Amd Volcanic Islands details

[tajoh111]

From Forbes: "Exclusive Interview: AMD's Matt Skynner Talks New Radeon Cards, Next-Gen Consoles, 7990 Criticism."

He gives a bit of information about AMD's upcoming GPUs.

That's around 425-435 mm^2 for (what I assume to be) Hawaii then. (I've seen different places say 551 mm^2 or 561 mm^2 for GK110.)

This die size is kind of worrisome. 20 percent increase in die size won't lead to a 40% increase in performance some are lead to believe.

[iMacmatician]

it is bull, if its Hawaii die size = GK110 - 30 percents, has the same die size like Tahiti, no more SPs, no more ROPs, it cant be true

Matt Skynner from AMD is saying BS?

[PedantOne]

Matt Skynner from AMD is saying BS?

i wish because with that die size cannot compete with GTX 780 or Titan

[LordEC911]

i wish because with that die size cannot compete with GTX 780 or Titan

Lawl. You are funny.

[Ace123]

Interesting comment about targeting enthusiasts and not "Ultra Enthusiasts". This is a mistake in my opinion.
Kelt Reeves founder of Falcon Nw had this to say on that subject.
?The Nvidia GeForce Titan is our #1 selling graphics card by far ? despite it being the most expensive,? Reeves reveals. ?What?s more, our customer for this card is actually buying an average of 1.3 Titans per PC. In other words, a huge number of them are buying two and even three of these monster GPUs in a single system.?

Someone should send this link to Matt skynner
http://www.forbes.com/sites/jasoneva...to-the-rescue/

[raghu78]

AMD Hawaii with a die size of 430 - 440 sq mm might be able to compete with GTX 780 on a clock for clock basis. But the fact is Nvidia will easily release a 2688 sp or even 2880 sp GTX 785 and put AMD a distant second. Its disappointing that AMD did not go for a bigger chip around 470 - 480 sq mm. AMD have once again failed to aggressively go for the GPU crown against Nvidia's top flagship. AMD is once again left to look at pricing and game bundles to improve its value proposition.

[boxleitnerb]

At 2560 SP they might be faster at similar clocks. As far as I've heard they will also clock higher than the GTX 780. It's still possible that they compete with Titan and even beat it by a percent or two.

[tajoh111]

i wish because with that die size cannot compete with GTX 780 or Titan

It might be able to compete, but I would imagine it would need to be clocked much closer to the thresholds of its clocking limits. Under air, Titan can get into the 1300mhz range with a bit more voltage and the right bios' currently. The highest I have seen from overclock.net is 1400 mhz under air. However stock clocks are 900mhz. I don't think next gen GCN(increased front end) is going to do AMD any favors for overclocking ability and we might see the opposite.

If Hawaii is clocked at 1050mhz-1100mhz. It should compete with the gtx 780(maybe even beat it slightly), however it will consume a lot of power doing so and not have nearly as much overclocking room as a gk110. Of course, if AMD blocks access to Nvidia for battlefield 4 which will surely be in reviews, we should see a unusual 20 percent lead(particularly if the game is dropped on last minute and reviewers use only a canned benchmark). Aside from drivers, the biggest things saving AMD lately in reviews are Gaming evolved games which Nvidia can't perform well in(sleeping dogs, Dirt showdown, Hitman Absolution). Add one title in you add 3 percent or so performance difference for the review. Add all three and you can see a close to 10 percent change in averages for a review. If battlefield 4 has initial performance similar to tomb raider canned benchmark when Nvidia wasn't given access to it, you can add another 3 percent to that average. Right now, Nvidia doesn't have any titles that act like sleeping dogs or dirtshowdown, with global illumination that cripple Nvidia performance.

I think AMD is going to go open air cooler for most of their hawaii line up considering how bad a blower fan worked on a 7970, particularly the ghz edition.

