[Beowulf] getting a phi, dammit.

[Dr Stuart Midgley]

Your guess is wrong.

Our experience with the Phi is streaks ahead of our experience with K10 (our code is single precision).  In the space of a few hours, you can have your codes running on the phi in native mode.  It is just a little computer sitting along.  ssh to it and your away.  You can get the entire GNU build system on it, run lustre or just about anything else that is on Linux.

It is VERY VERY cool.

You just can't do that with a K10/K20.  You have to completely re-architect your code/algorithm.

If your worried about cache-coherency, don't go near a K10/K20…  their memory model is completely screwed… and still has all the same short comings as the Phi from sitting on PCIE.

At least with the Phi, within a single system, its completely cache-coherent… all 60 cores (240 hardware threads).

If you want programming information, grab

    http://www.amazon.com/Intel-Xeon-Coprocessor-Performance-Programming/dp/0124104142


The future?  It won't be that long before you get a Phi as the main cpu in a system.  I suspect the biggest issue is getting enough fast memory (16 channel GDDR5 is expensive… and its hard to get lots of that near the cpu).

--
Dr Stuart Midgley
sdm900 at sdm900.com




On 28/02/2013, at 9:55 AM, Vincent Diepeveen <diep at xs4all.nl> wrote:

> 
> On Feb 27, 2013, at 10:01 PM, Mark Hahn wrote:
> 
>> 
>> thanks for all the replies.  so far, the impression I get is:
>> 
>> - Intel isn't interested in resellers, only a few OEMs.
>> - even OEMs are not really on top of things.
>> - there many claims of Phi being quite picky about PCIe and BIOS.
>> - OEMS generally want to sell you a $5k server with that $3k Phi.
>> - some OEMs offer surprisingly dense solutions (not necessarily
>>   a good thing...)
>> 
>> I find the PCIe/BIOS stuff quite dubious, though I suppose Phi is  
>> still
>> elatively immature.  maybe the next version will be the one that  
>> will be
>> generally applicable.  I suspect by contrast that I can buy a K20 card
>> and drop it into most any gen1/2/3 PCIe x16 slot and it'll work  
>> (assuming
>> I provide the right power and cooling, of course.)
>> 
>> all this seems rather surprising to me.  I know HPC isn't the center
>> of the universe, but to me it looks like Intel is executing Phi  
>> poorly.
> 
> Just my guess:
> 
> They just can't compete with the K20. The way intel chooses to  
> compete is launch
> their thing long before they can deliver it. We all remember how they  
> perfectly played
> that game with itanium2 as well.
> 
> By the time the news 'drips' through to the public about the  
> successtories of their fanboys,
> carefully selected, they might be able to produce it cheap enough to  
> deliver it in lager quantities i guess.
> 
> Let's face the additional production problems you have with a cache  
> coherent CPU
> which practically will see action in the same manner as how you'd  
> operate a gpu.
> 
> Production price of it must be immense higher than of K20.
> 
> If it takes this long to produce the Xeon Phi with this much delay -  
> just think of the next generation
> of it. Will it be cache coherent?
> 
> If not - then you can throw away a bunch of codes you write for Xeon  
> Phi now.
> 
> So what's the long term future of it?
> 
> A new architecture? Called Xeon GPU?
> 
>> NVidia has a huge headstart, and frankly, AMD's HSA stuff seems a lot
>> more coherent and thought out (though obviously still vapor-phase.)
>> 
>> _______________________________________________
>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin  
>> Computing
>> To change your subscription (digest mode or unsubscribe) visit  
>> http://www.beowulf.org/mailman/listinfo/beowulf
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf