No please look them up, I'll wait
I argue (and will continue to do so) that die sizes don't dictate financial performance, get your memories straight. Why is there so much anger in this thread anyway, sheesh....
No please look them up, I'll wait
I argue (and will continue to do so) that die sizes don't dictate financial performance, get your memories straight. Why is there so much anger in this thread anyway, sheesh....
There is no difference.
What about power consumption?
*Runs away* (Sorry I'm missing digi
and trying to make up for his absence)Ahh yes well I found this where you attack die size relevancy sounding quite a different tune than now to be sure, not going to waste any more time searching: http://forum.beyond3d.com/showpost.php?p=1607087&postcount=1559
It certainly gives you enough wiggle room I'm sure. So have at it.
It's just so funny and predictable to me though how die size never mattered, and anyone suggesting as much was rebutted, to some when AMD had the edge, but suddenly it's of paramount importance again. But, such is life.
BTW, back in that post you were arguing all about how Nvidia was better because even though their die size was kinda shite, they got more performance per theoretical flop and after all that was all that really mattered.
Bet you dont dedicate much posting effort to that argument going forward, eh?
Of course they dont, but they do dictate chip cost to a significant degree, a point you continually tried to imply wasn't true through shady arguments ("since nvidia was more profitable just look at their financials, obviously I now imply die size has no bearing on cost!").
It certainly gives you enough wiggle room I'm sure. So have at it.
It's just so funny and predictable to me though how die size never mattered, and anyone suggesting as much was rebutted, to some when AMD had the edge, but suddenly it's of paramount importance again. But, such is life.
BTW, back in that post you were arguing all about how Nvidia was better because even though their die size was kinda shite, they got more performance per theoretical flop and after all that was all that really mattered.
Bet you dont dedicate much posting effort to that argument going forward, eh?
Of course they dont, but they do dictate chip cost to a significant degree, a point you continually tried to imply wasn't true through shady arguments ("since nvidia was more profitable just look at their financials, obviously I now imply die size has no bearing on cost!").
You're not going to like this, but parameters such as # of ROPs etc. are probably more useful to estimate performance than area size. It is very hard to argue against this, otherwise Nvidia would have destroyed AMD in the past generations. So, yes, die size is most irrelevant wrt performance in the face of major architectural differences. But he doesn't say that it's not important for other reasons: of course it is, since it determines GMs. The question is: can you compensate this efficiency loss/higher base cost one way or the other. Nvidia has historically been able to do that in a very impressive way, AMD has not. This makes it less painful for Nvidia to be at a perf/mm2 disadvantage than AMD. It all depends on what part you're discussion.Rangers said:
What's interesting this time around is that perf/mm2 seems to be very close between the 2 parties, after correcting for compute features of Tahiti. We already know what the changes in GCN, let's hope we'll also get more info from Nvidia.
It's also interesting that Kepler still seems to be performing better per number of units (# shaders, MCs, etc) than GCN, as has always been the case in the past, despite obvious major changes. Would love to see more insight into that.
Last edited by a moderator:
Well yes I have large amounts of wiggle room because that statement doesn't at all contradict anything I've said in this conversation. I commented on the "improvement" in perf/mm from Fermi->Kepler and GK104's positioning as a result.
The "compute burden" argument can be made for individual SKUs like Tahiti or GF110 but it's quite clear that architecturally Kepler is leaps and bounds above Fermi in the perf/mm and perf/w metrics. In other words, a big, power hungry, compute encumbered Kepler chip should still be impressively faster than GF110 given what we know about GK104. Yet you couldn't say it's going to be x% faster than GK104 just because it's x% bigger - that's the argument I was making in that post you linked.
All meta-discussions aside: you couldn't be more wrong about this. Architecture probably determines 80% of the final area, if not more. Layout/ram density and other low level optimizations can only help so much. A pure VLIW architecture without operand crossbar in front of the ALU (as often seen in DSPs) will kill an architecture that doesn't place such constraints on the compiler and the programmer, in terms of area.Albuquerque said:
As for tangible differences for the user: larger area means higher leakage.
You did care about perf/W right?
All meta-discussions aside: you couldn't be more wrong about this. Architecture probably determines 80% of the final area, if not more. Layout/ram density and other low level optimizations can only help so much. A pure VLIW architecture without operand crossbar in front of the ALU (as often seen in DSPs) will kill an architecture that doesn't place such constraints on the compiler and the programmer, in terms of area.
