pack them tightly enough you need fans again, just like the Parallela Mini-Cluster you linked to.

The issue of low absolute performance, too much overhead with many boards and very low local memory could be relatively easily corrected by just building larger chips (it's a scalable architecture, after all!). But don't hold your breath for that until a new money stream opens. And for this to happen the design probably needs a killer application. A single one is enough.. but as far as I know it's not there yet (I don't mean ideas, but rather useable solutions).

MrS
____________
Scanning for our furry friends since Jan 2002

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2671
Credit: 752,534,974
RAC: 459,487
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 33965 - Posted: 21 Nov 2013 | 22:00:53 UTC - in response to Message 33961.

When not crippled the DP performance is the same as the SP performance. Titan, for example, does not have DP crippled and therefore has excellent DP performance.

BTW: that's not quite true, high-performance CPUs run DP at 1/2 of SP speed. Making DP faster would require significantly more transistors. And if you built such a chip it would also be faster at SP. I don't think the ratio can get any better than 1/2 without going to inefficient designs.

And Titan at 1/3 is indeed very good here, but so are AMDs Cayman and Tahiti at 1/4, whereas the brand-new Tahiti pays tribute to its gaming-focus and throttles back to 1/8. Which I think is hardware-related, not software crippled like for GK110 and GTX780/Ti.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 33967 - Posted: 21 Nov 2013 | 23:01:15 UTC

Thaks very much for your aswers, I learned a lot.

I read the Einstein thread you posted ETA. Very, very interesting. That's was exactly what I was looking for.

PD. It made me go back in time... when I programmed PIC and Motorola 68000... very long time ago... Does Asemmbler still exists?? :op

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2671
Credit: 752,534,974
RAC: 459,487
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 33989 - Posted: 23 Nov 2013 | 16:32:24 UTC - in response to Message 33967.

Does Asemmbler still exists?? :op

Of course it does.. and I think everyone is happy not avoid it as much as possible :D
I programmed a simple state machine in assembler at university. Oh dear!

MrS
____________
Scanning for our furry friends since Jan 2002

Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 34027 - Posted: 26 Nov 2013 | 8:17:47 UTC - in response to Message 33989.

They start delivering next week. We will see soon what is up to come.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2671
Credit: 752,534,974
RAC: 459,487
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 34126 - Posted: 4 Dec 2013 | 20:40:26 UTC - in response to Message 34027.

Wow, that was probably close to being bankrupt.. let's keep our fingers crossed everything works out for them and the backers! They did sound optimistic in the past as well..

MrS
____________
Scanning for our furry friends since Jan 2002

Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 34243 - Posted: 12 Dec 2013 | 15:06:16 UTC - in response to Message 34126.
Last modified: 12 Dec 2013 | 15:09:14 UTC

They have the first 200 boards. Start shipping.



Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 34696 - Posted: 16 Jan 2014 | 23:22:58 UTC

Ericsson (the Swedish company) and Carmel Ventures (a Israeli VC firm) have made a significant investment in Adapteva of 3.6M$.

Seems this is going to be an arquitectur to watch.

Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 34697 - Posted: 17 Jan 2014 | 8:46:29 UTC - in response to Message 34696.

An interesting demo on parallelization with Parallella.

Post to thread

Message boards : Multicore CPUs : Parallella