Kepler performance

[sdlvx]

Originally Posted by Shadow Concept

Do you think this will lead to AMD moving the performance segment goalposts like NVidia did?

Before Kepler, Pitcairn was the peak of "pure gaming" oriented cards, and of course Tahiti was the top card with the immense feature set. Then NV moves the "pure gaming" goalpost by creating a much bigger "pure gaming" card, not far from the size of Tahiti, but with the same design goals as Pitcairn.

From my observations, GCN still has quite the upper hand, so if AMD were to create a card close to or the same size as Tahiti/Kepler, but with Pitcairn/680 style design goals (+ any slight improvements they've made in the meantime), they should, at the same die size, stomp the current Kepler into the ground on all metrics.

I guess the problem with that is it would push the "full featured" top card in a die size area that AMD might prefer to avoid. But with the size and feature set of BigGK, I can see them doing their usual trick of building a die much, much smaller, getting 90% of the performance at 2/3rd the price and laughing all the way to the bank.

I think moving the performance segment goal posts and dropping GPGPU from gaming cards is the smartest thing to do if all you do is make GPUs given the cost of TSMC nodes as they shrink. As I said before, NV has been pushing CUDA and GPGPU for a very long time.

http://www.beyond3d.com/content/news/304

We are coming up on the five year anniversary of CUDA. If there hasn't been a killer consumer application using CUDA in five whole years, there's not ever going to be one. GK104 and NV making GK104 GTX 680 makes me feel like NV is finally admitting defeat in the consumer GPGPU space. Don't get me wrong, they've got GPGPU covered in other markets quite extensively, but for the consumer it doesn't matter.

It looks like GPGPU in the consumer space is a bad investment. But it's definitely not for AMD. AMD is aiming for HSA and they want the GPU and CPU to work very closely with each other. Intel and NV have nothing to gain from doing that with each other. Intel and NV are about to start competing with each other when Knight's Corner comes around.

Given AMD's goals, I don't see them dropping GPGPU in consumer cards, while it makes sense for NV. The work with the GPU architectures will pay off hugely if AMD plays everything right in the end.

All you have to do is look at AMD's existing hardware to know what direction they're going towards. Bulldozer sucks at floating point and is awesome at integer math. GCN is awesome at floating point and sucks at integer math. It only makes sense given AMD's stated goals of HSA that these two product derivatives should eventually exist on the same chip and work to execute the same code for impressive performance.

The only remotely decent real world app for consumers (I'm not talking about F@H and stuff) was CUDA transcoding, and GTX680 has NVENC. If they've replaced their strongest GPGPU consumer app with something like that they've abandoned the consumer market.

[prender]

I've got a what if scenario that I've done a little thinking about. It may be just downright illogical and stupid thinking on my part, as I'm not exactly as technical minded as some of you guys on here seem to be, but just humour me...

What if the GTX680 wasn't exactly the real "Kepler" that nVIDIA was going to release? What if the GTX680 was a backup plan, if the real Kepler cards were too difficult at this point in time for nVIDIA to manufacture?

Is that even possible? I don't know, just a feeling I've got. Probably nothing, maybe indigestion. Too much bacon, perhaps. What if the entire Kepler series were supposed to still have their GPGPU functionality, but they decided to keep a back up plan - a card that was stripped of most of its compute fuctions, clocked high, in order to try and compete with Tahiti, at least from a gaming perspective, and to keep their investors/share holders, etc happy, until they can get their real Kepler cards out. What if it was just a gamble for nVIDIA that happened to pay off. As in the GTX680 performing above their own expectations.

This isn't intended to be trolling, or baiting or anything of that sort. I was just in the bath the other day, and the idea came to me. I come up with all sorts of strange thoughts and ideas whenever the bath, or the shower, is involved.

[Relayer]

Originally Posted by prender

I've got a what if scenario that I've done a little thinking about. It may be just downright illogical and stupid thinking on my part, but just humour me...

What if the GTX680 wasn't exactly the real "Kepler" that nVIDIA was going to release? What if the GTX680 was a backup plan, if the real Kepler cards were too difficult at this point in time for nVIDIA to manufacture?