As for tangible differences for the user: larger area means higher leakage.
None of which has any bearing on perf per millimeter. I'm back to the same argument I started with, while your goalposts have moved four times.
Again, I'm now done with this OT segment, so please carry on.
Jawed
Legend
A more interesting question is "why didn't NVidia do it this way from day 1, i.e. G80".
leoneazzurro
Regular
I don't assume that GK104 is the flagship and never said that (I only wonder when this big chip will see the light). You are assuming instead that GK104 will beat soundly 7970 across all the board (and not in selected benchmarks à la HAWX
) "by 10%" and also that the 7970 stock is all AMD is capable of with this generation of chips.
Last edited by a moderator:
Like with everything else in life there's a lot to be said for experience.
I'm sure we could ask the very same question about Kepler 6 years from now. I don't assume that GK104 is the flagship and never said that (I only wonder when this big chip will see the light). You are assuming instead that GK104 will beat soundly 7970 across all the board (and not in selected benchmarks à la HAWX
Not at all. It doesn't have to "soundly beat" it for us to see that the balance has shifted from last generation. Like I said before I think "GK106" v Pitcairn will be an interesting matchup.
Sandra gpgpu bench against tahiti AND pitcain:
http://www.theinquirer.net/inquirer...hed-amds-mid-range-radeon-hd-7870-gpu-compute
Wonder if the clock is really 700 or 1000 here..
http://www.theinquirer.net/inquirer...hed-amds-mid-range-radeon-hd-7870-gpu-compute
Wonder if the clock is really 700 or 1000 here..
Jawed
Legend
If hot-clocking G80 is easier than not-hot-clocking Kepler, then I'm all ears.Like with everything else in life there's a lot to be said for experience.
But I doubt it very much.
Perhaps it's just a matter of hot-clocking being a good trade-off on 65nm but not on 28nm, due to different process characteristics.
It might also be related to the fact that power is becoming a bigger concern than die size.
http://blog.renderstream.com/2012/03/pre-order-gtx-680-and-the-latest-intel-xeon-based-systems/
GTX 680 is almost here!
NVIDIA’s latest consumer GPU, the GTX 680 has some very impressive specs and we think it will be very well received by the 3D industry. The GTX 680 is said to have an amazing 1536 CUDA cores compared to the 512 CUDA cores of the popular GTX 580! Based on our previous benchmarks we anticipate the GTX 680 to be about 3X faster at GPU rendering compared to the GTX 580.
This means that we can pack an incredible amount of computational power into our machines. For example, our RenderStream desktop when equipped with 2 GTX 680s will be like having a system with 6 of the previous generation 580 cards – a truly amazing GPU rendering experience for the price. We expect to get you some rendering benchmarks on the GTX 680 ASAP
GTX 680 is almost here!
NVIDIA’s latest consumer GPU, the GTX 680 has some very impressive specs and we think it will be very well received by the 3D industry. The GTX 680 is said to have an amazing 1536 CUDA cores compared to the 512 CUDA cores of the popular GTX 580! Based on our previous benchmarks we anticipate the GTX 680 to be about 3X faster at GPU rendering compared to the GTX 580.
This means that we can pack an incredible amount of computational power into our machines. For example, our RenderStream desktop when equipped with 2 GTX 680s will be like having a system with 6 of the previous generation 580 cards – a truly amazing GPU rendering experience for the price. We expect to get you some rendering benchmarks on the GTX 680 ASAP
I'm not buying the "new architecture" bit just yet. So far it looks like an optimized GF114 Fermi on clock roids.
@Jawed, maybe 28nm wide+slow simply has better overall perf/watt/mm characteristics than narrow+fast on previous nodes.
It's worth noting that the entire chip is now running at much higher clocks. If rumors pan out the regfile, scheduler, tmus, rops, geometry etc will be running at speeds approaching G80's hot-clock.
@Jawed, maybe 28nm wide+slow simply has better overall perf/watt/mm characteristics than narrow+fast on previous nodes.
It's worth noting that the entire chip is now running at much higher clocks. If rumors pan out the regfile, scheduler, tmus, rops, geometry etc will be running at speeds approaching G80's hot-clock.
