The new coprocessors, not enough bang for the buck?

[vitchilo]

So I was looking at the news and I saw that Intel was launching it's new Xeon Phi line of products and I was thinking about building the ultimate crunch rig... so I looked up at the numbers....

AMD Firepro S10 000
5.91 TFLOPs in single precision
1.48 TFLOPs in double precision
$3,600
$609/TFLOP in single precision
$2,432/TFLOP in double precision

(I didn't put all NVIDIA's products since all the others are ridiculous high priced compared to what they offer TFLOP-wise same goes for Intel)
NVIDIA K10
4.58 TFLOPs in single precision
0.19 TFLOPs in double precision
$2,500
$545/TFLOP in single precision
$13,157/TFLOP in double precision

Intel Xeon Phi 3100 (special version for computing coming out later in 2013 not fully announced yet)
Over 1 TFLOP in double precision (let's say 1.1TFLOP)
No info on single precision
Less than $2,000 (let's say $1,800)
$1,636/TFLOP in double precision

4x Radeon 7970 GHZ edition
15.16 TFLOPs in single precision
3.78 TFLOPs in double precision
$1,556
$102/TFLOP in single precision
$411/TFLOP in double precision

4x NVIDIA GTX680
12.36 TFLOPs in single precision
$1,720
$139/TFLOP in single precision

It's obvious that the 7970 are way cheaper than anything else... for GPU crunching.

For CPU crunching, the Xeon Phi is much cheaper than a bunch of i7, xeons or Opteron rigs considering a i7-980 at 4.5 GHZ is around 95 GFLOPs.

So as long as crunching requires GPUs, GPUs are way cheaper than anything NVIDIA, AMD or Intel are offering for the moment.

If CPU crunching is your thing, the upcoming Intel Xeon Phi 3100 will be the best bang for the buck. And when they come out, I might buy a bunch of them...

[Buckeye]

I don't think we will be able to buy them for some time, they are all sold out and not available to the general public anyway... yet.

[Snow Crash]

The professional cards from AMD, NVIDIA, or Intel (assuming you can even get them) are not worth the money for use on any of the BOINC projects.

The Intel cards, besides not available to the public (yet) will still need the science apps we use for crunching to be recompiled and I doubt any BOINC project will bother as there will likely be some number less than 10 people running them on volunteer projects.

I'm not saying it wouldn't be awesome if the project scientists did code what we want and the hardware vendors provided pro cards a decent prices but neither is likely in the near future. I would suggest that once you pick a project, find out what hardware runs best for that project ... it is not the same for all of them.

[vitchilo]

Really? Coprocessors couldn't crunch normal x86 data??? Is that sure or it's a guess?

[Snow Crash]

While the Intel Xeon Phi are x86 cores ... they are simplified x86 cores which don't have all of the same functionality as CPU ... here is an example ...

http://software.intel.com/sites/defa...27364001en.pdf
page 659

In 64 bit mode, the Intel? Xeon Phi™ coprocessor supports a subset of the Intel 64 Architecture instructions.
The following summarizes Intel 64 Architecture instructions that are not supported in the Intel? Xeon Phi™
coprocessor:
• Instructions that operate on MMX registers
• Instructions that operate on XMM registers
• Instructions that operate on YMM registers
GPR and X87 Instructions Not Supported in the Intel? Xeon Phi™ coprocessor
CMOV CMPXCHG16B FCMOVcc FCOMI
FCOMIP FUCOMI FUCOMIP IN
INS INSB INSD INSW
MONITOR MWAIT OUT OUTS
OUTSB OUTSD OUTSW PAUSE
SYSENTER SYSEXIT

[Movieman]

Just asked to borrow one from a certain company where I know a few people..Cross your fingers

[BlindFreddie]

While the Intel Xeon Phi are x86 cores ... they are simplified x86 cores which don't have all of the same functionality as CPU ... here is an example ... ...

AFAIK, very few or no WCG science programs use any of the esoteric x86 instructions, including any SIMD (SSE etc) ones, and they may run with the x86 architecture provided by the Xeon Phi.
The Intel blurbware for the Xeon Phi claims that programmers won't have to rewrite their x86 code for it. ( Introducing Intel? Xeon Phi? Coprocessor )

I think the biggest difficulty would be to get the O/S and BOINC client to treat the X-Phi cores as slow but ordinary cores that can run ordinary tasks with standard O/S file I/O facilities, plus (optional) console screen access for the BOINC tasks' graphics.
Maybe the BOINC developers and/or Intel and/or some Linux hackers will get something going that can achieve this. BOINC may be a big enough chunk of parallel computing and be a big enough chunk of Intel's target market for the X-Phi for Intel to push this.
Now where's that good Doctor when we need him?

[[XC] Oj101]

X-Phi

Made by Creative? Just remember that GLOPS is a theoretical figure and you can't really do a straight comparison. It's almost akin to saying a Pentium 4 Extreme Edition 3.73 GHz will be faster than a Core i7-3770K because it runs at a higher speed (ok, ignoring Turbo).