Is that even possible? I don't know, just a feeling I've got. Probably nothing, maybe indigestion. Too much bacon, perhaps. What if the entire Kepler series were supposed to still have their GPGPU functionality, but they decided to keep a back up plan - a card that was stripped of most of its compute fuctions, clocked high, in order to try and compete with Tahiti, at least from a gaming perspective, and to keep their investors/share holders, etc happy, until they can get their real Kepler cards out. What if it was just a gamble for nVIDIA that happened to pay off. As in the GTX680 performing above their own expectations.

This isn't intended to be trolling, or baiting or anything of that sort. I was just in the bath the other day, and the idea came to me. I come up with all sorts of strange thoughts and ideas whenever the bath, or the shower, is involved.

I would think the GK104 is exactly as planned. Probably clocked a little higher than if the GK110 would have come along as planned. Reduced compute functionality in the 2nd level chips isn't a new thing.

[Shadow Concept]

Originally Posted by sdlvx

I think moving the performance segment goal posts and dropping GPGPU from gaming cards is the smartest thing to do if all you do is make GPUs given the cost of TSMC nodes as they shrink. As I said before, NV has been pushing CUDA and GPGPU for a very long time.

http://www.beyond3d.com/content/news/304

We are coming up on the five year anniversary of CUDA. If there hasn't been a killer consumer application using CUDA in five whole years, there's not ever going to be one. GK104 and NV making GK104 GTX 680 makes me feel like NV is finally admitting defeat in the consumer GPGPU space. Don't get me wrong, they've got GPGPU covered in other markets quite extensively, but for the consumer it doesn't matter.

It looks like GPGPU in the consumer space is a bad investment. But it's definitely not for AMD. AMD is aiming for HSA and they want the GPU and CPU to work very closely with each other. Intel and NV have nothing to gain from doing that with each other. Intel and NV are about to start competing with each other when Knight's Corner comes around.

Given AMD's goals, I don't see them dropping GPGPU in consumer cards, while it makes sense for NV. The work with the GPU architectures will pay off hugely if AMD plays everything right in the end.

All you have to do is look at AMD's existing hardware to know what direction they're going towards. Bulldozer sucks at floating point and is awesome at integer math. GCN is awesome at floating point and sucks at integer math. It only makes sense given AMD's stated goals of HSA that these two product derivatives should eventually exist on the same chip and work to execute the same code for impressive performance.

The only remotely decent real world app for consumers (I'm not talking about F@H and stuff) was CUDA transcoding, and GTX680 has NVENC. If they've replaced their strongest GPGPU consumer app with something like that they've abandoned the consumer market.

You've nicely summed up what I've been feeling for a while. However unlike Intel, NVidia has no advantage over AMD in the part they compete against - in fact I think the opposite - as I said before, the fact that NVidia's stripped out gaming card has to be almost the same size as AMDs full feature card, with significantly higher clocks (and a boost on top of that) just to beat it by a small margin speaks volumes about the engineering gap between the two. If AMD were to strip out the extra features and remove 128bits of memory controller in exchange for higher clocked memory as well as increasing the clocks to match 680, then you would have a card at 680's performance at a much smaller size.

Nvidia has made a win here with product placement, but considering Kepler is supposed to be some miracle architecture - it still is a long way behind in many areas when you compare it to the competition more closely. (I'm talking about the architecture itself, not 680 vs 7970). People at NVidia must be getting increasingly concerned about this, I can't be the only one that sees it.

[esrever]

I doubt nvidia can make their big compute card efficient. The 680 really draws about as much power as the 7970 and its performance/watt isn't nearly as good as the 7870.

The 7870 is about 30% more efficient than the 7970 in per/watt. 7970 with about 25% more performance and 60% more power consumption
the 7970 is about 70% larger than the 7870.
numbers taken from tpu Asus OC review for stock 7970 and 7870.

going from the gk104 to gk110 is about the same die area difference between the 7870 to 7970. So gk110 would be 25% faster than the gk104 and consume 60% more power.
this is a minimal as the 7870 is actually faster than the gk104 in gpgpu.

so a 312w gpu with a 520mm die for the gk110. nvidia will probably market it as 250w. Might be the 480 all over again if nvidia can't get their chip to be more efficient than AMD. considering the 195w 680 is already hot, I wonder how hot the 300w+ gk110 will be like.

