Intel Introduces The Xeon Phi Product Family

H

HardOCP News

[H] News
Joined
Dec 31, 1969
Messages
0
During the International Supercomputing Conference in Hamburg, Germany, Intel highlighted the rapid adoption of the Intel® Xeon® processor E5 product family on the top500 list of most powerful supercomputers with 45 systems ranked on the list only 3 months after technology introduction.

The company also announced the new brand name for all products based on Intel® Many Integrated Core architecture - the Intel® Xeon® Phi™ product family - new technology that is aimed to bring industry leading performance and energy efficiency when processing highly parallel applications. Raj Hazra, Intel VP and general manager of High Performance Computing at Intel shares his thoughts here on how Intel® Xeon® Phi™ product family will help to accelerate the pace of discovery and innovation and pave the way to Exascale computing by 2018.
 
I don't get it, so this is like GPGPU processing from Intel?

But they don't have any even halfway decent GPU's?
 
Hmm, interesting...

fhjbzN.jpg
 
It's Intel's go at GPGPU sorta. This is what came out of Larrabee...It's a giant processor loaded with ">50" (from what I've read) slightly updated Pentium (original) cores. It's supposed to be a coprocessor that's already compatible with x86 code that just needs some tweaks.
 
Grebuloner said:
It's Intel's go at GPGPU sorta. This is what came out of Larrabee...It's a giant processor loaded with ">50" (from what I've read) slightly updated Pentium (original) cores. It's supposed to be a coprocessor that's already compatible with x86 code that just needs some tweaks.
Click to expand...

They did this with Larrabee and it was aborted so let's see how this works...

I don't see the purpose of selling x86 to HPC in a GPU format. It's not exactly a necessary ISA at that level.

It's GPGPU x86 style, basically.
 
So wait, if I'm reading this right... They're just using a modern process and stacking original Pentiums to brute force computational power in a GPU package?
 
techrat said:
So wait, if I'm reading this right... They're just using a modern process and stacking original Pentiums to brute force computational power in a GPU package?
Click to expand...

I'm not sure what the core architecture is but

The basis of the Intel MIC architecture is to leverage x86 legacy by creating a x86-compatible multiprocessor architecture that can utilise existing parallelization software tools.[16] Programming tools include OpenMP, OpenCL,[24] Intel Cilk Plus and specialised versions of Intel's Fortran, C++ and math libraries.[25]
Design elements inherited from the Larrabee project include x86 ISA, 512-bit SIMD units, coherent L2 cache, and ultra-wide ring bus connecting processors and memory.
The Knights Corner instruction set documentation is available from Intel.
Click to expand...

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

Basically another Larrabee but they're hoping with less fail. What I don't understand is why in the hell you want x86 at all when developers hated the original. This is just an Intel extension of "x86 everywhere" into places where it doesn't make sense. You're not exactly in need of legacy support in HPC. They've slipped a bit in market share to IBM who has regained ground over 2011 and first half of 2012.

1_Knights%20Corner%20specs_689w.jpg

http://www.brightsideofnews.com/new...nces-xeon-phi2c-50-core-22nm-coprocessor.aspx

Perf-per-watt and price are going to matter quite a bit. x86 legacy... I dunno if that's the same strong selling point it was several years ago.
 
techrat said:
So wait, if I'm reading this right... They're just using a modern process and stacking original Pentiums to brute force computational power in a GPU package?
Click to expand...

That seems right but can't be...how can 50 or so original Pentium processors = the power of 9000+ Pentium processors?
 
maboblivion said:
Bringing back coprocessors cause old is new!
Click to expand...

Funny thing... it's what Intel had to do to get back into the game after the double disaster of the P4 and the Itanium. They went back to the Pentium-M, which was based off the Pentium 3 and tweaked it. Core was born.

All that R&D and all they can do is retweak old ideas. Something's not right with Intel.
 
Am I missing something or can I add this to my computer for more cores?
 
JMCB said:
Am I missing something or can I add this to my computer for more cores?
Click to expand...

No.

1 - It's Workstation (high end), server and HPC crowd only meaning you'll likely get no support.

2 - It's likely expensive as hell.

3 - You probably can't buy it via retailer/OEM unless you buy a workstation/server or HPC.
 
MavericK96 said:
Haha, no kidding. Seems like they're doing alright "retweaking old ideas" to me.
Click to expand...

Pylon said:
They've still been crushing AMD for the last 5 years, so I think it's going on fine for them.
Click to expand...