[LordEC911]

Interesting comment about targeting enthusiasts and not "Ultra Enthusiasts". This is a mistake in my opinion.
Kelt Reeves founder of Falcon Nw had this to say on that subject.
?The Nvidia GeForce Titan is our #1 selling graphics card by far ? despite it being the most expensive,? Reeves reveals. ?What?s more, our customer for this card is actually buying an average of 1.3 Titans per PC. In other words, a huge number of them are buying two and even three of these monster GPUs in a single system.?

Someone should send this link to Matt skynner
http://www.forbes.com/sites/jasoneva...to-the-rescue/

Yes, because when you are designing/building/selling ultra enthusiast overpriced highend botique PCs to a niche market it is surprising that the most expensive consumer GPU is their #1 seller?

*SNIP*

You are making a lot of assumptions about an almost completely unknown entity.

Edit- So assuming somewhere around 430-440mm2, 5.5-6b transistors, compared to the 4.31b on Tahiti.
I would guess somewere around 5.8b transistors.

[Iconyu]

Interesting comment about targeting enthusiasts and not "Ultra Enthusiasts". This is a mistake in my opinion.
Kelt Reeves founder of Falcon Nw had this to say on that subject.
?The Nvidia GeForce Titan is our #1 selling graphics card by far ? despite it being the most expensive,? Reeves reveals. ?What?s more, our customer for this card is actually buying an average of 1.3 Titans per PC. In other words, a huge number of them are buying two and even three of these monster GPUs in a single system.?

Someone should send this link to Matt skynner
http://www.forbes.com/sites/jasoneva...to-the-rescue/

If AMD can produce another 4870, then I don't think they have much to worry about.

[tajoh111]

Yes, because when you are designing/building/selling ultra enthusiast overpriced highend botique PCs to a niche market it is surprising that the most expensive consumer GPU is their #1 seller?


You are making a lot of assumptions about an almost completely unknown entity.

Edit- So assuming somewhere around 430-440mm2, 5.5-6b transistors, compared to the 4.31b on Tahiti.
I would guess somewere around 5.8b transistors.

Something like a 20 percent increase in die size isn't going to be a magic bullet for 40 percent more performance, particularly at the clocks 7970 ghz is currently set at and how much performance it has.

I remember you were particularly optimistic with the performance difference between 5870 and 6970. That had a 16.4% difference in die size and we got something along that line for the performance difference(you expected much greater).We probably get less than that 16.4% Probably a bit below it if we take an average across the board.

With this 430ish die size coming from a particularly reliable source, Tahiti isn't going to get the massive spec jump like between gk110 and gk104. So it can't afford the low clocks like gk110. However I am guessing it can't particular afford to keep clocks high considering the relatively high power consumption of a 7970 ghz and the increased die size over 7970, increasing the complexity of the gcn cores. The better front end might increase utilization and efficiency of resources, but powers going to go up quite a bit like the 5870 to 6970(this went up more than the performance in some reviews).

I still stick by my 18% performance increase over a ghz edition(it might be greater if Battlefield 4 is mega optimize benchmark). But for the most part in GPU design history, the larger the size of a chips predecessor, the less each companies has been able to wring out for its next gen, particularly on the same process.

Somehow getting 40% more performance from only a 20 percent increase in die size on the same process is unprecedented. And with GCN not having a lead for performance per die size this generation and this relatively modest increase in die size, even with increased efficiency, its going to be hard for AMD to make the Iceberg you so desire.

[jvc]

Something like a 20 percent increase in die size isn't going to be a magic bullet for 40 percent more performance, particularly at the clocks 7970 ghz is currently set at and how much performance it has.

I remember you were particularly optimistic with the performance difference between 5870 and 6970. That had a 16.4% difference in die size and we got something along that line for the performance difference(you expected much greater).We probably get less than that 16.4% Probably a bit below it if we take an average across the board.

With this 430ish die size coming from a particularly reliable source, Tahiti isn't going to get the massive spec jump like between gk110 and gk104. So it can't afford the low clocks like gk110. However I am guessing it can't particular afford to keep clocks high considering the relatively high power consumption of a 7970 ghz and the increased die size over 7970, increasing the complexity of the gcn cores. The better front end might increase utilization and efficiency of resources, but powers going to go up quite a bit like the 5870 to 6970(this went up more than the performance in some reviews).

