Forget the GPU, time for Knights Corner.

Message boards : Number crunching : Forget the GPU, time for Knights Corner.

To post messages, you must log in.

1 · 2 · Next

AuthorMessage
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 7483
Credit: 91,062,256
RAC: 46,492
Australia
Message 1171602 - Posted: 17 Nov 2011, 8:03:47 UTC


In 1997 1 Teraflop of computing power required 9,298 PII Xeon systems. In 2011 it can be done with one chip.

"Intel says that its MIC architecture benefits from the ability to run existing applications without the need to port the code to a new programming environment."

Knights Corner


Grant
Darwin NT

ID: 1171602 · Report as offensive
Treasurer

Send message
Joined: 13 Dec 05
Posts: 109
Credit: 1,569,762
RAC: 0
Germany
Message 1171606 - Posted: 17 Nov 2011, 8:49:54 UTC - in response to Message 1171602.  

Correct me, but isnt something you put in a PCIe slot and capable of 1TF called a HD4850? At least thats what my BOINC says after every start.


ID: 1171606 · Report as offensive
Profile jason_gee
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 24 Nov 06
Posts: 7235
Credit: 87,241,527
RAC: 5,612
Australia
Message 1171607 - Posted: 17 Nov 2011, 8:57:14 UTC - in response to Message 1171606.  
Last modified: 17 Nov 2011, 8:59:07 UTC

Correct me, but isnt something you put in a PCIe slot and capable of 1TF called a HD4850? At least thats what my BOINC says after every start.


The key difference vaguely mentioned in the article is they are talking double precision. My understanding is that both the nVidia & AMD/ATI Flops shown at Boinc startup are estimates of single precision peak flops, whereas the article also suggests sustained throughput (rather than peak theoretical) as well.

Jason
"It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change."
Charles Darwin

ID: 1171607 · Report as offensive
Treasurer

Send message
Joined: 13 Dec 05
Posts: 109
Credit: 1,569,762
RAC: 0
Germany
Message 1171608 - Posted: 17 Nov 2011, 9:04:55 UTC - in response to Message 1171607.  

Ok than maybe the HD5970(2× 464 DP) or HD6990(2× 637,5 DP) is more likely to beat this chip. Or are that also "only" peaks?


ID: 1171608 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 7483
Credit: 91,062,256
RAC: 46,492
Australia
Message 1171609 - Posted: 17 Nov 2011, 9:09:43 UTC - in response to Message 1171608.  


Unlike present GPUs which only have the one processing core, the Knights Corner is (initially) a 50 core device.

I'd be interested to see an article on it's actual architecture- i supsect it's a result of Intels's Larrabee project. Instead of it being used for graphics, they put it towards number crunching instead.


Grant
Darwin NT

ID: 1171609 · Report as offensive
Profile jason_gee
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 24 Nov 06
Posts: 7235
Credit: 87,241,527
RAC: 5,612
Australia
Message 1171610 - Posted: 17 Nov 2011, 9:16:31 UTC - in response to Message 1171608.  

Ok than maybe the HD5970(2× 464 DP) or HD6990(2× 637,5 DP) is more likely to beat this chip. Or are that also "only" peaks?


My understanding of the formula used in all cases is 'Peak theoretical', which can also be called 'Marketing Flops'. Of course whether the article's claims of sustained TeraFlop performance pan out to a similar sortof artificial number remains to be seen... I'd envisage a number of practical problems related to the speed of memory subsystems etc.

Jason.
"It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change."
Charles Darwin

ID: 1171610 · Report as offensive
P. Coutsos

Send message
Joined: 30 Sep 06
Posts: 17
Credit: 16,103,421
RAC: 6,883
Greece
Message 1171611 - Posted: 17 Nov 2011, 9:21:43 UTC - in response to Message 1171610.  

Have you seen the picture of the Intel-Guy (Dr. Rajeeb Hazra), where he holds the thing? Can you imagine how much power this device will consume? I guess it is 600W+...

ID: 1171611 · Report as offensive
Profile jason_gee
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 24 Nov 06
Posts: 7235
Credit: 87,241,527
RAC: 5,612
Australia
Message 1171612 - Posted: 17 Nov 2011, 9:26:04 UTC - in response to Message 1171611.  

Have you seen the picture of the Intel-Guy (Dr. Rajeeb Hazra), where he holds the thing? Can you imagine how much power this device will consume? I guess it is 600W+...


I was thinking either that's a very big Chip, or a very small man, LoL.

22nM tri-gate is supposed to be relatively cool & efficient, will be interesting to see. I also still doubt they would have fit it into the regular desktop 95 or 135 Watt TDP envelope though.

