Potential of the CELL Processor for Scientific Computing

AirRaid Mach 2.5 wrote:

> PDF article: http://www.cs.berkeley.edu/%7Esamw/projects/cell/CF06.pdf
>
> web article: http://www.hpcwire.com/hpc/671376.html

Interesting.
They state
" On average, Cell is eight times faster and at least eight times more
power efficient than current Opteron and Itanium processors, despite the
fact that Cell's peak double precision performance is fourteen times
slower than its peak single precision performance. If Cell were to
include at least one fully utilizable pipelined double precision
floating point unit, as proposed in their Cell+ implementation, these
speedups would easily double."

but rather than think about new HW, (which is vaporware), the Authors
could have also looked at ways of mixing the two precisions, for
Future work ?.

eg For that scientific SW that is convegence based, perhaps
a two step Software system, that uses the 14x faster 32 bit floats on
inner loops, and 64 bits on outer & final calculations, could also give
them the ~speed double ? - but on silicon they can actually get

-jg

Half the average Joe doesn't know what that means

Jim Granville wrote:
> AirRaid Mach 2.5 wrote:
>
>> PDF article: http://www.cs.berkeley.edu/%7Esamw/projects/cell/CF06.pdf
>>
>> web article: http://www.hpcwire.com/hpc/671376.html
>
> Interesting.
> They state
> " On average, Cell is eight times faster and at least eight times more
> power efficient than current Opteron and Itanium processors, despite the
> fact that Cell's peak double precision performance is fourteen times
> slower than its peak single precision performance. If Cell were to
> include at least one fully utilizable pipelined double precision
> floating point unit, as proposed in their Cell+ implementation, these
> speedups would easily double."
>
> but rather than think about new HW, (which is vaporware), the Authors
> could have also looked at ways of mixing the two precisions, for
> Future work ?.
>
> eg For that scientific SW that is convegence based, perhaps
> a two step Software system, that uses the 14x faster 32 bit floats on
> inner loops, and 64 bits on outer & final calculations, could also give
> them the ~speed double ? - but on silicon they can actually get
>
> -jg
>

"Jim Granville" <no.spam.TakeThisOut@designtools.co.nz> wrote in message
news:4477917e@clear.net.nz...
> AirRaid Mach 2.5 wrote:
>
>> PDF article: http://www.cs.berkeley.edu/%7Esamw/projects/cell/CF06.pdf
>>
>> web article: http://www.hpcwire.com/hpc/671376.html
>
> Interesting.
> They state
> " On average, Cell is eight times faster and at least eight times more
> power efficient than current Opteron and Itanium processors, despite the
> fact that Cell's peak double precision performance is fourteen times
> slower than its peak single precision performance. If Cell were to
> include at least one fully utilizable pipelined double precision floating
> point unit, as proposed in their Cell+ implementation, these speedups
> would easily double."
>
> but rather than think about new HW, (which is vaporware), the Authors
> could have also looked at ways of mixing the two precisions, for
> Future work ?.
>
> eg For that scientific SW that is convegence based, perhaps
> a two step Software system, that uses the 14x faster 32 bit floats on
> inner loops, and 64 bits on outer & final calculations, could also give
> them the ~speed double ? - but on silicon they can actually get
>
> -jg
>

by these numbers PS3 should be somewhere around 2-3x more powerful than
360???

I agree with you on the speculation of the tweaked config. Eventually at the
end of the day, it all comes down to the programmer writing code that takes
advantage of the hardware.

"Gayness" <gung.RemoveThis@hoe.com> wrote in message
news:e58dbk$9rr$1@enyo.uwa.edu.au...
> Half the average Joe doesn't know what that means

Yes, but any old idiot knows that 8.14 is better than 0.59 (not actual
figures, just for reference)


>
> Jim Granville wrote:
>> AirRaid Mach 2.5 wrote:
>>
>>> PDF article: http://www.cs.berkeley.edu/%7Esamw/projects/cell/CF06.pdf
>>>
>>> web article: http://www.hpcwire.com/hpc/671376.html
>>
>> Interesting.
>> They state
>> " On average, Cell is eight times faster and at least eight times more
>> power efficient than current Opteron and Itanium processors, despite the
>> fact that Cell's peak double precision performance is fourteen times
>> slower than its peak single precision performance. If Cell were to
>> include at least one fully utilizable pipelined double precision floating
>> point unit, as proposed in their Cell+ implementation, these speedups
>> would easily double."
>>
>> but rather than think about new HW, (which is vaporware), the Authors
>> could have also looked at ways of mixing the two precisions, for
>> Future work ?.
>>
>> eg For that scientific SW that is convegence based, perhaps
>> a two step Software system, that uses the 14x faster 32 bit floats on
>> inner loops, and 64 bits on outer & final calculations, could also give
>> them the ~speed double ? - but on silicon they can actually get
>>
>> -jg
>>

