[bit-tech] Radeon HD 7000-series rumoured for May production

[tigersty1e]

Just so we are clear here.

6970 = 176GB/s
6950 = 160GB/s
6870 = 134GB/s
5870 = 153GB/s
5850 = 128GB/s

4890 = 124GB/s
4870 = 115GB/s
5770 = 76GB/s

Now look at the 5770 vs the 4870/90. The is a huge gap in bandwidth between the cards, but performance is very similar. Same SPs, very similar clocks. So how do you explain that.

Looking at the 6950, 5870 and 6870. The 6870 is closer to the 5870 in performance than the 5870 is to the 6950, but bandwidth favors the 5870 heavily against the 6870. How do you explain that too?

AMD have been using GDDR5 since 2008 with the 4870, I'm pretty sure they can get their memory controller working with higher speeds when the 7000 series is released. Who knows what kind of speeds the controller in Cayman can handle as it can easily max out the 6ghz of the GDDR5 that the 6970/50 come with.

comparing the 4890 to 5770 is more comparable considering they both have the same core speed. performance is not very similar. the 4890 significantly beats the 5770. the jump from 5770 to 4890 is similar to the jump from 4890 to 5850.

you have to realize that everybody loves to rave about their equipment. the 5770 is probably the most bought card from that gen and "everybody" kept saying their 5770 was as fast as a 4890.

again, where are you getting your numbers? the 5870 is closer to the 6950 than it is to the 6870.

http://www.anandtech.com/bench/Product/294?vs=293
http://www.anandtech.com/bench/Product/290?vs=294

amd using ddr5 since the 4870 is exactly my point. they vastly improved their controller on the 2nd go-around. 900 to 1200 (33%)....1200 to 1375 (15%)....1375 to ...definitely not 1500.

the perfect storm with die shrink, mem speed increase is what gave us nive 4890 to 5870 boost.

 

[RussianSensation]

comparing the 4890 to 5770 is more comparable considering they both have the same core speed.

the perfect storm with die shrink, mem speed increase is what gave us nive 4890 to 5870 boost.

I don't agree. HD5770 and HD4890 have identical texture and pixel fill-rate, yet HD5770 only has 62% of HD4890's memory bandwidth. How does the HD5770 outperform the HD4890 in some of the latest games then?

Also, AlienBabelTech did a nice analysis on the HD5770 which clearly shows it's more sensitive to GPU bottlenecking, not memory bandwidth.

Finally, you propose that memory bandwidth is critical in extracting even more performance out of next generation of cards. But if memory bandwidth was so critical for the Cayman architecture, why is it GTX580's performance drops from 16% to only 3% by going from 1920x1200 8AA to 2560x1600 8AA despite having 192GB/sec of memory bandwidth vs. HD6970's 176GB/sec?

If memory bandwidth was critical, then GTX580 would pull ahead more than 16% since that bottleneck would be felt even more with higher resolution + AA.

 

[Dean]

I don't agree. HD5770 and HD4890 have identical texture and pixel fill-rate, yet HD5770 only has 62% of HD4890's memory bandwidth. How does the HD5770 outperform the HD4890 in some of the latest games then?

Also, AlienBabelTech did a nice analysis on the HD5770 which clearly shows it's more sensitive to GPU bottlenecking, not memory bandwidth.

Finally, you propose that memory bandwidth is critical in extracting even more performance out of next generation of cards. But if memory bandwidth was so critical for the Cayman architecture, why is it GTX580's performance drops from 16% to only 3% by going from 1920x1200 8AA to 2560x1600 8AA despite having 192GB/sec of memory bandwidth vs. HD6970's 176GB/sec?

If memory bandwidth was critical, then GTX580 would pull ahead more than 16% since that bottleneck would be felt even more with higher resolution + AA.

I would assume at that res the actual frame buffer size would come into play.

 

[Silverforce11]

No. 1.5GB vram is enough for single monitor res, its totally not the reason why the 6970 ~equals a gtx580 at high res. The 69xx ROPs are twice as efficient, they just perform better at high res.

Bandwidth is not at all limiting to the 69xx series. You can go crazy on the mem OC and performance won't buldge much. This is why AMD runs the memory chips well below specs, it's just not needed. However, this does not mean that bandwidth won't matter in 79xx. If it's double up on everything, SP and ROPs, a huge boost in bandwidth is needed to keep it all running efficiently. By the time 79xx comes out, 7Gbps gddr5 will be the norm. There's room for 35-40% bandwidth gains (vs double SP/ROP) if they stick with 256 bit. Will it be enough? Maybe... but maybe not.

Have to look at it this way: 79xx doesn't need to add any new feature. dx11 is set. The architecture can handle tessellation loads fine now and has no inherent weakness which needs fixing. They can double up and give a huge perf jump. As for HPC stuff, cayman does it fine in terms of perf/watt. I can't see them doing anything wild but stick to their efficiency goals. perf/watt, perf/mm all the way.

 

[Arkadrel]

If they want to make smaller chips again, then they ll likely stick with 256bit bus....

If AMD can get maybe just ~10% more speed on memory, at 256bit, theyd be around ~192 GB/s memory bandwidth with 256bit bus (assumeing GDDR5 ~1500mhz mark).

Question is if thats enough? risk a card ending up abit memory starved vs saveing a bit on product costs via smaller die size?

....by the time 79xx comes out, 7Gbps gddr5 will be the norm.

7Gbps GDDR5

does that mean 7000/4 = 1750mhz speeds? (vs our current 1375mhz speeds)

256bit x 1,750 / 8 x 4 = ~224 GB/s memory bandwidth


it would be a BIG jump going from 1375mhz -> 1750mhz (like a ~25%+ jump)

 

[Silverforce11]

I don't think a 25% jump in bandwidth will be enough if everything else is doubling. I'm fully expecting AMD to make a cayman sized GPU on 28nm. They have shown no signs in recent time of backing down and giving up trying for the performance crown.

 

[Lonyo]

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.

 

[Silverforce11]

You can OC 6970 memory to ~6Gbps. Except it doesn't offer much performance gains at all. It doesn't need the extra bandwidth. Even if you OC the core without OC on memory you still get nice perf gains. At default, it has more bandwidth than it needs.

These assumptions of complex memory controllers don't really apply to 79xx because we have no idea what the architecture is going to be yet.

Did AMD somewhere made a statement that they had difficulty with GDDR5 over 6Gbps (For Cayman)??

 

[tviceman]

I don't agree. HD5770 and HD4890 have identical texture and pixel fill-rate, yet HD5770 only has 62% of HD4890's memory bandwidth. How does the HD5770 outperform the HD4890 in some of the latest games then?

Also, AlienBabelTech did a nice analysis on the HD5770 which clearly shows it's more sensitive to GPU bottlenecking, not memory bandwidth.

Finally, you propose that memory bandwidth is critical in extracting even more performance out of next generation of cards. But if memory bandwidth was so critical for the Cayman architecture, why is it GTX580's performance drops from 16% to only 3% by going from 1920x1200 8AA to 2560x1600 8AA despite having 192GB/sec of memory bandwidth vs. HD6970's 176GB/sec?

If memory bandwidth was critical, then GTX580 would pull ahead more than 16% since that bottleneck would be felt even more with higher resolution + AA.

You may have a point, but CoDBlops is such a poor, poor benchmark example to use in illustrating your point.

 

[Arkadrel]

@Silverforce11

You think AMD should aim for 380mm^2 for its topend?

And just double up on everything?

Is that even needed by anyone atm? imagine something thats as fast as 580 SLI on a single card.... what game needs that atm? Thats what a 380mm^2 28nm chip would be like.

I think AMD should aim at just makeing them smaller, aim for 150% of 580's performance, and make a chip thats like ~280mm^2 that they can sell for 150% 580's prices, and make huge profits on, until Nvidia has a 6xx series card out, and then AMD just lowers prices and competeings on performance/cost.

That would be the smarter bussines move to make. That would also mean theyd be sticking with 256bit bus for their cards again.

 

[tviceman]

@Silverforce11

You think AMD should aim for 380mm^2 for its topend?

And just double up on everything?

Is that even needed by anyone atm? imagine something thats as fast as 580 SLI on a single card.... what game needs that atm? Thats what a 380mm^2 28nm chip would be like.

I think AMD should aim at just makeing them smaller, aim for 1.5x 580's performance, and make a chip thats like ~280mm^2 that they can sell for 1.5x 580's prices, and make huge profits on, until Nvidia has a 6xx series card out, and then AMD just lowers prices and competeings on performance/cost.

That would be the smarter bussines move to make. That would also mean theyd be sticking with 256bit bus for their cards again.

They won't be able to double the number of cores at 28nm and still keep the same die size as before on 40nm. If they double the cores and everything else stays the same, the die size will be bigger than cayman.

 

[yacoub]

So while these may enter production in May, it's not likely they'll be out before Q3 and thus any significant price drops on 69xx series cards won't be for quite a while, eh?

 

[Madcatatlas]

So while these may enter production in May, it's not likely they'll be out before Q3 and thus any significant price drops on 69xx series cards won't be for quite a while, eh?

I think the 6950 has already significantly dropped down in price. According to one of the latest HardOCP reviews, the price range for 6950s is 200 to 300 with rebates
http://www.hardocp.com/article/2011/04/18/asus_eah6950_engtx570_directcu_ii_vid_card_review/9

 

[AtenRa]

Those things don't take away the intrigue, theory, thought-process, and contrast/compare by looking at R770/R790 in relation to Barts, because of the fact they have the same die size. Those things AMD did to Cypress is how they got the Barts die size down to the same as R770. Which is the entire point that is trying to be made. The point is not how the 7900 will compare to Barts.

We cant compare HD48xx to Barts HD68xx because they have the same Die Size. HD4890 has 956M transistors and Barts HD6870 has 1.7B transistors. They have the same die size because HD4890 have been made with the old 55nm process and HD6870 it is made at the latest 40nm process.

With 77% more transistors, Barts HD6870 cant be compared to HD4870/90

 

[Topweasel]

We cant compare HD48xx to Barts HD68xx because they have the same Die Size. HD4890 has 956M transistors and Barts HD6870 has 1.7B transistors. They have the same die size because HD4890 have been made with the old 55nm process and HD6870 it is made at the latest 40nm process.

With 77% more transistors, Barts HD6870 cant be compared to HD4870/90

I think you continue to miss his point. AMD had a homerun with the 4800 series in creating a top notch performance card with a smallish footprint. Something that they have tried to keep even while increasing the footprint.

With Barts you have a video card that is they set out to match the die size of that original 4800 series card. Its why they named it the 6800, in that power/performance/chip per wafer design they did previously. To basically say that while the 6900 series isn't about hitting the top performance, and the 5800 series was about staying competitive, that their sweet spot is still that 250ish mm^2 die size. Hell chances are that the 7900 series if coming out soon, with Nvidia still 6 months or more behind, that it will be aimed once again at that die size and this doubling up that people talk about won't come till 8k series a year from now.

At least that's my interpretation. I think it is very likely that the 7900 series will be a 250mm part that will give AMD the performance crown for a substantial time, where they can price it a $500 because its the fastest and ramp it all the way down to $200 if Nvidia ships a competing product while they design a chip 50% larger for the second go around.

 

[RussianSensation]

To basically say that while the 6900 series isn't about hitting the top performance, and the 5800 series was about staying competitive, that their sweet spot is still that 250ish mm^2 die size.

HD5800 series had no competition from September 2009 to April 2010. To say it was released to "stay" competitive is not very accurate. Even now the HD5870 is pretty much as fast as a stock HD6950 (that's 1.5 years after its release!). When AMD has major new releases, the performance difference is enormous (i.e., 8500 --> 9700Pro, 9800Pro --> X800XT, X800XT --> X1800XT, HD3870 --> HD4870 --> HD5870). GTX580/HD6970 are "refresh" parts which were marketed as "new generation". However, they should have been called GTX485 and HD5890 (while the HD6970 is indeed a new architecture, it was still limited by 40nm process; so technically it shouldn't be considered a new generation from a performance stand-point). Therefore, in my mind I look at it as HD5870 --> HD7970 as another major new release from AMD. Historically, such major releases generally have been accompanied by a 70-90% performance increase.

Also, there is no evidence to support the notion that 250-260mm^2 is the "sweet spot for AMD" for high-end videocards. If you look at post #78, you can see that they utilized die sizes of less than 260mm^2 three times, while they utilized die sizes of greater than 300mm^2 on four occurences in recent history. If anything, there is a slightly greater chance that the GPU die size will exceed 300mm^2 again.

the 7900 series if coming out soon, with Nvidia still 6 months or more behind, that it will be aimed once again at that die size and this doubling up that people talk about won't come till 8k series a year from now. I think it is very likely that the 7900 series will be a 250mm part that will give AMD the performance crown for a substantial time

HD7900 series likely won't ship until Q3 2011. Right now all we have are some indication that AMD will unveil "details" of HD7000 series in June. Based on that HD7xxx series is more likely to ship closer to Q3 2011, which means NV will not be 6 months behind.

Where they can price it a $500 because its the fastest and ramp it all the way down to $200 if Nvidia ships a competing product while they design a chip 50% larger for the second go around.

1) That's now how AMD operates. They don't release $500 high-end GPUs and reduce pricing on them to $200 when/if NV releases a faster card.

2) How are they going to increase the size by another 50% for the "second go around" in 2012? 28nm is here to stay for a while. 22nm is not expected until 2013. This is why HD7970 series should be AMD's biggest release since HD5870 series.

Based on rumours, Kepler is going to be a significantly faster, brand new GPU architecture from NV. Performance per watt is expected to increase 3-4x compared to Fermi. Double precision performance will also increase significantly. There is just no way AMD can do a half-baked 28nm "refresh" with HD7970 series if they want to stay in the game. If I were to guess at this point, I would say HD7970 may end up 70% faster than HD5870 at 1920x1200 4AA and 80-90% faster at 2560x1600. And even that may not be enough to beat out Kepler.

 

[AtenRa]

I think you continue to miss his point. AMD had a homerun with the 4800 series in creating a top notch performance card with a smallish footprint. Something that they have tried to keep even while increasing the footprint.

With Barts you have a video card that is they set out to match the die size of that original 4800 series card. Its why they named it the 6800, in that power/performance/chip per wafer design they did previously. To basically say that while the 6900 series isn't about hitting the top performance, and the 5800 series was about staying competitive, that their sweet spot is still that 250ish mm^2 die size. Hell chances are that the 7900 series if coming out soon, with Nvidia still 6 months or more behind, that it will be aimed once again at that die size and this doubling up that people talk about won't come till 8k series a year from now.

At least that's my interpretation. I think it is very likely that the 7900 series will be a 250mm part that will give AMD the performance crown for a substantial time, where they can price it a $500 because its the fastest and ramp it all the way down to $200 if Nvidia ships a competing product while they design a chip 50% larger for the second go around.

Since 28nm will double the transistor density, If we shrink Cayman (HD6970) at 28nm it could be close to 190-200mm2, down from 389mm2 (40nm).
By the way that will be a nice Middle End card later on

Do you think that 50-60mm2 more will add that much in order to get the performance crown from GTX580 and by how much ?? 10-20% ??

And then 4-6 months later Kepler will come and will obliterate the 79xx, no way
I see a 300-350mm2 from AMD this round.

 

[cusideabelincoln]

We cant compare HD48xx to Barts HD68xx because they have the same Die Size. HD4890 has 956M transistors and Barts HD6870 has 1.7B transistors. They have the same die size because HD4890 have been made with the old 55nm process and HD6870 it is made at the latest 40nm process.

With 77% more transistors, Barts HD6870 cant be compared to HD4870/90

You can compare them. Looking at the differences is one of the objectives when comparing two different things. This is contrast. You must compare to contrast. Get it? Got it? Good. Bolding your words doesn't make what you said true.

 

[RussianSensation]

You can compare them. Looking at the differences is one of the objectives when comparing two different things. This is contrast. You must compare to contrast. Get it? Got it? Good. Bolding your words doesn't make what you said true.

Actually, I also don't understand what point you are trying to get across.

Based on your assessment of HD4870 --> HD6870, what does this comparison yield in terms of the expected performance increase from HD6970 to HD7970?

You seem to acknowledge the facts that HD6870 does not have dual graphics engines, does not have HD69xx's ROPs, does not have Cayman's memory controller, does not support double precision, etc. Then, what meaningful information can you derive from looking at the HD6870 that will allow us to make a reasonable prediction of HD7970's performance, itself based on the Cayman and not Barts architecture? All of those features resulted in a much larger die space on the 6970; and they will make their way into the 7970.

Take a step back and look at HD6870 vs. HD5870. After removing 480 SPs, 24 TMUs and reducing memory bandwidth by 12.5% compared to the 5870, the 6870 is only about 10% slower. We may perhaps conclude that the original Cypress VLIW-5 architecture was not a very balanced one. However, this doesn't tell us anything about HD7970's architecture or die space.

 

[Rezist]

Everyone assumes they want to double everything, if indeed at a 256bit bus they become memory limited maybe they will try to design a chip that is maxed out at that speed. The 6950 and 6970 aren't bandwidth limited yet so they can still go faster with there existing memory controller and GDDR5.

I expect 28nm will be smoother then 40 nm was but were going to have 2 "gens" of cards on 28nm for sure. So I expect launch with a card somewhere between 250-300 mm for the high end and a year later follow that up with either a 384bit card at 400mm+ or stick with 256bit at 400mm+ whatever makes sense.

 

[Dudler]

snip...

Based on rumours, Kepler is going to be a significantly faster, brand new GPU architecture from NV. Performance per watt is expected to increase 3-4x compared to Fermi. Double precision performance will also increase significantly. There is just no way AMD can do a half-baked 28nm "refresh" with HD7970 series if they want to stay in the game. If I were to guess at this point, I would say HD7970 may end up 70% faster than HD5870 at 1920x1200 4AA and 80-90% faster at 2560x1600. And even that may not be enough to beat out Kepler.

Not to derail the thread, but your reading that nVidia marketing chart wrong. The increase in perf/watt is from TESLA to Kepler. Aka Gtx280-285.

That means they brag about increasing the performance per watt only 4 times even when the process node has undergone two full node shrinks. Each full node roughly shrinks 50% die size but more importlantly for that chart ups the perf per watt, so if they are bragging about Kepler being 4 times more efficient than tesla, we are going to see Fermi 3.0.

Just look at the 5770, about 90 W at gaming load

http://tpucdn.com/reviews/HIS/HD_5770/images/power_peak.gif

while performing around the same as the 4870. The 4870 uses around 160W.(No reference but TPU didn't measure card consumption so far back. Isle minus peak gives around 150W though and I add 10W for idle)

One shrink and they doubled their performance per watt. Now add another shrink 40nm-->28nm and we have the 4 time better perf per watt. Kepler is not going to be special unless they come up with an better architecture. Ofc I am assuming Kepler will be 28nm. If it comes out on 40nm I will be impressed with their perf per watt increase.

 

[Topweasel]

Since 28nm will double the transistor density, If we shrink Cayman (HD6970) at 28nm it could be close to 190-200mm2, down from 389mm2 (40nm).
By the way that will be a nice Middle End card later on

Do you think that 50-60mm2 more will add that much in order to get the performance crown from GTX580 and by how much ?? 10-20% ??

And then 4-6 months later Kepler will come and will obliterate the 79xx, no way
I see a 300-350mm2 from AMD this round.

But that would only be the case if they used Cayman. Right now is first time they have used 200mm plus card for the "middle end" cards and I think we are the ones backwards. Its more that cayman is the ultra high end series and that the 6800 is supposed to be the just plain high-end chip. I think the 6800 series is the spiritual replacement for the 4800 series. That the die shrunk card will be the Bart chip and that we won't see a "shrunken" cayman till the 8k series where I would expect I 300+ mm high end chip. I think that will become their new system. It also ramps down the increase in video power a tad to let software catch up.

 

[Idontcare]

I expect 28nm will be smoother then 40 nm was but were going to have 2 "gens" of cards on 28nm for sure. So I expect launch with a card somewhere between 250-300 mm for the high end and a year later follow that up with either a 384bit card at 400mm+ or stick with 256bit at 400mm+ whatever makes sense.

Yikes that's a tall order. 40nm -> 28nm changes are even more complex than the changes that were implemented in order to enable the 55nm -> 40nm transition.

They won't repeat the same mistakes, that is for sure, but there will be plenty of new ones to discover for the very first time.

 

[purefun1965]

28 nm is the last of big gaming die sizes. Nvidia is moving to hpc and amd is moving to on die graphics. The gaming market is not using features now. why build more? Nvidia needs hpc because they dont have x86. Amd needs to move smaller due to netbooks and such. The gaming market cant sustain new features if they wont implement them. They cater to consoles.

 

[purefun1965]

To quote a singer songwriter from the sixty's Bob Dylan " The times they are a changing"