How Important are FLOPS to Gaming Performance?

Benq's post makes more sense. (comparision logic)

 

The comparison is flawed - for starters, he's comparing all of Xbox's flops, CPU+GPU, with just PS2's CPU flops. But regardless of that anyway, they were so different systems as to make such comparison very difficult. PS3 and X360 are more similar than that, architecturally e.g. a FULL gpu in both etc - and the GPU is what makes the biggest contribution to peak theoretical floating point figures. You can't directly map between the PS2/Xbox floating point situation and the one that exists with PS3 and X360.

 

It is not flawed in the sense PS2 didnt have a GPU like you said yourself so there is no extra flops that PS2 can take advantage of, yet it keeps very nicely with Xbox & Gamecube.

The point that BenQ was trying to make is the flop performance to console power is not like mph to cars.

 

The GS is there..however you want to count its floating point performance, I'm sure the number would be non-neglible. When counting current GPU floating point performance, manufacturers don't just look at vertex work which is what the EE was doing.

I think it'd be difficult to come up with any analogy that would perfectly satisfy, at this level.

 

It's flawed in a sense that he's comparing fixed floating point (for the most part) performance with programmable floating point performance.

 

The idea is that if a console's performance can be expressed in a single figure, namely FLOPS, and the relative performance of those consoles can be judged by that and their games, it presents the idea that reall the next-gen will be next to identical.

Summing up ALL the PS2's and XB's float ops, it would seem XB had up to 13x the power, but the difference between systems wasn't so great. Then summing up ALL the PS3's and XB360's FLOPS, if the numbers given are accurate, 1 TFLop vs 2 TFlop is only 2x the power difference, which going by current gen means negligable.

The logic is sound. Therefore I would be inclined to consider that the information this premise is founded on is wrong. I'd be surprised if 13x the number crunching performance couldn't achieve more over PS2 than XB managed. I would have to question the validity of that GPU flop count perhaps, or maybe those are peak Flops rates that XB could never get close to due to BW limits? That is perhaps the most likely explanation from what I hear of XB architecture.

I would say that if one machine can access 2x the flop performance, regardless of peak figures, over it's rival, it should shine through that that is the case in it's games.

 

I agree with what you are saying, Shifty. The topic title is a classic question, but the discussion was predicated on an entirely faulty basis from the start- that CPU FLOPS and GPU FLOPS can just be added together seamlessly and interchangeably.

 

that's true but it's exactly the same as the E3 commentary that had people after the show claiming that PS3 is indisputably 2x more powerful than X360.

 

People do a lot of stupid things. If you know better, then you shouldn't do the same stupid things.

Now if you break down comparisons to CPU to CPU and GPU to GPU, and still come out to "2x more powerful", maybe that's worthwhile to toot your horn a bit.

 

good point

 

I can't see how one can predict what the end result will be simply from looking at the max FLOP ratings of the two systems. Considering that the method to achieving this performance is substantially different in both systems, wouldn't it be more important to see what perfomance can be achieved and sustained in different situations? The FLOPS advantage can be minimized in certain cases by making good use of the strenghts of the architecture. It seems to me that Xenon+C1 were designed with eachother in mind, more so than Cell+RSX. This tighter integration may help it overcome some of this FLOPS advantage. On the other hand, the 7 SPEs in Cell can also be put to good use in certain types of applications where brute force can help. I don't think it's as simple as saying one console is 2X the power of the other. This would imply that PS3 has the same capabilities of the X360, and at the same efficiency, which is not the case. Many details for both haven't been made public yet, so it's hard to see what level of performance can be expected from each.

 

how dare you bring common sense into a discussion.

 

How is Xenon+c1 designed "more so" with each other in mind? Particularly, in what unique way will they be working in tandem?

 

Considering that RSX was "Plan B" as far as a GPU for the PS3, I don't see how Cell could have been designed with the RSX in mind. This doesn't mean that nVidia can't design RSX to take advantage of Cell, which I sure hope they will...

Two specific examples of how I see Xenon+C1 working together is through the use of the VMX units (with their custom instructions) and the GTU and MEMEXPORT functionality built into C1 (as mentioned in Dave's article). Don't get me wrong though, I'm not saying Cell+RSX will not be able to accomplish similar functions. It just seems it might be done in more of a "brute force" manner... Like I stated above, until we know for sure what nVidia is including in RSX, we can't know exactly how elegant/efficient (or not) the architecture will be...

 

Of course, you sound unnecessarily pessimistic about it.

Regarding the "MEMEXPORT functionality", is it not just xenon's ability to directly access system RAM? If that were to be a "unique" functionality, I would suggest that in contrast, RSX is said to be able to access its own DDR with a dedicated bandwidth, as well as XDR directly.

However, I might be able to learn from you how custom VMX units would be able help xenon + c1 work together. Would you be able to clarify?

 

I was under the impression that the PS2 had a comparitively horrible GPU (compared to Xbox and Gamecube) and it was the FLOPS power of the CPU that allowed it to compete with the other systems in terms of graphics. Am I wrong?

 

Not being pessimistic, unless I missed something in Cell architecture (which was specifically designed for RSX), then I have no reason to believe otherwise. Like I said, I'm sure nVidia's not just simply "throwing something together" and calling it a day. In the end, Cell+RSX will work well together...

Seems like you are over-simplifying what MEMEXPORT can do.
Take a look at the article and you will see the difference...
And yes as I stated, Cell+RSX may be able to accomplish something similar in their own way.

There's been some discussion on this already. Just do a quick search. Plus, it seems we're veering the thread off topic...

 

It is not an over-simplification if that is in fact, what it does. Perhaps this
is off topic, although it is somewhat related to what you've brought to the table - your suggestion that xenon and c1 will work better together, beyond the FLOPs race.

 

http://www.beyond3d.com/articles/xenos/index.php?p=10