Brenden D. Chase wrote:

>
> "Jim Granville" <no.spam.RemoveThis@designtools.co.nz> wrote in message
> news:4477917e@clear.net.nz...
> > AirRaid Mach 2.5 wrote:
> >
> > > PDF article:
> > > http://www.cs.berkeley.edu/%7Esamw/projects/cell/CF06.pdf
> > >
> > > web article: http://www.hpcwire.com/hpc/671376.html
> >
> > Interesting.
> > They state
> > " On average, Cell is eight times faster and at least eight times
> > more power efficient than current Opteron and Itanium processors,
> > despite the fact that Cell's peak double precision performance is
> > fourteen times slower than its peak single precision performance.
> > If Cell were to include at least one fully utilizable pipelined
> > double precision floating point unit, as proposed in their Cell+
> > implementation, these speedups would easily double."
> >
> > but rather than think about new HW, (which is vaporware), the
> > Authors could have also looked at ways of mixing the two
> > precisions, for Future work ?.
> >
> > eg For that scientific SW that is convegence based, perhaps
> > a two step Software system, that uses the 14x faster 32 bit floats
> > on inner loops, and 64 bits on outer & final calculations, could
> > also give them the ~speed double ? - but on silicon they can
> > actually get
> >
> > -jg
> >
>
> by these numbers PS3 should be somewhere around 2-3x more powerful
> than 360???


It's atleast 5-10x more powerful depending on how you set up everything.
I said this, what, 2-3 years ago?

As with the PS2, the design of PS3 isn't just to be more powerful and
efficient, it is to be more flexible.

That article is on HPC which is only one algorithm advantage.
Look here at the advantage in graphics against other processors for Ray
Tracing. (It's 35x btw)

see table 13.
http://www-128.ibm.com/developerworks/power/library/pa-cellperf/

>
> I agree with you on the speculation of the tweaked config. Eventually
> at the end of the day, it all comes down to the programmer writing
> code that takes advantage of the hardware.


Floating Point hardware will always be faster than Non Floating Point
hardware at calculating floats now matter how good a programmer you
are. And considering all the top programmers will be working with Sony
instead of MS we should be seeing HUGE differences in technology
between the PS3 and Xbox360. Expect some titles to make the Xbox360
look like a Dreamcast.


--

Dinesh Manocha at UNC Chapel Hill hosted a workshop last week covering
"Computing at the Edge with Commodity Processors" (although, it'd be
hard to argue that Cell is currently "commodity" until the PS3 shows
up... unless one likes to pay the early adopter premium offered by IBM
and Mercury Computer Systems.)

Yes, Cell does have some interesting potential and companies like
Mercury are developing improved math libraries. Other interesting
developments include program synthesis packages (FFTW, ATLAS) that may
generate improved code given the specific input problem (Keshav
Pingali's research at Cornell), as well as "metaprogramming" packages
like RapidMind's current software (Mike McCool's startup.)

One of the more interesting questions is whether Cell really has
significant advantages over GPU-accelerated computations. Clearly Cell
offers more flexibility over GPU in terms of overall programming.
However, upgrading my GPU is easier than a forklift Cell upgrade. And,
Havok's game engine demonstrates respectable game physics computations
on the GPU, so why do Cell if GPUs can do the job?

OTOH, rumor has it that game engines haven't taken advantage of the
Cell SPEs and that is one of the larger causes of the PS3 release.


-scooter

Oh, yeah, I almost forgot: There are a lot of tricks that have to be
played to make the SPEs run and run fast. Communication with the SPEs
is exclusively via DMA transfers. It takes 275 cycles to set up DMA
to/from the SPE to off-chip memory, so latency hiding is a first class
process. The SPE's local store has to be aggresively managed (all 256K
of it), so, the developer has to be careful how DMAs are batched.

Basically, programming for the Cell SPE isn't trivial or pretty.


-scooter