GK104 / Tahiti die size comparison

[tviceman]

Looks like GK104 is a little smaller than the 340-350 mm^2 estimation that it was pegged at, being around 320mm^2 in this comparison. I don't think Nvidia would give a card with this GPU the name of gtx680 unless it can generally be faster than an hd7970. Faster and smaller? Hmmmmm.....

Originally posted at http://forum.beyond3d.com/showthread.php?t=58668&page=84

 

[GaiaHunter]

To be faster and smaller NVIDIA would have to/did improve performance/mm² quite a bit and probably transistor density as well.

Will be interesting, if it is true, how NVIDIA did achieve that - compromise or engineering breakthroughs.

 

[BallaTheFeared]

AMD had to redesign for GPGPU, and they went to a larger bus than previous generations.

That could account for some additional space required for AMD.

No clue what Nvidia kepler will be, but I wouldn't try to compare it directly to a shrunk Fermi just yet.

 

[LOL_Wut_Axel]

If that's true, then good job by engineering at NVIDIA. Tahiti is at 352mm^2, though, so not that huge of an improvement in comparison.

 

[RussianSensation]

If that's true, then good job by engineering at NVIDIA. Tahiti is at 352mm^2, though, so not that huge of an improvement in comparison.

What if GK104 = GTX680 (moved the naming 1 level up ala HD6900 from HD5800), has 320-340mm^2 die with performance ~ HD7950 and then

GK110 = GTX690 is a 500-550mm^2 monster, and is about 40-50% faster than HD7950/GTX580.

The problem with GK104 being the highest end card is 256-bit memory bus. This also contradicts every rumor under the sun. Every source I can think of in the last 3-6 months reported that GK104 was a Performance replacement not the Enthusiast line and that GK110 is the highest end single-GPU version that will be released later.

 

[Keysplayr]

What if GK104 = GTX680 (moved the naming 1 level up ala HD6900 from HD5800), has 320-340mm^2 die with performance ~ HD7950 and then

GK110 = GTX690 is a 500-550mm^2 monster, about be 40-50% faster than HD7950/GTX580.

The problem with GK104 being the highest end card is 256-bit memory bus. This also contradicts every rumor under the sun. Every source I can think of in the last 3-6 months reported that GK104 was a Performance replacement not the Enthusiast line and that GK110 is the highest end single-GPU version.

So that would mean GK110 would be like:

http://www.youtube.com/watch?v=pqFM6Xounvw

 

[RussianSensation]

So that would mean GK110 would be like:

http://www.youtube.com/watch?v=pqFM6Xounvw

Not really. It's more like HD7950 being 5% faster than GTX580 and barely faster than mid-range HD7870 is the jaw drop. There is no way AMD is naive enough to believe that Kepler would only be 20% faster than GTX580. NV has never bested its previous high-end by only 20% <-- AMD knows it, almost everyone knows it. AMD simply launched HD7970 early to maximize profits @ $550 and it worked very well for them since Kepler is late by almost a quarter. AMD needs $ far more than NV does. It was a good strategy for them.

However, it's reasonable to believe AMD never expected HD7970 to compete with the highest end Kepler. Instead of spending 3 more months trying to get 1100mhz HD7970 out, they went with the 'first mover' strategy and thus were able to price it at $550. The fact that there are shortages of supply and driver problems means they rushed the GPU to market. But they didn't care since they knew it only needed to be 20-25% faster to safely secure the $550 price.

It's probably a certainty that there has to be a better version of HD7970 coming, by Q3 or whenever, especially given how well HD7970 overclocks.

 

[Homeles]

Tahiti is 352mm², not 365mm²...

The relative transistor density on Pitcairn is really neat. I was quite impressed by AMD cramming twice the power of Cape Verde into a die significantly less than half the size.

It'll be interesting to see how Nvidia fares. I really wish I had some architectural details to absorb, but there isn't crap out there on Kepler's arch right now. Hopefully some information is released soon...

 

[kidsafe]

Why would I believe GK104 to be any better than Tahiti when...

1) It has a smaller die.
2) It has less memory and associated bandwidth.
3) It only has a 5-phase power design.

 

[Olikan]

It's probably a certainty that there has to be a better version of HD7970 coming, by Q3 or whenever, especially given how well HD7970 overclocks.

just to add...

