backtop


Print 39 comment(s) - last by acompsys.. on Nov 29 at 1:09 PM


Rajeeb Hazra, General Manager of Intel's Technical Computing Group holding “Knights Corner”
Over 1 TeraFLOPS on a single chip

GPGPU and cloud computing have been hot topics for the last several years. Intel has shown off several designs like Larrabee and the Single-chip Cloud Computer in the past. However, it is Knights Corner that will be the firm's first commercial product to use the Many Integrated Core (MIC) architecture. The co-processor will be offered as a PCIe add-in board.

The MIC concept is simple: Use architecture specifically designed to process highly parallel workloads, but ensure compatibility with existing x86 programming models and tools.

This would give MIC co-processors the ability to run existing applications without the need to port the code to a new programming environment, theoretically allowing maximum CPU and co-processor performance simultaneously with existing x86 based applications. This would dramatically save time, cost and resources that would otherwise be needed to rewrite them to alternative proprietary languages.

AMD and NVIDIA have been trying to do with their latest architectures by enabling support for languages like C++, but Intel wants to challenge them in this potentially lucrative market.

Knights Corner will be manufactured using Intel’s latest 3-D Tri-Gate P1270 22nm transistor process and will feature more than 50 cores. Intel demonstrated first silicon of Knights Corner at the SC11 conference yesterday. The co-processor wowed the crowd by delivering more than 1 TeraFLOPS of double precision floating point performance.

The firm also touted its "commitment to delivering the most efficient and programming-friendly platform for highly parallel applications", and showed off the benefits of the MIC architecture in weather modeling, tomography, protein folding, and advanced materials simulation at its booth.

There is no timeframe on when Knights Corner will enter production or be available to customers.


Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: holy hell..
By kattanna on 11/16/2011 10:39:33 AM , Rating: 3
yes, their video cards.

thats what this is going up against, GPU computing, not your desktop CPU computing.

and i cant wait to see the BOINC apps being able to use this thing, that be sexy


RE: holy hell..
By ViroMan on 11/16/2011 11:36:07 AM , Rating: 2
GPU computing for bionc like apps usually offers many times more performance then a CPU. Im willing to bet 3-5 GPUs will beat 50 CPUs on a board. Although I don't know what the wattage for the CPU card would be, 5 GPUs sure as hell going to need an external PSU.


RE: holy hell..
By kattanna on 11/16/2011 11:58:28 AM , Rating: 2
you can already get teraflop computing within your desktop

http://www.nvidia.com/object/personal-supercomputi...

using a pair of these cards. If intel can do the same using one addon card, and if the price is right, then they could do very well in this area.


RE: holy hell..
By benny638 on 11/16/2011 12:42:05 PM , Rating: 2
515 Gigaflops is what the Tesla is able to do in double precision. The Intel drop in card can do 1000+ Gigaflops in double precision. This makes the intel card twice as fast. Impressive to say the least.


RE: holy hell..
By Khato on 11/16/2011 12:55:33 PM , Rating: 2
Actually, the Tesla C2050 can do 515 DP GFLOP peak. According to this nvidia presentation (http://www.nvidia.com/content/GTC-2010/pdfs/2057_G... it looks like they get around 360 GFLOP DGEMM. Hence the single Knights Corner chip is delivering around 3x the performance of the Tesla C2050. It'll be quite interesting to see what production performance looks like... not to mention power consumption.


RE: holy hell..
By Samus on 11/16/2011 2:20:22 PM , Rating: 2
Not true.

My GTX480's in 2x SLI rank up 615 gigaflops in cudaprof, and this configuration is nearly two years old. GTX580's should scale to around 750 gigaflops, and this is something available RIGHT NOW.

Tesla has been able to achieve a Teraflop in 3xSLI configuration for years.

Intel offering a chip that can do this is simply an entra point to the sector, not a groundbreaking achievement. However, knowing Intel, they will dominate the field in power consumption.


RE: holy hell..
By EJ257 on 11/16/2011 2:54:02 PM , Rating: 2
True but Intel is doing this on a single chip vs. the SLI configuration which still requires two or three GPUs. I'm sure Intel is also working on a way to make multiples of these chip work in parallel like SLI/X-Fire.