Jason

"It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change."
Charles Darwin

ID: 1171612 · Report as offensive
Team kizb

Send message
Joined: 8 Mar 01
Posts: 219
Credit: 3,709,162
RAC: 0
Germany
Message 1171619 - Posted: 17 Nov 2011, 11:28:09 UTC

So is this somthing that we'll be able to put in our PC in a couple years?


My Computers:
Blue Offline
Green Offline
Red Offline

ID: 1171619 · Report as offensive
Profile HAL9000
Volunteer tester
Avatar

Send message
Joined: 11 Sep 99
Posts: 6093
Credit: 155,190,353
RAC: 48,969
United States
Message 1171631 - Posted: 17 Nov 2011, 12:54:38 UTC

I have been wanting "co-processors" to stick in my computer for the past 15 years. All I have to say is that "it is about freaking time".

GPU computing is good, but why haven't they made generalized "compute" boards. Where they could strip off all of the video things that you don't need to processing data.

The previous generation of this they showed off was "Knight's Ferry" which had 32 cores and ran 128 threads. IIRC it sucked up around 250-300w. So you would need 2 8 pin PICe connectors for one of those cards. I'm not sure if it was using older 40 or 32nm tech. so the new one might use less power today.

This also seems like what they were doing with the GPU they were making a few years ago. Where they just shoved a bunch of low power x86 cores into a single chip.


SETI@home classic workunits: 93,865 CPU time: 863,447 hours

Join the BP6/VP6 User Group today!

ID: 1171631 · Report as offensive
Brkovip
Avatar

Send message
Joined: 18 May 99
Posts: 274
Credit: 144,414,367
RAC: 0
United States
Message 1171632 - Posted: 17 Nov 2011, 12:56:44 UTC - in response to Message 1171619.  

So is this somthing that we'll be able to put in our PC in a couple years?


If you can afford it. It might cost as much as your house when it comes out.

ID: 1171632 · Report as offensive
Profile MarkJProject Donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 08
Posts: 978
Credit: 33,084,424
RAC: 7,157
Australia
Message 1171633 - Posted: 17 Nov 2011, 13:11:18 UTC
Last modified: 17 Nov 2011, 13:23:03 UTC

I wonder if they will release it as a PCIe card? We could stuff these in a machine instead of a bunch of GPU's.

Does it support AVX? :-)

Some stuff from intel.com...

Standard programming models
A key advantage for developers using Intel MIC products is the ability to use standard, existing programming tools and methods. Intel MIC architecture combines many Intel CPU cores onto a single chip. These cores can be programmed using standard C, C++, and FORTRAN source code. The same program source code written for Intel Many Integrated Core products can be compiled and run on a standard Intel Xeon processor. The familiar programming model removes developer training barriers allowing the developer to focus on the problems rather than software engineering.

For example, CERN OpenLab, the European Organization for Nuclear Research, was able to use the Knights Ferry kit to migrate a complex benchmark written in C++ code to the new architecture in just a few days. According to Sverre Jarp, CTO of the CERN open lab, "The familiar hardware programming model allowed us to get the software running much faster than expected."

Creating a many core Intel architecture
The MIC project is the fruit of three research streams: The 80-core Tera-scale research chip program, the single-chip cloud computer initiative, and the Larrabee many-core visual computing project. The result is a fundamentally new architecture that uses the same tools, compilers, and libraries as the Intel Xeon processors. Since Intel processors are used in over 80% of the world’s supercomputers, programmers can continue to work in familiar territory when creating software for the MIC architecture.

Initial MIC products
The first MIC product is codenamed "Knights Corner," and development kits, codenamed "Knights Ferry," are already shipping to select software developers.

BOINC blog

ID: 1171633 · Report as offensive
.clair.

Send message
Joined: 4 Nov 04
Posts: 1300
Credit: 39,812,964
RAC: 24,885
United Kingdom
Message 1171790 - Posted: 18 Nov 2011, 1:18:50 UTC

Also on Rosetta@home number crunching thread Intel unveils 1 teraflop chip with 50-plus cores

And two links from that thread :-

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

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

The slides from the presentation give an idea of its low power consumption.
And may be available next year to those with a very fat wallet . . .

ID: 1171790 · Report as offensive
Profile MarkJProject Donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 08
Posts: 978
Credit: 33,084,424
RAC: 7,157
Australia
Message 1172128 - Posted: 19 Nov 2011, 6:52:43 UTC
Last modified: 19 Nov 2011, 6:59:13 UTC

Here is another link with pictures.

You'd sweare thats a nvidia reference card they are holding. Apparently the demo ones are based on 40nm chips and yes its a PCIe card.