With the Tahiti XT chip sitting at a 925MHz clock speed out of the box, you could be forgiven for thinking there wouldn't be much headroom left in the chip. According to Zvika Greenstein, AMD director of Product Management for Discrete Graphics, AMD "made a conscious decision to leave a lot of overclocking headroom."
(...)
We would have thought that having a 1GHz GPU at launch would have really pushed Nvidia to compete when it brings Kepler to market in a couple of months' time, but it looks as though AMD is leaving that option to its motherboard manufacturers and their factory overclocked offerings.

http://www.techradar.com/news/computing-components/graphics-cards/amds-southern-islands-graphics-explained-1066145?src=rss&attr=all

 

[Silverforce11]

AMD had to redesign for GPGPU, and they went to a larger bus than previous generations.

There's no indication that their new architecture is less efficient per mm2 or per w than VLIW, infact, its more efficient.

So the caveat that somehow they redesigned for GPGPU means its inefficient. In terms of overall gaming load, VLIW was very inefficient, not using 25-33% of its shaders.

Quote: "We would have thought that having a 1GHz GPU at launch would have really pushed Nvidia to compete when it brings Kepler to market in a couple of months' time, but it looks as though AMD is leaving that option to its motherboard manufacturers and their factory overclocked offerings."

This is such a stupid design decision, as if not setting a high standards would therefore mean your competitor will go along with that and not push their limits too. In benchmarks, the relevant info is stock vs stock performance, only enthusiasts care about overclocking headroom. Most ppl would not even factor that bonus in their purchase decisions.

 

[BallaTheFeared]

There's no indication that their new architecture is less efficient per mm2 or per w than VLIW

Nobody said it wasn't.

 

[wahdangun]

Why would I believe GK104 to be any better than Tahiti when...

1) It has a smaller die.
2) It has less memory and associated bandwidth.
3) It only has a 5-phase power design.

yeah, especially nvdia always have bigger die size than amd.

and to be honest I'm quite impressed with HD7870, when everyone said it will have lower performance than HD 6970 but the reality is it can even outperform HD 7970 in crysis !!!!!!!!!

 

[LOL_Wut_Axel]

What if GK104 = GTX680 (moved the naming 1 level up ala HD6900 from HD5800), has 320-340mm^2 die with performance ~ HD7950 and then

GK110 = GTX690 is a 500-550mm^2 monster, and is about 40-50% faster than HD7950/GTX580.

The problem with GK104 being the highest end card is 256-bit memory bus. This also contradicts every rumor under the sun. Every source I can think of in the last 3-6 months reported that GK104 was a Performance replacement not the Enthusiast line and that GK110 is the highest end single-GPU version that will be released later.

You could make the same argument for the rumored Tenerife.

Right now, GK104 will be Enthusiast no matter how many people want to deny it because it'll be the fastest GPU from NVIDIA for a good 6+ months. It'll be Performance when GK110 cards are released.

Also, GK110 will bring about the new 700 series.

Like I've said numerous times, GK104 is engineered to be 50-65% faster than GF114 and Tahiti is 60% faster than GF114. Keep that in mind.

 

[SlowSpyder]

Not really. It's more like HD7950 being 5% faster than GTX580 and barely faster than mid-range HD7870 is the jaw drop. There is no way AMD is naive enough to believe that Kepler would only be 20% faster than GTX580. NV has never bested its previous high-end by only 20% <-- AMD knows it, almost everyone knows it. AMD simply launched HD7970 early to maximize profits @ $550 and it worked very well for them since Kepler is late by almost a quarter. AMD needs $ far more than NV does. It was a good strategy for them.

However, it's reasonable to believe AMD never expected HD7970 to compete with the highest end Kepler. Instead of spending 3 more months trying to get 1100mhz HD7970 out, they went with the 'first mover' strategy and thus were able to price it at $550. The fact that there are shortages of supply and driver problems means they rushed the GPU to market. But they didn't care since they knew it only needed to be 20-25% faster to safely secure the $550 price.

It's probably a certainty that there has to be a better version of HD7970 coming, by Q3 or whenever, especially given how well HD7970 overclocks.

I pretty much agree with you completely. AMD just had to get out a 28nm chip that was 'good enough' that could be sold at high prices. At 925MHz the 7970 out paces any single GPU, but it has huge overclocking headroom. This makes me believe they found a speed that worked well enough for yields and was also fast enough to sell at a high price point. When the GK110 arrives, I imagine AMD will have something (89xx?) coming out soon after, or maybe even before. Just my $.02.

 

[MrTeal]

yeah, especially nvdia always have bigger die size than amd.

and to be honest I'm quite impressed with HD7870, when everyone said it will have lower performance than HD 6970 but the reality is it can even outperform HD 7970 in crysis !!!!!!!!!

Huh? Link?