W/o any compute enhancements 1.5b trans might just be enough for an extra 1024 SP and 16 ROPs. This coupled with arhitecture enhancements and less redundancy and we could be in for a surprise. Fingers are crossed.

[LordEC911]

Something like a 20 percent increase in die size isn't going to be a magic bullet for 40 percent more performance, particularly at the clocks 7970 ghz is currently set at and how much performance it has.

Where does this 40% increase keep coming from? Why are we comparing Tahiti clocks and performance? Hawaii is a different ASIC with an evolved/tweaked architecture.

I remember you were particularly optimistic with the performance difference between 5870 and 6970. That had a 16.4% difference in die size and we got something along that line for the performance difference(you expected much greater).We probably get less than that 16.4% Probably a bit below it if we take an average across the board.

Average is/was at least +20% over 5870. Ibiza performance targets got leaked and mixed in with the Cayman leaks...
I have apologized about that on numerous occasions.

With this 430ish die size coming from a particularly reliable source, Tahiti isn't going to get the massive spec jump like between gk110 and gk104. So it can't afford the low clocks like gk110. However I am guessing it can't particular afford to keep clocks high considering the relatively high power consumption of a 7970 ghz and the increased die size over 7970, increasing the complexity of the gcn cores. The better front end might increase utilization and efficiency of resources, but powers going to go up quite a bit like the 5870 to 6970(this went up more than the performance in some reviews).

We will have to wait and see. There are numerous optimizations/efficiency gains that have been added to GCN since Tahiti was released.

I still stick by my 18% performance increase over a ghz edition(it might be greater if Battlefield 4 is mega optimize benchmark). But for the most part in GPU design history, the larger the size of a chips predecessor, the less each companies has been able to wring out for its next gen, particularly on the same process.

That is great, you do that.

Somehow getting 40% more performance from only a 20 percent increase in die size on the same process is unprecedented. And with GCN not having a lead for performance per die size this generation and this relatively modest increase in die size, even with increased efficiency, its going to be hard for AMD to make the Iceberg you so desire.

~30% more performance with a 51% larger die. GK110 isn't exactly holding a mm2/performance lead.

[tajoh111]

W/o any compute enhancements 1.5b trans might just be enough for an extra 1024 SP and 16 ROPs. This coupled with arhitecture enhancements and less redundancy and we could be in for a surprise. Fingers are crossed.

Simply adding more 1024 more cores and 16 more ROP without changing the architecture would lead to a loss for AMD. Per core GCN is less efficient than Nvidia, core scaling(look at a gtx 650ti and gtx 680 performance difference). Also there is a good chance more bottle necks would pop up without doubling everything else. Simply adding more 50%GCN cores + ROP would not lead to close to a 50% increase in performance.

There's a good chance if AMD somehow managed to double up every spec on the 7970(keeping the architecture the exact same) like gk104 to Gk110, it would still lose(particularly because a 600mm2 Plus die would need low clocks) because of inefficiency in scaling up the cores. A change/improvement in architecture is the better route because of the relatively large die size of the 7970 and AMD not having the a Professional market to support the wasteful nature of large dies. The Professional market is how Nvidia affords the R and D and poor yields of big chips. Particularly in this market, doubling up specs lead to more impressive performance gains in applications vs gaming ones.

[tajoh111]

Where does this 40% increase keep coming from? Why are we comparing Tahiti clocks and performance? Hawaii is a different ASIC with an evolved/tweaked architecture.


Average is/was at least +20% over 5870. Ibiza performance targets got leaked and mixed in with the Cayman leaks...
I have apologized about that on numerous occasions.


We will have to wait and see. There are numerous optimizations/efficiency gains that have been added to GCN since Tahiti was released.


That is great, you do that.


~30% more performance with a 51% larger die. GK110 isn't exactly holding a mm2/performance lead.

40% over a ghz edition is what is needed for Hawaii to beat titan and not simply be comparable.

