[wccftech] Another leaked Kepler performance slide 760/770/780/790...

[BallaTheFeared]

Oh personal attacks, nice touch.

There was nothing ignorant about my comment, I was aware of the performance per watt increase. My thoughts were it didn't mean anything, can't you figure that out? First off the cost difference to run is nothing, secondly the 470/480 were the worst performance per watt 40nm chip released (probably because of the high leakage that they had to accept just to get something out).

Improving on a flawed (performance per watt) design which was performing poorly in that area because they had to rush it out isn't impressive.


You've been talking to too many AMD people, performance per watt on a performance desktop is like putting E85 in a race car. Sure it's nice, but it's not a factor.

 

[3DVagabond]

thats fine that you are not impressed. you saying "the 570 wasn't much of an improvement over the 470/480, slightly better performance per watt, but nothing to write home about" was inaccurate though. please show me one other time where a card became cooler, quieter, raised clocks, increased sp count, increased tmu count and became 25% faster using the same wattage while staying on the same core and process.

I'm with you overall. I don't even know why we're comparing the 470 with the 570 in the first place. The 570 was the improved 480. The 560 ti 448 is the "next gen" 470. That's even a more impressive improvement, IMO. Although, to be honest, I'm not really surprised that these cards improve substantially over the prior ones. The prior gen (470/480) was really really bad perf/w. nVidia really missed the mark and needed improvement. Of course, much of the improvement was done with software, not because the chip was dramatically more efficient.

Back on topic, Kepler should be much better. Assuming all is well with the chips. They're dropping down a full node and also dropping "hot clocks". The shaders will be running at 1/2 the speed relative to what Fermi does. Those things alone will make large improvements.

 

[toyota]

why would dropping the shader clocks on Haswell make larger improvements?

 

[3DVagabond]

why would dropping the shader clocks on Haswell make larger improvements?

OK, I'm confused. Haswell?

 

[Arkadrel]

OK, I'm confused. Haswell?

Haswell is the codename for a processor microarchitecture to be developed by Intel's Oregon team as the successor to the Sandy Bridge architecture.

Its just a typo, he ment Kepler.

 

[toyota]

OK, I'm confused. Haswell?

Haswell is the codename for a processor microarchitecture to be developed by Intel's Oregon team as the successor to the Sandy Bridge architecture.

Its just a typo, he ment Kepler.

yes sorry I meant Kepler. I was just looking at the Haswell thread and had that in my mind.

 

[3DVagabond]

Haswell is the codename for a processor microarchitecture to be developed by Intel's Oregon team as the successor to the Sandy Bridge architecture.

Its just a typo, he ment Kepler.

Ok, if that's all it is. I honestly wanted to make sure I wasn't missing something. So, am I to understand that shader clocks makes no difference in power usage? I thought that it does.

 

[Arzachel]

Back on topic, Kepler should be much better. Assuming all is well with the chips. They're dropping down a full node and also dropping "hot clocks". The shaders will be running at 1/2 the speed relative to what Fermi does. Those things alone will make large improvements.

Unless Kepler is a ground-up new arch (think VLIW -> GCN), that would be a horrid thing to do, because it would lead to a insane die size and around half performance per shader. What hotclocking does, is that it allows shaders to run at double the speed of the core clock, and Nvidia has used it since G80 iirc. It allows them to trade off power efficiency for transistor efficiency, because hotclocking shaders takes up much less die space than just doubling the shader count, but power draw increases exponentially with clocks(iirc, correct me if I'm wrong).

Nvidia dropping hotclocked shaders would either mean a total shader rework or TSMC 28 nm process just refusing to go past 2 Ghz.

 

[3DVagabond]

Unless Kepler is a ground-up new arch (think VLIW -> GCN), that would be a horrid thing to do, because it would lead to a insane die size and around half performance per shader. What hotclocking does, is that it allows shaders to run at double the speed of the core clock, and Nvidia has used it since G80 iirc. It allows them to trade off power efficiency for transistor efficiency, because hotclocking shaders takes up much less die space than just doubling the shader count, but power draw increases exponentially with clocks(iirc, correct me if I'm wrong).

Nvidia dropping hotclocked shaders would either mean a total shader rework or TSMC 28 nm process just refusing to go past 2 Ghz.