 

[3DVagabond]

yeah, especially nvdia always have bigger die size than amd.

and to be honest I'm quite impressed with HD7870, when everyone said it will have lower performance than HD 6970 but the reality is it can even outperform HD 7970 in crysis !!!!!!!!!

I think we'll find that those games the 7870 outperforms the 7900's in is because of the 12.2 drivers. Most sites won't have rerun the benches on the 7900 cards. They'll have gone with the release driver scores.

 

[Mopetar]

Huh? Link?

It's a joke of sorts. One of the review sites had a Crysis benchmark where the 7970 was using older drivers so the 7870 outperformed it because of the performance increases from the newer drivers.

 

[BallaTheFeared]

It's a joke of sorts. One of the review sites had a Crysis benchmark where the 7970 was using older drivers so the 7870 outperformed it because of the performance increases from the newer drivers.

This was proven?

 

[Bryf50]

You could make the same argument for the rumored Tenerife.

Right now, GK104 will be Enthusiast no matter how many people want to deny it because it'll be the fastest GPU from NVIDIA for a good 6+ months. It'll be Performance when GK110 cards are released.

You've said this in every thread. Its a totally pointless argument of semantics. We understand what your saying. Great. But your arguing with no one.

 

[Homeles]

This was proven?

Could be a combination of higher clocked geometry engines (bottle neck with Crysis 2 has gotta be with tessellation) with the aforementioned driver differences.

 

[BallaTheFeared]

What's the status of the tess units on these compared to 79xx, also has AMD updated the tess factors in their "optimized" settings, and do they disclose that information in their driver releases?

 

[notty22]

What's the status of the tess units on these compared to 79xx, also has AMD updated the tess factors in their "optimized" settings, and do they disclose that information in their driver releases?

TH reporting texture issues, so there could be games going on there.
If so, one of the sites, that have found these things before will uncover them.
http://www.tomshardware.com/reviews/radeon-hd-7870-review-benchmark,3148-3.html

As we go through the benchmarks, you&#8217;ll see that the Radeon HD 7800s do an incredibly good job of applying Adaptive anti-aliasing to our Skyrim benchmark sequence. The numbers were so impressive, in fact, that we went out of our way to snap screenshots in order to validate the resulting image quality, too. As it turns out, they do; Adaptive AA on the 7800s looks similar to the Radeon HD 6900s, putting our minds at ease.
While we were scrutinizing those Adaptive AA results, however, we couldn&#8217;t help but notice that some textures appeared noticeably blurrier in the Radeon HD 7800 screenshots. We first assumed a setting had changed in the game or driver. But double-checking proved that wasn&#8217;t the case. Further investigation showed that the Radeon HD 7800-series cards match the Radeon HD 6900&#8217;s crisper output if the Catalyst A.I. texture filtering quality slider is moved from its default (Quality) to the highest (High Quality) setting.
So, to be clear, using the exact same 8.95.5 driver at its default settings, the Radeon HD 7800s deliver blurrier textures than the Radeon HD 6900s. Take a look:
The differences are not colossal, and you probably wouldn&#8217;t notice them during game play (we didn&#8217;t). But they're easily identifiable in screen shots. We don&#8217;t want to overstate the impact of what we&#8217;re seeing. But, on the other hand, we take reductions in image quality seriously because the slope is slippery, and we&#8217;ve seen this before.
This issue came up very late in our testing. We asked AMD for comment, but don&#8217;t have an official response as of yet. Moreover, questions remain: Are the Radeon HD 7800 cards enjoying higher performance as a result of an optimization? Could this be an unintentional bug? How much better would the Radeon HD 6900s look if we also bumped their Catalyst A.I. slider up to High Quality? We absolutely plan to answer all of those questions after we collect more data. And we&#8217;ll update the story once we get some more feedback from AMD.

 

[Homeles]

What's the status of the tess units on these compared to 79xx, also has AMD updated the tess factors in their "optimized" settings, and do they disclose that information in their driver releases?

I think so? I don't think the 7870 drivers have been released, although I could be wrong.

And they've got the same number of geometry units - 2. Since this is clocked higher, that could make a difference, although I doubt the difference would be that large. I did pick the most extreme example though.

 

[wahdangun]

Could be a combination of higher clocked geometry engines (bottle neck with Crysis 2 has gotta be with tessellation) with the aforementioned driver differences.

nope, its not because tessellation because if that was true then HD 6970 can't compete with GTX 480. but look in that graph the evergreen and cayman card neck to neck with fermi GF100 and GF110.

and if it was because new driver then imagine if HD 7970 was re-bench with this new driver then the performance jump will be quite high.