Cell server maker promises pint-sized powerhouse
- Thread starter avaya
- Start date
Mintmaster
Veteran
I wonder how hard it is to get these Cell-based servers to work anywhere near their claimed rate for rendering. There's a lot of branching and non-streaming memory access (which is latency heavy) in ray-tracing, and Intel/AMD designers have taken great pains to minimize the effect of these things rather than increase throughput.
As for the other applications, I'm wondering if GPUs will be better at many of these parallelizable tasks, especially with ATI possibly setting a trend by putting more emphasis on math ability in R580, and also providing scatter ability for GPGPU coders.
I'm not saying I have all the answers, but it seems like there's a very small niche where Cell will truly excel.
As for the other applications, I'm wondering if GPUs will be better at many of these parallelizable tasks, especially with ATI possibly setting a trend by putting more emphasis on math ability in R580, and also providing scatter ability for GPGPU coders.
I'm not saying I have all the answers, but it seems like there's a very small niche where Cell will truly excel.
How about you finding a supercomputer that uses SP for actual supercomputing work?
This announcement is pure BS and marketing spin to make it sound like this these superduper TFLOPS are actually useful for supercomputing. It seems you are the only one on this board who actually WANTS to believe this marketing BS scrambling to its defense.
PC-Engine said:How about you finding a supercomputer that uses SP for actual supercomputing work?
This announcement is pure BS and marketing spin to make it sound like this these superduper TFLOPS are actually useful for supercomputing. It seems you are the only one on this board who actually WANTS to believe this marketing BS scrambling to its defense.
I think you're the only one who doesn't want to believe it. Too bad. Must be lonely in your corner.
There are a few weakness with CELL's SP floating point. First, it's not IEEE 754 compliant. Well, that's about all.
For some works, mostly signal processing and (perhaps) rendering, double precision or IEEE 754 compliance are not very important. For almost all other works like simulation, physics modeling, etc. double precision is very important. Full IEEE 754 compliance is good, but generally not a strict requirement. However, since different people have difference requirement on different IEEE 754 features, it's better to have full compliance.
CELL's DP is IEEE 754 compliant, but it's about 7x to 14x slower than its SP capacity.
For some works, mostly signal processing and (perhaps) rendering, double precision or IEEE 754 compliance are not very important. For almost all other works like simulation, physics modeling, etc. double precision is very important. Full IEEE 754 compliance is good, but generally not a strict requirement. However, since different people have difference requirement on different IEEE 754 features, it's better to have full compliance.
CELL's DP is IEEE 754 compliant, but it's about 7x to 14x slower than its SP capacity.
DudeMiester
Regular
Well didn't they say they were making a second version of the Cell where DP was only 1/2 as fast as SP? Even then say 10TFLOPs is insane for a single rack!
If the "2x slower DP version" CELL is used, it'd be more interesting.
Note that BlueGene/L's theoretical performance is 5.7TFLOPS per rack and it runs LINPACK at about 4.3TFLOPS per rack. I think a CELL based server would have lower efficiency (although probably has comparable or even higher raw performance).
Note that BlueGene/L's theoretical performance is 5.7TFLOPS per rack and it runs LINPACK at about 4.3TFLOPS per rack. I think a CELL based server would have lower efficiency (although probably has comparable or even higher raw performance).
The smaller the units are, the more cramped the chips will be i'd suppose. They'd need some rather extreme cooling i guess.
Very impressive indeed, whatever the haters say. But then again, it was expected that PC-Engine would come in to downplay this.
Sony could end world hunger and find a cure for AIDS and he still would downplay them.
Very impressive indeed, whatever the haters say. But then again, it was expected that PC-Engine would come in to downplay this.
Sony could end world hunger and find a cure for AIDS and he still would downplay them.
Cartoon Corpse
Veteran
life isn't all about flops you know. if you flip a flop, you're either correcting your course or being a coward, it's hard to tell these days. flip flops, now that's where it's at. mmmm beach.
chavvdarrr
Veteran
funny.
Pointing to an obvious Cell-related problem (ie full precision floating point ops speed), makes some people... angry. Whats wrong in opinting problems in existing Cell-design?
What do you think about Linus Torvalds who wrote thet " Cell looks like fish and is programmed like a cow" ? (copyright Douglas Adams )
Cell has unique characteristics, question is if +'s will outweight -'s
We'll wait and see. No need to start a holy crusade if someone does not believe that Cell is best thing mankind ever created ....
Pointing to an obvious Cell-related problem (ie full precision floating point ops speed), makes some people... angry. Whats wrong in opinting problems in existing Cell-design?
What do you think about Linus Torvalds who wrote thet " Cell looks like fish and is programmed like a cow" ? (copyright Douglas Adams
) Cell has unique characteristics, question is if +'s will outweight -'s
We'll wait and see. No need to start a holy crusade if someone does not believe that Cell is best thing mankind ever created ....
PC-Engine said:They were at a supercomputing event boasting TFLOPS of performance. I don't think they were there to play chess.
Boasting TFLOPS of single precision performance. Maybe they were trying to confuse some people, but I doubt anybody who is in the business of buying supercomputers is that easily confused.
I'm sure they weren't the only ones there showing off computers that couldn't compete in the supercomputer benchmarks. Technically speaking a computer isn't a supercomputer unless it was at some point (almost) the fastest in the world anyway.
Sometimes I just can't help it.
As long as a computer system has the DP FLOPS performance to be listed in the top500 at some point in time, it was/is a supercomputer.
The current slowest supercomputer is rated at 2.944 DP TFLOPS.
http://www.top500.org/system/7912
None of what Mercury Computer Systems showed exceeded 3 DP TFLOPS so it's not a supercomputer and until somebody builds one using their servers with enough performance to be placed in the top500, it will not be a supercomputer.
