[Beowulf] K Computer built for speed, not use

[Vincent Diepeveen]

On Oct 11, 2012, at 5:16 PM, Lux, Jim (337C) wrote:

> A good example of specialized computing would be the work that DE  
> Shaw is doing.  They're building specialized ASICs (or FPGA  
> programs) tailored for their chemistry applications, to get  
> multiple order of magnitude improvement over general purpose  
> solutions.  Clearly they think it's worthwhile: it's a high risk,  
> high reward sort of industry

Well it's a good example of what HPC is - taking a look into the  
future by spending some money on number crunching hardware.

With FPGA's they probably see further into the future than equivalent  
priced supercomputers can.Note that you need quite capable
hardware engineers with respect to FPGA's and especially with respect  
to memory controllers, in order to beat nowadays GPU's,
but that's undoubtfully what they do.

When having a good fpga design and i talked for prices of mass  
producing it, in itself printing a 1000 fpga chips is just 50 dollar  
a piece.
Then add $100 for each pci-card and some for SRAM on card, depending  
upon how much SRAM you want.

So in itself to crunch with fpga is cheaper than equivalent GPU's,  
provided you've got a capable network as probably the bandwidth
required is going to be huge in case of internode communication.

Of course the costs of the team and development cost of the capable  
FPGA is in the many millions of dollars a year.

Kind Regards,
Vincent

>
> Jim Lux
>
>
> -----Original Message-----
> From: beowulf-bounces at beowulf.org [mailto:beowulf- 
> bounces at beowulf.org] On Behalf Of Mark Hahn
> Sent: Wednesday, October 10, 2012 9:38 PM
> To: Beowulf Mailing List
> Subject: Re: [Beowulf] K Computer built for speed, not use
>
>> Any general purpose system will inevitably underperform for some
>> people, and many might argue that the art of managing such a  
>> project is
>> making sure everyone squawks equally loud about how the stake is  
>> being
>> driven into their heart.
>
> I think of it from the other direction: a specialized machine would  
> need to demonstrate really significant savings.  an astrophysics  
> machine might actually use many times fewer transistors than a  
> commodity processors (as one of Makino's slides shows), but could  
> it do anything else?  perhaps some forms of MD.  but making  
> transistors work harder is, today, not necessarily a winning  
> strategy, since managing heat is probably the dominant concern.
>
> if a domain somehow looks like it's on the verge of a breakthrough,  
> and specialized hardware is going to give a 100x advantage, well,  
> that sounds pretty good.  if a dedicated cosmology box is going to  
> bring only a 3x speedup and still be 1e6 away from resolving  
> important processes, well, I say "get in line for a general-purpose  
> account"...
>
> though this argument also hinges somewhat on the expectation of  
> demand being bursty _across_ domains.
>
> regards, mark hahn.
> _______________________________________________
> 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