• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Xeon Phi discussion

colinstu

2[H]4U
Joined
Oct 11, 2007
Messages
3,563
I'm sure most of you have been hearing the random news about these pretty neat sounding boards Intel will be selling.

Looking online some more, sadly it looks like these will be going for $2000 a pop if not more.

http://www.intel.com/content/www/us...h-performance-xeon-phi-coprocessor-brief.html

Each is 60 real cores, however there are 4 threads for each core, which equates to 240 threads. Clocked at a little over 1GHz. I think I remember hearing that each core is pretty similar to the Core 2 architecture, not exactly certain. These are x86_64 based though so no need for CUDA/any of that kinda stuff.

What I'm really curious is if someone loaded up a 4 slot PCIe mobo with 4 of these cards. Theoretically wouldn't F@H be able to run over all of these cards at once? Someone did that linux command to see info about the processor and all the cores showed up through there... just like a traditional processor. The link I have above says that "up to 8 coprocessors per host server" ...hmm. Each has about a 300w TDP too.

Anyways... expensive food for thought... really wish someone (a review website or company) would try out folding on one sometime. Should be neat!
 
If I get one... it is on my list of things to try. Believe me... at the Intel/ TACC press event I was thinking it.
 
From what I understood these things essentially ran a tiny linux VM on the card which has access to all the CPU's. So multiple cards in a system would just have multiple VM's, not onr big one. I could be wrong though.

I mean you can get a teraflop on one of these babies, and if they keep making them we could see one with 128CPU's before too long... The fact they are launching this on 22nm means they are serious about this product, though.

I can only imagine they would be pretty stinkin fast. Now, to properly use them a new core would be required as they have really wide sse units on them that I believe use new/different instructions, and to get max throughput you need to use them.

8GB of GDDR5 w/ 320GB/sec bandwidth available. That's just sick.
 
Last edited:
The article I read stated they would not be sold comercially.
Sounds like the only way to personally own one is to wait for it to show up on the 'Bay.
Though I could have misinterpreted the article.
 
Looking online some more, sadly it looks like these will be going for $2000 a pop if not more.

Directly from Intel's e-mail:
The Intel Xeon Phi coprocessor 3100 product family will be available during the first half of 2013 with recommended customer price below $2,000.

Also @Core32 - If you cannot purchase them at Newegg/ Amazon or somewhere similar, not a hug deal in this case. Likely you would want the 'A' version for active cooling anyway. Big difference is that you buy this type of thing from a VAR.
 
Directly from Intel's e-mail:


Also @Core32 - If you cannot purchase them at Newegg/ Amazon or somewhere similar, not a hug deal in this case. Likely you would want the 'A' version for active cooling anyway. Big difference is that you buy this type of thing from a VAR.

If the shrouds can be taken off w/o messing with the warranty, I think people here would rather mod in some 120mm fans themselves or watercooling :D;)

And if a VM can stretch over multiple coprocessor cards, then I see people buying 7-8slot PCIe mobos and some flexible PCIe x16 extenders ;) (unless it's found out FAH doesn't scale that well over 1 or 2 of these)
 
If the shrouds can be taken off w/o messing with the warranty, I think people here would rather mod in some 120mm fans themselves or watercooling :D;)

And if a VM can stretch over multiple coprocessor cards, then I see people buying 7-8slot PCIe mobos and some flexible PCIe x16 extenders ;) (unless it's found out FAH doesn't scale that well over 1 or 2 of these)

I am sure Intel has these suckers locked down, but imagine if you could overclock them... I bet you could get to the 1.2-1.3Ghz range with a little bit of voltage and good cooling. That would friggin haul ass!
 
Anand just posted something too.

j20Ns.png


Max # of cores to be disclosed during 3100 series launch... hmm. Well seeing that it seems like each core gets 0.5MB of cache, think it's safe to say those will feature a max of 57 cores. Yay math..
 
Max # of cores to be disclosed during 3100 series launch... hmm. Well seeing that it seems like each core gets 0.5MB of cache, think it's safe to say those will feature a max of 57 cores. Yay math..

But the interesting thing is that TDP... quite a bit higher, less cores, less RAM, so it must be quite a bit higher clock speed!
 
pj said "Yea but the guy who wrote that is not too bright ;-)" before... thanks emailed subscription notifications ;)

:) both may be true :-P

Also, you are right on with the 57 core count thing. 28.5/.5. Clocks are the big question.
 
