what could be done in realtime on a 16 TFLOPs workstation ?

[Guden Oden]

The companies expect that a one rack Cell processor-based workstation will reach a performance of 16 teraflops or trillions of floating point calculations per second.

They "expect" for this to happen, so there are no such workstations yet, maybe not even in a near future. Anyway, since when were workstations rackmounted? Sounds more like a mainframe to me. Why this liberal definition of what a workstation is anyway. Typically a workstation is a computer in an ordinary desktop case, so let's start calling server cabinets desktops. Hell, why not call them luggable laptops, that's even better!

 

[randycat99]

Floortop on coaster wheels?

 

[Entropy]

Numbers like 16 TeraFlops are of the Guaranteed Not To Be Exceeded variety. There is no way that workstation will be able to get even remotely close to such numbers even for the most suitable 0.1% of codes.

That doesn't mean it won't be a calculational monster. I hope it will be, and I hope it has tools that will allow it to be harnessed for scientific computation, and that IBM will market it towards the scientific community. But there has been no buzz about it, in any circles I know.

And for that to happen, i.e. for the machine to be much use as anything else than a glorified graphics ASIC, there are requirements that need to be fulfilled that typically don't come cheap or easily. There is a reasonable number of people around with an interest in high-performance computing, who would be very interested in what this CELL workstation is suggested to be. They typically have access to pretty deep pockets as well. I find it difficult to believe that these 16 TFlops are particularly useful for anything but graphics. It would be very interesting if they were though.

 

[chachi]

Numbers like 16 TeraFlops are of the Guaranteed Not To Be Exceeded variety. There is no way that workstation will be able to get even remotely close to such numbers even for the most suitable 0.1% of codes.

I guess it depends on what you consider a workstation is, and what a rack is. This picture shows BlueGene/L:

That's four sets of four racks. The second picture from the linked BlueGene/L webpage shows a closeup of one set of racks, in it you can see that each one holds 16 (2U?) blades.

Maybe some people are losing sight of the difference between what it's possible for Cell as a technology to do and what will eventually go into the console? For example BlueGene/L does ~70 Tflop/s, roughly twice as many as the previous #1, the Earth Simulator. All things being equal, a Cell equivalent of BlueGene/L would be 256 Tflop/s.

So what does that mean for PS3? It's hard to say without specifics, like how many Cell processors are in a blade on that rack? If they followed the BlueGene/L model and have 64 processors per blade with 16 2U-ish blades per rack, and a rack is 16 Tflop/s then it'd work out to about 16 Gflop/s per processor. We know that Cell is a multicore design, how many cores will they put in the PS3?

Of course that could be way off and it's really just some weird hyped up number that's nowhere near the real world performance. That'd be disappointing. It might make for a less interesting console spec but the impact on computing would be much greater the other way. IBM and Sony could make a lot more money selling that hardware to people who would happily pay much more than a few hundred bucks for it. A real 16 Tflop/s system would be in the top 20 of the list of most powerful supercomputers in the world.

 

[one]

So what does that mean for PS3? It's hard to say without specifics, like how many Cell processors are in a blade on that rack? If they followed the BlueGene/L model and have 64 processors per blade with 16 2U-ish blades per rack, and a rack is 16 Tflop/s then it'd work out to about 16 Gflop/s per processor. We know that Cell is a multicore design, how many cores will they put in the PS3?

The 16Tflops number of the Cell workstation is not by LINPACK benchmark, so you can't compare TOP500 supercomputers with it. Most likely it's a theoretical peak by a certain benchmark test which matches Cell WS's typical usage, rather than dated LINPACK test. Then the LINPACK benchmark is far from 'the real world performance' anyway, by 2 reasons - LINPACK itself is old and cluster-type supercomputers can't use processors efficiently.

 

[Wunderchu]

Re: what could be done in realtime on a 16 TFLOPs workstatio

I was just wondering what could be acomplished on the Cell based workstations that are coming down the road.

we have info that already, as expected, the first prototype Cell based workstations are already out, pushing 2 TFLOPs. call it late 2004.

by late 2005 or early 2006, or at the latest, mid 2006, I'd expect the 16 TFLOPs Cell workstations to be out. just think what could be done on one of those, assuming it has the proportional graphics processing capabilities to back it up. I mean an equally large amount of graphics processing resources, more than PS3 will have.

we've seen a little bit of what the ~97 GFLOP GSCube could do:
*1.2 billion flat shaded polys/sec.
*many tens of millions of fully featured polys/sec
*maybe upto 300 million semi featured polys/sec
(from what i recall of the guys that did realtime Antz demo)

a 16 TFLOP workstation has ~167x more theoretical FP speed than GSCube and over 2500x more than PS2.

I don't want to hype Cell too much, but it would be / will be exciting to see what can be done on those forthcoming 16 TFLOPs workstations. it's much more believable than say, distributed computing for realtime rendering over the internet.

It would be so incredibly cool if you could group consumer PS3s together to get the same performance as those workstations. going by what was done with those 60 or 70 PS2s in Illinois at UIC, I don't think what I am thinking of is too far fetched.



sorry, excuse this poorly written mess of a post...just thinking of the possibilities

whatever can be done, it should be quite amazing , IMO (apologies for stating the obvious )