Of course, it's all rumors, but that's supposed to be exactly what's happening. Supposedly they can't clock the shaders high enough to keep the hot clock. Thus the major increase in "reported" shader counts.

This was originally reported on Chiphell

 

[BallaTheFeared]

Wow... The speculation is getting crazy...

2304 cuda cores at 950MHz = X4500... I only get X4000 with SLI and that's faster than a 1125MHz 7970...

300 TDP, that's quite a bit higher than the 7970

Here is the pic from Chiphell:

 

[Arzachel]

Of course, it's all rumors, but that's supposed to be exactly what's happening. Supposedly they can't clock the shaders high enough to keep the hot clock. Thus the major increase in "reported" shader counts.

This was originally reported on Chiphell

+ D: If that is true, then Nvidia is various degrees of screwed, depending on when they found that out. No wonder they're banging the efficiency drum, there is no chance of them getting anywhere near high-end performance without an overhaul of the arch, if they lose the hotclocks.

 

[Idontcare]

Unless Kepler is a ground-up new arch (think VLIW -> GCN), that would be a horrid thing to do, because it would lead to a insane die size and around half performance per shader. What hotclocking does, is that it allows shaders to run at double the speed of the core clock, and Nvidia has used it since G80 iirc. It allows them to trade off power efficiency for transistor efficiency, because hotclocking shaders takes up much less die space than just doubling the shader count, but power draw increases exponentially with clocks(iirc, correct me if I'm wrong).

Nvidia dropping hotclocked shaders would either mean a total shader rework or TSMC 28 nm process just refusing to go past 2 Ghz.

Since you asked, power increases linearly with clocks.

However, increasing the clockspeed does tend to require increasing the voltage, in addition to the operating temperature increasing.

Both of which feed back into the power-consumption equation and result in rapidly increasing power consumption that is a quadratic function.

(If curious, see here for an analysis of the 2600K, the basic device physics are the same, CPU or GPU)

 

[3DVagabond]

Wow... The speculation is getting crazy...

2304 cuda cores at 950MHz = X4500... I only get X4000 with SLI and that's faster than a 1125MHz 7970...

300 TDP, that's quite a bit higher than the 7970

Here is the pic from Chiphell:

I'm sure that, assuming this chart is at all true, performance are estimates. As far as anyone knows, there are no GK100's (aka: BigGK) to even be tested.

 

[IonusX]

I'm sure that, assuming this chart is at all true, performance are estimates. As far as anyone knows, there are no GK100's (aka: BigGK) to even be tested.

indeed let us not forget that by the time we get to argue about this sea islands will be on the horizon for amd.. so amd will be able to smack it in the back if it cant in the front.

 

[Dark Shroud]

I have to say this thread has been a entertaining read.

Now then Nvidia could be in some real trouble here, I just hope AMD doesn't mess this up and makes some money. Either way we're seeing another delayed FERMI launch.

 

[AtenRa]

Wow... The speculation is getting crazy...

Here is the pic from Chiphell:

So NVIDIA will change its Cuda cores with Kepler ??

If the above specs are true, it seams that Kepler cores can do 2 ops per cycle (much like AMDs) and not 4 ops per cycle that we had so far up until Fermi.

One more thing, it seams that GK100 and GK104 have the same architecture with 96 cores paired with 4 TMUs, but GK106 and lower have 96 cores paired with 8 TMUs.
Not only that but GK100 and GK104 would be different dies with the same architecture.

Edit: I have to say that i find it very odd for NV to change its cores now, i believe that could have tremendous implications with software (Cuda etc). Even AMD with GCN kept the same analogy of VGPRs and instructions/operations between the new ALUs and old VLIW.

 

[BallaTheFeared]

It's all speculation, unless as others have said they can't get the 2GHz+ clocks on the shaders this move doesn't make a lot of sense.

 

[Arzachel]

Since you asked, power increases linearly with clocks.

*pic*

However, increasing the clockspeed does tend to require increasing the voltage, in addition to the operating temperature increasing.

Both of which feed back into the power-consumption equation and result in rapidly increasing power consumption that is a quadratic function.

(If curious, see here for an analysis of the 2600K, the basic device physics are the same, CPU or GPU)

Thanks for the clarification! I probably should have remembered some of that from back when you posted about how power draw scales with voltage in one of the Bulldozer speculation threads.

 

[3DVagabond]

Since you asked, power increases linearly with clocks.

However, increasing the clockspeed does tend to require increasing the voltage, in addition to the operating temperature increasing.

Both of which feed back into the power-consumption equation and result in rapidly increasing power consumption that is a quadratic function.

(If curious, see here for an analysis of the 2600K, the basic device physics are the same, CPU or GPU)

So, just so I'm clear, would cutting shader clocks (not necessarily the GPU clocks) reduce power usage, as far as you know? I thought it would, but Toyota has made me question that belief.

 

[Keysplayr]

Sure it would. Use extremes. If you increased shader clocks to 70 bazillion MHz, would it use more power than if they were run at 2000MHz? Common sense says, of course it would. By how much? No idea.

 

[bunnyfubbles]

Wow... The speculation is getting crazy...

2304 cuda cores at 950MHz = X4500... I only get X4000 with SLI and that's faster than a 1125MHz 7970...

300 TDP, that's quite a bit higher than the 7970

Here is the pic from Chiphell:

TDP would be quite a bit higher, but so would performance, thats a monster pixel fillrate, would be looking at a bare minimum of 50% faster in just about every instance than a 580 with specs like that.

 

[blackened23]

If you click on that link to the chiphell forum, all of the posters there (translated) are saying the slides are fake. Gotta love google chrome translate.

 

[Keysplayr]

Fake. If you click on that link to the chiphell forum, all of the posters there (translated) are saying the slides are fake. Gotta love google chrome translate.

Do they give any reasoning? Or are they just calling "fake" out of hand?
What are they saying?

 

[Lonyo]

Do they give any reasoning? Or are they just calling "fake" out of hand?
What are they saying?

Here's some reasoning why at least some of it is fake: Those specs show 5.8GHz on the RAM of the GK104 to get that memory bandwidth on that bus and 5GHz on the GK100.

Now, I'm not saying it's impossible, but NV can barely manage to break 4GHz currently on any card, so unless they have made a massive breakthrough, I don't see 5.8GHz happening. 5GHz is not unreasonable, since AMD have managed 5.5GHz on 384-bit and 256-bit cards, but NV seem to struggle with getting high speed RAM working on their cards.

Now that's only one minor spec and always subject to change, but IMO it's living in dreamworld if you claim 5.8GHz stock clock on GDDR5, especially for a company which can't even manage 5GHz at the moment.

That doesn't mean it's all fake, but it means that at least some of it is from a dreamworld.


(The above is based on the following:
http://www.anandtech.com/show/4061/amds-radeon-hd-6970-radeon-hd-6950/11)

Cracking open the 6970 we find the PCB with the Cayman GPU at the center in all its 389mm2 glory. Around it are 8 2Gb Hynix GDDR5 chips, rated for 6Gbps, 0.5Gbps higher than what the card actually runs at. As we’ve said before the hardest part about using GDDR5 at high speeds is the complexity of building a good memory bus, and this continues to be the case here. AMD has made progress on getting GDDR5 speeds up to 5.5Gbps primarily through better PCB designs, but it looks like hitting 6Gbps and beyond is going to be impractical, at least for a 256bit bus design. Ultimately GDDR5 was supposed to top out at 7Gbps, but with the troubles both AMD and NVIDIA have had, we don’t expect anyone will ever reach it.

 

[3DVagabond]

Here's some reasoning why at least some of it is fake: Those specs show 5.8GHz on the RAM of the GK104 to get that memory bandwidth on that bus and 5GHz on the GK100.

Now, I'm not saying it's impossible, but NV can barely manage to break 4GHz currently on any card, so unless they have made a massive breakthrough, I don't see 5.8GHz happening. 5GHz is not unreasonable, since AMD have managed 5.5GHz on 384-bit and 256-bit cards, but NV seem to struggle with getting high speed RAM working on their cards.

Now that's only one minor spec and always subject to change, but IMO it's living in dreamworld if you claim 5.8GHz stock clock on GDDR5, especially for a company which can't even manage 5GHz at the moment.

That doesn't mean it's all fake, but it means that at least some of it is from a dreamworld.

I believe that GK106 is a real live chip in the wild. These could possibly come from the 106 specs extrapolated across the expected range of chips.