Titan chips are pretty much bad gk110 chips. They usually range in the 60-74% range ASIC range and are not fully enabled. They are likely the left over parts from the pro market.

Take a fully enabled gk110 and take a better asic you could see a 50 percent performance gain over a ghz edition while pushing the chips similarly hard(a fully enabled gk110 vs tahiti xt). EVGA superclocked Gtx 780 with ACX coolers already have close to a 40% increase in performance over a 7970 ghz edition. And this is with 25% percent of gk110 cores disabled(2304 vs 2880).

Gk110 is so under-clocked its not even funny(when was the last time this big of a chip was able to top performance per watt charts). Them still using a super quiet blower fan vs an open air for such a large chip means they have a lot left in the tank.

[m_jones_]

Where are you pulling out this figure that Titan is 40% faster than the 7970, from most of the benchmarks i have looked at this really is not the case especially for the ghz model.

[dasa]

Where are you pulling out this figure that Titan is 40% faster than the 7970, from most of the benchmarks i have looked at this really is not the case especially for the ghz model.

maybe 40% if at the same clock speed as titan and since 7970ghz is clocked 20% faster...
there is some games where titan does win by around 40% although on average it may be closer to 25%
so it would need to be 40% faster to win outright

[GoldenTiger]

Gk110 is so under-clocked its not even funny(when was the last time this big of a chip was able to top performance per watt charts). Them still using a super quiet blower fan vs an open air for such a large chip means they have a lot left in the tank.

Spot-on... it's hilariously underclocked considering what virtually all of the cards can reach (Sandy-Bridge-like easy OC's). Personally I'm using a no-throttle BIOS and running at 1280mhz core fulltime boost, with 7048mhz mem (tested to ensure it's not error-correcting and is still boosting performance), for a 24/7 OC with a voltage bump. Card's not whisper-quiet at those speeds but it's not what I'd call loud either (ACX-cooler Superclock eVGA card). Voltage isn't even crazy either to be honest...

I also normally run for gaming a whisper quiet profile at 1150c and around 1500rpm or less fan speed on the acx, 7.1ghz mem. While the other is stable I am a quiet freak so I value silence over the performance bump. This profile runs at 1.175v.

[tajoh111]

maybe 40% if at the same clock speed as titan and since 7970ghz is clocked 20% faster...
there is some games where titan does win by around 40% although on average it may be closer to 25%
so it would need to be 40% faster to win outright

Titan is somewhere around 32-33 percent faster than a 7970 ghz at the highest resolutions. Particularly if you have a wider range of games to dilute the effect of something like dirt showdown. If hawaii is 2 percent faster than titan, most would not consider it faster. Most people will consider it more or less the same speed. It needs to have something tangible and to me that is 40 percent faster than a 7970ghz.

[Mechanical Man]

40% over a ghz edition is what is needed for Hawaii to beat titan and not simply be comparable.

Titan chips are pretty much bad gk110 chips. They usually range in the 60-74% range ASIC range and are not fully enabled. They are likely the left over parts from the pro market.

Take a fully enabled gk110 and take a better asic you could see a 50 percent performance gain over a ghz edition while pushing the chips similarly hard(a fully enabled gk110 vs tahiti xt). EVGA superclocked Gtx 780 with ACX coolers already have close to a 40% increase in performance over a 7970 ghz edition. And this is with 25% percent of gk110 cores disabled(2304 vs 2880).

Gk110 is so under-clocked its not even funny(when was the last time this big of a chip was able to top performance per watt charts). Them still using a super quiet blower fan vs an open air for such a large chip means they have a lot left in the tank.

You really think AMD has been sitting on its ass, doing nothing? I dont get this thinking. We cant know what kind of modifications and bottleneck removals has AMD been doing compared to GCN 1.0. 7870 was on par performance per watt basis against GTX680. 7970 was not, becouse of the compute overhead. I dont see how they would release same performing GPU on same process node. It would make no sense what so ever. They have had lot of time to better the the desing and remove bottlenecs and power consumption. Also 28nm process is way more mature and their understanding of that process is way better at this point.