Imagine where Intel could be if they weren't brute forcing old ideas to 'crush' AMD but actually brought new ideas to the table? AMD completely did away with CISC and brought incredible performance gains (and 64bit) before Intel had to scramble to catch up. Intel just seems more content to rest on their laurels whenever possible.
 
People seem to forget that stock factory core clocks are up against a wall, the scaling is outward....

If they have a platform which can scale threaded operations simply and efficiently it will be excellent... no need to write CUDA for basic threading....
 
techrat said:
Imagine where Intel could be if they weren't brute forcing old ideas to 'crush' AMD but actually brought new ideas to the table? AMD completely did away with CISC and brought incredible performance gains (and 64bit) before Intel had to scramble to catch up. Intel just seems more content to rest on their laurels whenever possible.
Click to expand...

Perhaps, but that's purely speculation. Who knows, maybe reworking old ideas turn out to be the best ideas? We don't and can't know really.
 
Meh, intel still playing catchup. ATI has had teraflop computing for years. Good chance you've already got that power in your computer right now.
 
MV75 said:
Meh, intel still playing catchup. ATI has had teraflop computing for years. Good chance you've already got that power in your computer right now.
Click to expand...

Not quite. 7970 is only ~950 GFlops double precision.
 
SRT8 said:
This is an Epic product. Good Job Intel!
Click to expand...

"Epic" as in "Extremely late to the game and more of a me-too release in attempt to salvage what was a miserable failure of an attempt to build a GPU", then I guess you would be right.
 
Where are you guys seeing that it's got a bunch of original Pentium chips in it? I read the published stuff I could find from Intel and it was pretty vague about what the core is based on.
 
SkribbelKat said:
Where are you guys seeing that it's got a bunch of original Pentium chips in it? I read the published stuff I could find from Intel and it was pretty vague about what the core is based on.
Click to expand...

https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)

TheWiki said:
Larrabee's x86 cores were based on the much simpler P54C Pentium design which is still being maintained for use in embedded applications.[12] The P54C-derived core is superscalar but does not include out-of-order execution, though it has been updated with modern features such as x86-64 support,[11] similar to the Bonnell microarchitecture used in Atom. In-order execution means lower performance for individual cores, but since they are smaller, more can fit on a single chip, increasing overall throughput. Execution is also more deterministic so instruction and task scheduling can be done by the compiler.
Click to expand...
 
This should be interesting...

If these are what they were supposed to be it should be ~50 cores on 22nm I believe?

I want to see folding benches..!
 
MavericK96 said:
Not quite. 7970 is only ~950 GFlops double precision.
Click to expand...

Depends on what type of FLOPS.
The HD4850 had 1TFLOP of single-precision.
I guess double precision is all the rage these days though, it's what the hip, kewl kids are doing now. :D
 
techrat said:
Imagine where Intel could be if they weren't brute forcing old ideas to 'crush' AMD but actually brought new ideas to the table? AMD completely did away with CISC and brought incredible performance gains (and 64bit) before Intel had to scramble to catch up. Intel just seems more content to rest on their laurels whenever possible.
Click to expand...
I can't say whether intel is lazy or not, but I don't think amd64 is a good example of how intel doesn't do anything. During this time intel was still refining and pushing their Itanium processors, which are arguably a far more radical deparature than what amd was doing with amd64, and furthermore was 64bit before amd clawhammer.

Obviously it wasn't very sucessful, but I think that's a different argument.
 
serpretetsky said:
I can't say whether intel is lazy or not, but I don't think amd64 is a good example of how intel doesn't do anything. During this time intel was still refining and pushing their Itanium processors, which are arguably a far more radical deparature than what amd was doing with amd64, and furthermore was 64bit before amd clawhammer.

Obviously it wasn't very sucessful, but I think that's a different argument.
Click to expand...

Itanium was HP's baby. HP had it nearly finished and then Intel was brought in to bolt on some bits and brand it because HP's executives didn't like how the companies previous server CPUs had done. And considering how Itanium has done in the market I think it's more an example of Intel beating a dead horse (because it's an Intel product so it's got to be good, right?) than an example of Intel producing anything new.

Don't get me wrong, Intel can be a fantastic company who comes up with innovative ideas it's just that most of those innovative ideas will end up killed on the bureaucratic ROI sword before they ever make it out of the company. Most of the stuff that's truly worthwhile ends up coming out of some dark corner of the company that they forgot was even there.
 
Red Falcon said:
Depends on what type of FLOPS.
The HD4850 had 1TFLOP of single-precision.
I guess double precision is all the rage these days though, it's what the hip, kewl kids are doing now. :D
Click to expand...