RE: holy hell..
By Belard on 11/17/2011 10:41:46 PM , Rating: 2
Really?! Where can you buy this SINGLE chip from intel? Which *IS* not a drop CPU replacement. It requires a PCIe slot.

If this intel chip make it to market in 1-2 years, AMD and Intel will have similar GPU systems up before then.

This is why Intel canceled their 3D video card plans... they don't have the drivers or technology to compete against AMD & Nvidia. They can't match or surpass them.


RE: holy hell..
By someguy123 on 11/16/2011 3:30:32 PM , Rating: 2
You're comparing 3 cards to a single knights ferry....


RE: holy hell..
By Aloonatic on 11/16/2011 3:46:48 PM , Rating: 2
To be fair, he's comparing 3 cards that actually exist and can be bought in the shops to one specially selected piece of silicon that only exists Intel's super secret research lair.

By the time that Intel's processor reaches the market, it'll be interesting to see what one Nvidia card could do, or how ever many that you could buy and put in a rig for the same money as the rig that you'll need to get those figures out of a Knights processor.

Still, at least it's good to see that the Larrabee time and money might not be going to waste after all.


RE: holy hell..
By someguy123 on 11/16/2011 3:49:44 PM , Rating: 2
I'm not saying intel will dominate, or that this card will even see the light of day within time to be relevant, but to say something like "we can already do this, but with 3 cards!" is silly. Similar to someone dropping a petaflop chip and saying "we've done that with supercomputers for years!"


RE: holy hell..
By Argon18 on 11/16/2011 11:49:44 PM , Rating: 3
knights ferry??? lmao i don't know why but that's funny!


RE: holy hell..
By kattanna on 11/16/2011 5:30:03 PM , Rating: 2
quote:
GTX580's should scale to around 750 gigaflops, and this is something available RIGHT NOW.


according to BOINC, my HIGH OC 560ti can get 901 Gflops peak


RE: holy hell..
By Da W on 11/16/2011 3:55:57 PM , Rating: 3
Telsa is 40nm, Intel's 50 core is 22 nm. Next generation Telsa will get there.


RE: holy hell..
By andre-bch on 11/16/2011 6:09:07 PM , Rating: 2
Lot's of misinformation in the above posts.

The highest performing single GPU tesla is M2090 which can do 665 DP GFLOPs.

Geforce GTX 4xx and 5xx cards' DP performance are intentionally limited to 1/8 of their SP performance. Teslas' DP is 1/2 of their SP performance.

Meaning; a GTX 480 can do 1344 SP GFLOPs, but only 168 DP. A tesla M2090 achieves 1331 SP GFLOPs, but 665 DP.

All in all, the upcoming 28 nm kepler, will at least have twice the DP performance of fermi and should be available in 1H 2012. Knights Corner will have a hard time competing with that.

http://en.wikipedia.org/wiki/Nvidia_Tesla

http://techreport.com/articles.x/18682


RE: holy hell..
By andre-bch on 11/16/2011 6:22:46 PM , Rating: 2
Forgot that 460, 550 and 560 are also GTX. Why isn't there an EDIT button again?!

Anyways, among geforces, GF100 and GF110 based cards have 1/8 DP number of their SP performance, for the rest, GF104, 106, 114, 108, etc., it is 1/12.

http://www.anandtech.com/show/4221/nvidias-gtx-550...

http://www.anandtech.com/show/3973/nvidias-geforce...


RE: holy hell..
By someguy123 on 11/16/2011 8:09:13 PM , Rating: 2
I can see nvidia outpacing them in raw performance, especially since there's no window of release, if ever released.

They're touting x86 compatibility, though. With GPGPU you're porting/building new CUDA/openCL software. Theoretically this chip would work with current software, which is very enticing for developers and for those of us using slow adobe software.


RE: holy hell..
By Khato on 11/17/2011 12:08:59 AM , Rating: 2
As I stated above, it's quite important to differentiate between theoretical and actual performance. The Tesla M2090 has a peak theoretical throughput of 665 GFLOPS. From what I've found on NVIDIA's own presentations for their C2050, that theoretical throughput on the M2090 will likely go down to around 470 GFLOPS on the common DGEMM. The 1TFLOPS demonstration on Knights Corner was on DGEMM, it's not a theoretical max.


"We don't know how to make a $500 computer that's not a piece of junk." -- Apple CEO Steve Jobs














botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki