Who will deliver the first 1TFLOPs retail GPU?

[Kanyamagufa]

Gotta go with nV...only if they've been more consistent and on target with their cycles. They just have a knack for releasing their products before AMD.

Fedex!

And for the record, that was brilliant.

 

[Techno+]

AMD once said that they are aiming for a 1 TFLOP CGPU for 2010 ( although the one TFLOP GPU will be out sooner).

It was something like this

3 GHZ
*
8 FLOPS\cycle
*
48
_________

1152 GFLOPS

So I voted for ATI since they seem to be taking this FLOPS thing more seriously

 

[Anarchist4000]

Well considering the missing MUL on G80 would it be wrong to say R580 has as much power as G80 and ATI is nearly a full generation ahead in the race? I'd consider it much more likely to break the barrier using a vector setup than scalar processors because of how compact they are.

 

[Razor1]

depends on how you look at it, theoretically yes the r580 has more horsepower then the g80, but utilization, efficiency wise, the g80 is better.

 

[no-X]

depends on how you look at it, theoretically yes the r580 has more horsepower then the g80, but utilization, efficiency wise, the g80 is better.

Depends how you evaluate efficiency. How fast would G80 be at R580's clocks? And don't forget, that G80 has 4x the texturing power (trilinear/anizo) of R580. G80 could be more effective in arithmetical operations, but R580 is more effective in texturing. But this only means, that G80's texturing engine is the same overkill as R580's pixel shading core

 

[Anarchist4000]

We're just arguing how many FLOPs each card has and who will reach 1 TFLOP first. Efficiency isn't really an issue unless the card is unable to utilize all that power given a best case scenario. Maybe we should define the question a bit more. Who will be the first to utilize 1 TFLOP while doing nothing but matrix*matrix operations or something along those lines? So the first to actually utilize 1 TFLOP in a synthetic benchmark.

 

[trinibwoy]

Depends how you evaluate efficiency.

IMO it makes sense to evaluate efficiency by perf/watt or actual perf/theoretical perf. In both cases G80 outdoes R580 if we use the GTS SKU as a point of comparison. It has a significant total flop deficit relative to R580 yet either beats it badly or is on par based on the workload. Not sure where you get that R580 is more effective at texturing.....

 

[no-X]

Not sure where you get that R580 is more effective at texturing.....

...actual perf/theoretical perf

performance: G80 is 60-70% faster on the average
theoretical trilinear fillrate: G80: 36,8 giga-samples/s - R580: 10,4 giga-samples/s => G80 is 250% faster theoretically, but only 70% practically, so even according to your definition of efficiency, R580 is more effective at texturing

 

[Arun]

performance: G80 is 60-70% faster on the average
theoretical trilinear fillrate: G80: 36,8 giga-samples/s - R580: 10,4 giga-samples/s => G80 is 250% faster theoretically, but only 70% practically

All you're proving IMO is that G80 is much more rarely bottlenecked by texturing, though - not that R580's texturing engine is more efficient...
I guess you could argue R580 has "too many ALUs" and the G80 has "too many TMUs", but I'd consider that a tad of an oversimplification; bottlenecks shift from game to game, and from frame to frame, after all...

 

[no-X]

You are right, of course, but this oversimplification isn't meaningless. All we wan't is fast graphic card which gives us balanced performance in majority of games. And comparision of theoretical parameters and practical results is the easiest way to find out, which graphics card is most balanced. As I remember, this way of comparision gives quite interesting results...

 

[Megadrive1988]

perhaps AMD-ATI's 1 billion transister GPU that Orton mentioned not long ago.

 

[trinibwoy]

performance: G80 is 60-70% faster on the average
theoretical trilinear fillrate: G80: 36,8 giga-samples/s - R580: 10,4 giga-samples/s => G80 is 250% faster theoretically, but only 70% practically, so even according to your definition of efficiency, R580 is more effective at texturing

Heh, efficiency and effectiveness are two different things. Effectiveness relates to simply getting the job done - not the efficiency involved in doing so. And why in the world would you isolate trilinear fillrate as some estimation of how performance would stack up in a real world gaming scenario? I'm sure if you use a synthetic trilinear or AF test you will see numbers closer to that 250% advantage.

 

[trinibwoy]

As I remember, this way of comparision gives quite interesting results...

I would remove clockspeed from there since architectures are so divergent now that performance/clock is pretty meaningless. Performance/watt is much more useful.

 

[Ailuros]

I think that Ailuros prob think wrt 4 FLOPs :

MADD = 2
MUL = 1
SF = 1

Total = 4 FLOPs.

Or he means that they have expanded the exta MUL into a MADD for a future G8x:

Dual-issue MADD = 2 + 2 (= 4 FLOPs),
(and he is ignoring the extra SF).

Yes for the latter; but purely speculative.

 

[phenix]

Who: Nvidia with G90
When: April 2008

 

[Oushi]

Who: Nvidia with G90
When: April 2008

Same Here

 

[no-X]

I would remove clockspeed from there since architectures are so divergent now that performance/clock is pretty meaningless. Performance/watt is much more useful.

here you are