Well that's what it claims on the slide (double precision) so that's what I'm going with.
 
This thing still doesn't make sense...

It DOES use ORIGINAL Pentium cores...speed unknown but the P54C ran at 75, 90, or 100 MHz...so HOW exactly are 50 of these cores FASTER than the 1997 super computer that ran 9000+ of the EXACT same CPU?!

I mean, granted this stand alone card is using DDR5 RAM and I'm sure has a much faster pipeline linking the cores together, but still 50 of the same cores is faster than thousands of them? Hmm...
 
JoeUser said:
This thing still doesn't make sense...

It DOES use ORIGINAL Pentium cores...speed unknown but the P54C ran at 75, 90, or 100 MHz...so HOW exactly are 50 of these cores FASTER than the 1997 super computer that ran 9000+ of the EXACT same CPU?!

I mean, granted this stand alone card is using DDR5 RAM and I'm sure has a much faster pipeline linking the cores together, but still 50 of the same cores is faster than thousands of them? Hmm...
Click to expand...

Clearly there's more to it than just shoving 50+ original Pentium cores into one chip.
 
JoeUser said:
This thing still doesn't make sense...

It DOES use ORIGINAL Pentium cores...speed unknown but the P54C ran at 75, 90, or 100 MHz...so HOW exactly are 50 of these cores FASTER than the 1997 super computer that ran 9000+ of the EXACT same CPU?!

I mean, granted this stand alone card is using DDR5 RAM and I'm sure has a much faster pipeline linking the cores together, but still 50 of the same cores is faster than thousands of them? Hmm...
Click to expand...

Well, they are likely not the identical cores, probably just based on them, and then shrunk down to modern process sizes like 22nm, with a bunch of optimizations to make them work better using these smaller processes.

Compared to their original 0.8 micron (800nm) process, a 22nm version will produce a lot less heat and clock much higher.

If we factor in just clock speed improvements 50 p5 cores would have to be clocked at 11.8Ghz in order to keep up with 9000 66mhz p5 cores, so that doesn't explain it all, but combined with improvements and optimization, it brings it within the realm of possibility.
 
This product was hinted months ago if not a few years ago.

First, Intel announced Larrabee but scrapped it. Larrabee used many modfiied, shrunk-down Pentium cores and was intended to be Intel's version of a graphics card.

That didn't pan out.

Fast forward several months later, Intel announced their capability of squeezing an incredible number of cores onto a single CPU die. But, this was using the same Pentium cores or its equivalent when it was announced. It was reported here on [H] and I read it again on Engadget.com.

As I read over again what Larrabee was and what it was to become and looking at the Xeon Phi, this is either going to fail or succeed depending on who welcomes an x86-based GPGPU product. If it's anything like Larrabee, programs would have to be re-compiled to work on it, and it would require specific compilers made for the Phi. It could technically run an operating system and programs work on it. And, if it's like Larrabee, it's capable of processing OpenGL and DIrectX code on it.

According to the Larrabee article on Wikipedia, Larrabee had a multiple, superscalar P54C Pentiums working together. It also had a 512-bit vectoring engine and a 1024-bit internal ring bus.

All in all, I believe this is what happens when you take the supercomputer of yesteryear since they were nothing more than many processors running in parallel/symmetrically and shrunk them down, tied them down together onto a single die, and tweaked it to run modern programs.

The Xeon Phi is nothing more than a supercomputer in your hands in other words. Whereas Nvidia and AMD have powerful GPUs to do similar if not more than Xeon Phi already in the past several years, Xeon Phi seems to hope to cover GPGPU and HPC as well as general execution of server programs on one product. I know Nvidia and AMD have planned or are planning to add C++ capability and execution and CPU-like features into their own GPUs, this is pretty much in my opinion, Intel's version of what AMD/Nvidia were planning before. I believe Kepler and GCN architectures (which did add C++ capability into their GPUs) would be Phi's nearest competitors.

It's up to the markets that this product will be aimed at whether they want another GPGPU and HPC product, and are willing to re-compile programs and code for it to work on the Phi.

Another issue is cost. From the sounds of it, this is not going to be cheap and it will most likely never be available on the consumer level. I don't think this is going to be cheaper than Nvidia's Tesla products by any means. Educational, science and government institutions that have a need for GPGPU and HPC, or a very, very small (literally) supercomputer are the kind of people that this product is going to be aimed at.
 
You must log in or register to reply here.
Back
Top