We'd have to ask Intel if they could make one available for Jason so he can start coding :-)

I wonder what nvidia are going to do to respond? While their cards have great single-precision capabilities, they might try and increase the double-precision on upcoming chips. Not too sure what they can do about the programming model though.

ID: 1172128 · Report as offensive
Wembley
Volunteer tester
Avatar

Send message
Joined: 16 Sep 09
Posts: 429
Credit: 1,608,832
RAC: 1,235
United States
Message 1172134 - Posted: 19 Nov 2011, 8:03:47 UTC



Does this make you drool? :p

We may be obsolete soon. If the SETI lab had one or two of these in the server closet they could probably crunch all the WU's themselves without wasting all this bandwidth by sending the WU's to us.

Who wants to start the fund drive?


ID: 1172134 · Report as offensive
zoom314
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 56753
Credit: 40,741,083
RAC: 4,920
United States
Message 1172176 - Posted: 19 Nov 2011, 14:05:45 UTC - in response to Message 1172134.  

Heck what are those 8 cards?

Any specs?


Pluto is still a planet

Beep! Beep!

ID: 1172176 · Report as offensive
Profile MarkJProject Donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 08
Posts: 978
Credit: 33,084,424
RAC: 7,157
Australia
Message 1172188 - Posted: 19 Nov 2011, 14:45:25 UTC - in response to Message 1172176.  

Heck what are those 8 cards?

Any specs?


From what I gather 50 core x86 cards running at 1.8Ghz based on 40nm fabrication.

The promised release ones are meant to be 80 core based upon 22nm fabrication.

ID: 1172188 · Report as offensive
.clair.

Send message
Joined: 4 Nov 04
Posts: 1300
Credit: 39,812,964
RAC: 24,885
United Kingdom
Message 1172286 - Posted: 19 Nov 2011, 20:17:00 UTC - in response to Message 1172188.  

Heck what are those 8 cards?

Any specs?


From what I gather 50 core x86 cards running at 1.8Ghz based on 40nm fabrication.

The promised release ones are meant to be 80 core based upon 22nm fabrication.


Thinking about what a tesla card cost`s,
One thing is certain, they will put a big hole in anyone`s credit card.
and that setup must be on a PCIe expansion board, which cost a fair bit,
There is a heck of a lot of dollars worth in that photo.

ID: 1172286 · Report as offensive
zoom314
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 56753
Credit: 40,741,083
RAC: 4,920
United States
Message 1172291 - Posted: 19 Nov 2011, 20:26:26 UTC - in response to Message 1172286.  

Heck what are those 8 cards?

Any specs?


From what I gather 50 core x86 cards running at 1.8Ghz based on 40nm fabrication.

The promised release ones are meant to be 80 core based upon 22nm fabrication.


Thinking about what a tesla card cost`s,
One thing is certain, they will put a big hole in anyone`s credit card.
and that setup must be on a PCIe expansion board, which cost a fair bit,
There is a heck of a lot of dollars worth in that photo.

Oh yeah, that's sounds about right, the pic has unobtainium in It, I guess, for Me at least. But I have more important problems right now...
Pluto is still a planet

Beep! Beep!

ID: 1172291 · Report as offensive
Profile HAL9000
Volunteer tester
Avatar

Send message
Joined: 11 Sep 99
Posts: 6093
Credit: 155,190,353
RAC: 48,969
United States
Message 1172295 - Posted: 19 Nov 2011, 20:41:46 UTC - in response to Message 1172291.  
Last modified: 19 Nov 2011, 20:43:42 UTC

Heck what are those 8 cards?

Any specs?


From what I gather 50 core x86 cards running at 1.8Ghz based on 40nm fabrication.

The promised release ones are meant to be 80 core based upon 22nm fabrication.


Thinking about what a tesla card cost`s,
One thing is certain, they will put a big hole in anyone`s credit card.
and that setup must be on a PCIe expansion board, which cost a fair bit,
There is a heck of a lot of dollars worth in that photo.

Oh yeah, that's sounds about right, the pic has unobtainium in It, I guess, for Me at least. But I have more important problems right now...

Looks like each one of those boars has an 8pin and a 6pin going to it. For at least 225w to each card, or 1800w of 12v goodness.

The next question is who makes that power configuration/chassis. Could be a custom Intel job for show, but it is a good setup for the massive GPU guys as well.

Edit: Here is the whole article where that image came from.
http://www.thinq.co.uk/2011/6/20/intel-pushes-hpc-space-knights-corner
SETI@home classic workunits: 93,865 CPU time: 863,447 hours

Join the BP6/VP6 User Group today!

ID: 1172295 · Report as offensive
1 · 2 · Next

Message boards : Number crunching : Forget the GPU, time for Knights Corner.


 
©2016 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.