Advanced search

Message boards : Graphics cards (GPUs) : 1 TFLOPS Intel Knights Corner: cGPU PCIe

Author Message
Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 206
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22532 - Posted: 16 Nov 2011 | 19:02:02 UTC
Last modified: 16 Nov 2011 | 19:11:25 UTC

Hello Something may one day be used here.

http://www.xbitlabs.com/news/cpu/display/20111115163857_Intel_Shows_Off_Knights_Corner_MIC_Compute_Accelerator.html

http://seattletimes.nwsource.com/html/technologybrierdudleysblog/2016775145_wow_intel_unveils_1_teraflop_c.html

http://news.softpedia.com/news/NVIDIA-Bested-Intel-Demos-1-TFLOPS-Knights-Corner-Accelerator-234814.shtml

http://www.xbitlabs.com/news/cpu/display/20111115223328_Intel_s_MIC_Compute_Accelerators_May_Adopt_QPI_Bus_Instead_of_PCI_Express.html

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22534 - Posted: 16 Nov 2011 | 21:19:32 UTC - in response to Message 22532.

Interesting.

Pro:
- 1 TFlops measured, no theoretical maximum - that's nothing to sneeze at
- presumably simpler programming than GPUs

Difficulties:
- Where to get the memory bandwidth from?
- Looks like it uses some socket and may use QPI for fast communication (essential).. however, multi socket Intel server boards are really expensive
- Sure it's superior to a Fermi.. but that's comparing 40 nm to 22 nm Tri-Gate, that's approximately 3 full process nodes difference!

MrS
____________
Scanning for our furry friends since Jan 2002

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22538 - Posted: 17 Nov 2011 | 16:25:20 UTC - in response to Message 22534.
Last modified: 17 Nov 2011 | 16:30:28 UTC

It uses PCIE, as with Knight's Ferry.
http://download.intel.com/pressroom/images/Intel_ISC_2010_Hamburg-KnightsFerry.jpg
http://www.thinq.co.uk/media-site/photos/knightsferry-demo.jpg

Here's a very similar review form around 5months ago,
http://www.thinq.co.uk/2011/6/20/intel-pushes-hpc-space-knights-corner/

The memory bandwidth might be high enough on an X79 board for 3930K/3960X processors (with quad channel 2133MHz DDR3), and PCIE3.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

MarkJ
Volunteer moderator
Volunteer tester
Send message
Joined: 24 Dec 08
Posts: 738
Credit: 200,909,904
RAC: 0
Level
Leu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22547 - Posted: 19 Nov 2011 | 6:40:13 UTC - in response to Message 22538.

It uses PCIE, as with Knight's Ferry.
http://download.intel.com/pressroom/images/Intel_ISC_2010_Hamburg-KnightsFerry.jpg
http://www.thinq.co.uk/media-site/photos/knightsferry-demo.jpg

Here's a very similar review form around 5months ago,
http://www.thinq.co.uk/2011/6/20/intel-pushes-hpc-space-knights-corner/

The memory bandwidth might be high enough on an X79 board for 3930K/3960X processors (with quad channel 2133MHz DDR3), and PCIE3.


Looking at that 1st photo you'd swear the thing is an Nvidia reference design GPU, right down to the fan on the end :-)

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22549 - Posted: 19 Nov 2011 | 17:11:23 UTC - in response to Message 22547.

Looking at that 1st photo you'd swear the thing is an Nvidia reference design GPU, right down to the fan on the end :-)


Haha!

If it's PCIe and with such a design it probably uses local memory. I was a bit deceived by the 1st link, where they speculated on it using QPI. However, even with full access to a socket 2011 memory bandwidth through a fast interface, the chip would likely be limited by high memory latency (long way to the memory) and much lower bandwidth than high end GPUs. And in this case performance would plummet if you used 2 of them. However, it seems like Itel is proposing compute servers using 8 of them. So yeah, definitely local memory.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22550 - Posted: 19 Nov 2011 | 17:38:00 UTC - in response to Message 22549.

Perhaps something along the lines of XDR2. Knights Ferry used 2 GB GDDR5.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22552 - Posted: 19 Nov 2011 | 22:03:28 UTC - in response to Message 22550.

Since they wouldn't have to compete with GPUs in terms of price/performance (rather with Teslas), they could easily afford a 512 bit interface, so that ordinary GDDR5 would suffice.

MrS
____________
Scanning for our furry friends since Jan 2002

MarkJ
Volunteer moderator
Volunteer tester
Send message
Joined: 24 Dec 08
Posts: 738
Credit: 200,909,904
RAC: 0
Level
Leu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22554 - Posted: 19 Nov 2011 | 23:41:48 UTC
Last modified: 19 Nov 2011 | 23:58:23 UTC

Reading the article with the photos in it here (the 2nd thinq.co.uk one) that skgiven posted, they say the prototype ones are based on 45nm and only 50 cores without tri-gate, which is probably why they have large power cables.

The production ones are supposed to be 22nm with tri-gate and 80 cores per card. Available in 2012.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22556 - Posted: 20 Nov 2011 | 0:53:42 UTC - in response to Message 22554.

Data can be transferred at up to 32 gigabytes/sec.

The Knights Ferry was supposed to use Quad HT, so I expect that would still be there with Knights Corner. IF it has 80 cores that would that mean 320 threads, and extrapolating from Knights Ferry, 20MB L1 cache.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 206
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22576 - Posted: 24 Nov 2011 | 18:07:22 UTC - in response to Message 22556.

Hello: Last comments from Intel to Nvidia, comparing yields of the two platforms. Greetings.

http://news.techeye.net/chips/intel-is-still-taking-the-mic

Profile oldDirty
Avatar
Send message
Joined: 17 Jan 09
Posts: 22
Credit: 3,805,080
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwatwatwatwat
Message 22643 - Posted: 6 Dec 2011 | 23:00:29 UTC

lol..intel says that GPGPU is not an optimal solution or not the best approach?
hell yea, thats wuy it's so famous and seem to push intel for the knights corner.
Oh yes intel, since Ati's xtx Cards ( R580 ) it was clear that simple VGA's can do so heavy operation in a right shaderlanguage, and the Giant sleeps well.
Now, 5-6 Years later, GPGPU is not the best approach, and oh yes, we have the x86 solutaion for all ya. LOL...


____________

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22649 - Posted: 7 Dec 2011 | 21:55:04 UTC - in response to Message 22643.

Intel isn't disputing the fact that an array of many simple execution units is great for some tasks. Both, GPUs and nights corner belong into this category. What they are disputing, however, is the way this power should be exposed to the software guys.

Sure, GP-GPU is popular.. but tell, where does it actually benefit common every-day tasks? It's actually not as wide spread as the hype around it might make it seem. What Intel is saying: we shouldn't need to develop and switch to new software tools in order to use this hardware power. Or deal with the numerous bugs in graphics drivers.

The easier you can make the crunching power accessible, the more wide spread its adoption will be. That's why you don't see many applications for the superior amount of raw horse power in current AMD GPUs.

Sometimes a bird in the hand is worth more than two in the bush.. which in this case means "I'd rather have somewhat less raw horse power in a Knights Corner than in a GPU, but be able to actually use it".

MrS
____________
Scanning for our furry friends since Jan 2002

Profile oldDirty
Avatar
Send message
Joined: 17 Jan 09
Posts: 22
Credit: 3,805,080
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwatwatwatwat
Message 22651 - Posted: 7 Dec 2011 | 22:31:06 UTC

Ok understand, but intel just come out with it's monster x86 PCIe Card very late, thats what i mean. If they wake up a little earlier, we will have much more Compute capabilities than today.
____________

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 6,169
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22652 - Posted: 8 Dec 2011 | 0:41:42 UTC - in response to Message 22651.
Last modified: 8 Dec 2011 | 0:41:55 UTC

Ok understand, but intel just come out with it's monster x86 PCIe Card very late, thats what i mean.

This is true. The real warning for Intel was that supercomputers aren't built only on CPUs anymore.

If they wake up a little earlier, we will have much more Compute capabilities than today.

It's not necessary true. Knights Corner is not intended for the home users, or gamers. In a highly multithreaded system, subprocess (whatever the manufacturer calls it) distribution is the key for the success of the architecture. We are in the beginning of a new era of computing, it'll take a couple of years until the various implementations improve, and find their best use.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22655 - Posted: 8 Dec 2011 | 11:27:41 UTC - in response to Message 22652.

Unfortunately the intended place for the Knights Corner is in supercomputers, data centers, and the cloud; where software that can utilize so many cores and threads already exists.

I agree that Intel missed an opportunity to produce and release such a device earlier for the general public. Specialist app developers, gaming and research programmers would have all benefited, as would Intel.

Yes it's only been over the last couple of years that general applications have been able to use many CPU cores/threads, but this is a definite failure by Intel to realize an opportunity. If Intel had the vision to develop basic versions of this card and actually release such devices to the general public, many crunchers would be using them, many universities, gamers, editors, and who knows what other fields would have found a use for them. Entrepreneurship was not given a chance, instead Intel is looking to embrace globalization in the form of the cloud.
While they might not be of much use in a desktop system for general computing use, Intel missed a chance to let a market develop itself; researchers, market-app developers, game programmers would have developed software to utilize the hardware, and they would have created a self-motivating market.

Instead Intel developed the Knight's Ferry, but did nothing with it. At this stage Intel has not released or realized anything. Intel developed a Ferry but missed the tide.
The cost of KC will no doubt discourage most, especially saying as NVidia has done so well with CUDA; it's an established research tool and the high end cards are also used with modeling, graphic design, editing, rendering. Even if Intel's card really is just around the corner, and for the GP (highly unlikely), both AMD and NVidia will be releasing new generations of GPGPU, some of which might even be affordable.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Post to thread

Message boards : Graphics cards (GPUs) : 1 TFLOPS Intel Knights Corner: cGPU PCIe

//