If Nvidia wants a manageable TPD chip, it would probably be barely faster than the 680 in gaming due to having to have much lower clock speeds.

[boxleitnerb]

Not really, it depends on the game. In many games the 680 uses quite a bit less power than the 7970. That a 500+mm2 die with strong DP performance cannot reach the efficiency of Pitcairn is clear.
I'm a bit more optimistic, though. Firstly, I expect GK110 to be 40% faster than the 680. Secondly, the GPU boost will come in handy in using the TDP headroom better.

[esrever]

It doesn't matter what the 680 pulls compared to the 7970.
Im comparing the 7970 to the 7870 and translating it the the 680 and the nvidia compute card.

efficiency goes down 30% and power consumption goes up by 60% moving from gaming to compute from AMD. I don't see how nvidia can magically make it a more efficient transition considering the 680 is farther from the gk110 than 7870 is from 7970 in compute.

http://www.techpowerup.com/reviews/N...TX_680/25.html
680 consumes almost the same amount power as the 7970 generally. 195w nvidia TPD = 250w AMD TPD= 22% less . 312w gk110 is nvidia TPD, in AMD TPD it would be 410w.

No wonder AMD didn't want to make a chip that big and settled with a smaller gaming chip and a smaller compute chip.

[distinctively]

Originally Posted by boxleitnerb

Not really, it depends on the game. In many games the 680 uses quite a bit less power than the 7970. That a 500+mm2 die with strong DP performance cannot reach the efficiency of Pitcairn is clear.
I'm a bit more optimistic, though. Firstly, I expect GK110 to be 40% faster than the 680. Secondly, the GPU boost will come in handy in using the TDP headroom better.

That's more like it! Some hard predictions coming out. I'll just put my clean record of never being right and predict a 25% increase of GK110 over the GK104. I do agree that GPU boost might come in handy to control TDP.

[kalelovil]

Originally Posted by esrever

Im comparing the 7970 to the 7870 and translating it the the 680 and the nvidia compute card.

efficiency goes down 30% and power consumption goes up by 60% moving from gaming to compute from AMD. I don't see how nvidia can magically make it a more efficient transition considering the 680 is farther from the gk110 than 7870 is from 7970 in compute.

Compute functionality isn't the only thing explaining Tahiti's relative inefficiency.
Charlie has mentioned that Tahiti has a considerable amount of redundancy (and probably takes a conservative approach to more than just the core clock) since it was the first ASIC on TSMC's 28nm process and AMD wanted to reduce their risks as much as possible.

Although if what Charlie has heard is correct, and Nvidia is having significant problems producing a chip not even 300m2 on TSMC's 28nm process, then big Kepler is still quite a way off (or will never reach the mass gaming market).

[boxleitnerb]

Originally Posted by esrever

It doesn't matter what the 680 pulls compared to the 7970.
Im comparing the 7970 to the 7870 and translating it the the 680 and the nvidia compute card.

efficiency goes down 30% and power consumption goes up by 60% moving from gaming to compute from AMD. I don't see how nvidia can magically make it a more efficient transition considering the 680 is farther from the gk110 than 7870 is from 7970 in compute.

http://www.techpowerup.com/reviews/N...TX_680/25.html
680 consumes almost the same amount power as the 7970 generally. 195w nvidia TPD = 250w AMD TPD= 22% less . 312w gk110 is nvidia TPD, in AMD TPD it would be 410w.

No wonder AMD didn't want to make a chip that big and settled with a smaller gaming chip and a smaller compute chip.

One game (Crysis 2) != generally.
Tahiti is so "slow" compared to Pitcairn because the front end is holding it back. Kepler is scaling well, it has a better front end. I believe we will see a good scaling with additional units from GK110. TDP is the big question. 250W, 270W, 300W?