How do you passively cool 300 watts?!?

Let's put it this way... not fully operational, but I had earplugs in. Still loud.

Actually 300w is not THAT bad. Figure it is like 2x Xeon E5's or Opteron 16 cores. The units are heavy though.

And thanks Pocatello!
 
This article was a good subject of conversation at work today. Really glad I knew about it beforehand. :)
 
I didn't think folding was even going to be possible on these cards. Lack of MMX, SSE, and AVX portions of the cpu...
 
I didn't think folding was even going to be possible on these cards. Lack of MMX, SSE, and AVX portions of the cpu...

Yeah, it will likely require a new core, so don't count on it

good work pj
 
How do you passively cool 300 watts?!?

You don't. The passively cooled cards are expected to be put in servers that will have much forced airflow through them, so they will, in-effect, be actively cooled.
 
Going to post a few points not in the Xeon Phi piece at Tom's on STH. That was the initial piece but I will go through the photo archive.

Reading through your comments on there... interesting that you have to SSH into those cards to access them. Looks like my (and others around here) dreams are ruined :p (dreams of running loads over multiple cards?...unless that actually is possible).
 
Reading through your comments on there... interesting that you have to SSH into those cards to access them. Looks like my (and others around here) dreams are ruined :p (dreams of running loads over multiple cards?...unless that actually is possible).

If each node is running a form of linux or a VM, someone will figure that out. But with the general lack of access to the community, most likely not in the near future.
 
"Xeon Phi 5110P" positioning high-end, 60 core, clocked at 1.053GHz, double-precision floating point performance 1.011TFlops, an instruction / data cache 1.875MB (per core 32KB), secondary cache 30MB (512KB per core) , equipped with sixteen channels ,512-bit bit wide the 8GB GDDR5 memory, the equivalent frequency of 5GHz, bandwidth of 320GB / s, and support for ECC. The entire card thermal design power of 225W and the passive cooling in a total of 16 on the 3213 count the Friends of molinxx two HPC Phi Each has Xeon Phi 5110P, 60C240T; x; including three Xeon Phi 5110P successfully received the 8102 BA16 package linux without any optimization, 240T ran out of the 720,000 + PPD optimization estimation performance 4P E5-4650 machine a higher low, a Xeon Phi 5110P ran less than 300 watts of power consumption in the hand is close to one hundred million of PPD will become a reality; cheer2; http://www.equn.com/forum/thread-35895-1-1.html


If that is legit... 720K PPD on 300W is absolutely MASSIVE!
I wonder what they could do if there was some optimization for those 512bit wide vector units too...

So, thats basically like 2x 4P opteron boxes in a single card?!? Even at $4000, it's ~somewhat price competitive!
 
Yeah, these things might actually be a viable option. If you can find where to buy one, that is.
 
So, thats basically like 2x 4P opteron boxes in a single card?!? Even at $4000, it's ~somewhat price competitive!

Less than one high-end 4p MC box - my worst one does around 725K ppd on 8102s. "Unoptimized" is intriguing, and 300W is definitely intriguing, and even the $2700 I saw somewhere is not way out of line.
 
Ah, I see, I though 4p amd boxes were doing a lot less, maybe I was thinking of 8101's. But, yeah, it is still pretty sweet.

So, who's buying one? :D
 
Ah, I see, I though 4p amd boxes were doing a lot less, maybe I was thinking of 8101's. But, yeah, it is still pretty sweet.

So, who's buying one? :D

Isn't the low-end ppd for a 4p g34 box to be around 300k-350k ppd? 4x low clocked 8 core chips folding a nasty WU right? (possibly not benefiting from quad channel & a proper linux configuration too)
 
Isn't the low-end ppd for a 4p g34 box to be around 300k-350k ppd? 4x low clocked 8 core chips folding a nasty WU right? (possibly not benefiting from quad channel & a proper linux configuration too)

That is probably right - I'm just saying that reported ppd is in line with a single pretty high end 4p G34 system. To build a new one from scratch that would produce >700K ppd (8102s), your cost would be well over $2700. So, the price seems right in line with current high-end stuff.

Find me some and I'll test them out.
 
Xeon Phi does not support SSE or SSE2. Nice try, though.
 
Back
Top