[BababooeyHTJ]

Spot-on... it's hilariously underclocked considering what virtually all of the cards can reach (Sandy-Bridge-like easy OC's). Personally I'm using a no-throttle BIOS and running at 1280mhz core fulltime boost, with 7048mhz mem (tested to ensure it's not error-correcting and is still boosting performance), for a 24/7 OC with a voltage bump. Card's not whisper-quiet at those speeds but it's not what I'd call loud either (ACX-cooler Superclock eVGA card). Voltage isn't even crazy either to be honest...

I also normally run for gaming a whisper quiet profile at 1150c and around 1200rpm or less fan speed on the acx, 7.1ghz mem. While the other is stable I am a quiet freak so I value silence over the performance bump. This profile runs at 1.175v.

No, 7950 is hilariously underclocked. My titan on the other hand wasn't stable at over 1167mhz and that was on water with a modded bios. That seemed to be about average. Most people running an oc like that aren't going to brag on the forums. I see it all of the time especially with cpus. Also GK110 runs hot as balls at 1.2v.

As is if you check some 7970 ghz edition and GTX780 benchmarks and then go look at the old 4870 reviews the difference is very similar. Price difference is also very similar.

[tajoh111]

No, 7950 is hilariously underclocked. My titan on the other hand wasn't stable at over 1167mhz and that was on water with a modded bios. That seemed to be about average. Most people running an oc like that aren't going to brag on the forums. I see it all of the time especially with cpus. Also GK110 runs hot as balls at 1.2v.

As is if you check some 7970 ghz edition and GTX780 benchmarks and then go look at the old 4870 reviews the difference is very similar. Price difference is also very similar.

You got a dud of a titan there. I am pretty sure you have been to overclock.net. With the right bios firmwares(which have been updated within the last month), everyone and I mean everyone is getting 1200mhz lately. Some are reaching 1300+. With water getting to or past 1400mhz is not unheard of.

I know the gtx 780 isn't titan, but pretty much everyone that owns a evga classified edition is getting to 1300 with a bit of work, some more than others.

Some of the early Titans got a really low asics and even then the lowest I have seen clocked is 1126 range or something. I have seen plenty of 7950 hit only these type of clocks to because of voltage locks.

But as far as clocks go, gk110 goes up higher percent wise than tahiti. Go to HWBOT and you will see the top gk110's top the top 7970's as far as pure frequency goes which is why it will be difficult for AMD to beat with hawaii.

[BababooeyHTJ]

I kept my eye on the forums when Titan was released. Most wouldn't do 1200mhz.

I have seen plenty of 7950 hit only these type of clocks to because of voltage locks.

I've owned four 7950s, every one did at least 1200mhz on water. thats a 50% overclock from stock clocks. or a 30% overclock over boost edition.

But as far as clocks go, gk110 goes up higher percent wise than tahiti.

Lol, ok.

[tajoh111]

I kept my eye on the forums when Titan was released. Most wouldn't do 1200mhz.



I've owned four 7950s, every one did at least 1200mhz on water. thats a 50% overclock from stock clocks. or a 30% overclock over boost edition.



Lol, ok.

At the beginning when Titan was released the cards clocked between 1126- 1226. But things have improved since then. More recent bios have come out and unlocked alot more potential. In addition, I am starting to see better asics making it into GK110 gaming cards. For a monster chip like gk110, your likely to see the biggest improvements overtime in regards to yields and clocking. Particularly when Nvidia professional lineup is eating the best chips. The most recent gk104's don't really clock better than the old ones, but its seems recently gk110 has gotten better.

Some 7950's are famous for not hitting high clocks. The ones that are know not to do this are a particular model of xfx double d edition which lock the volts and don't use a reference PCB. It is a popular model because of its low price still. I believe there are certain sapphire cards that have limited overclocking potential because of the PCB and voltage locking.

http://www.xtremesystems.org/forums/...Vapor-X-Review

Here is a guide. But with later 7950's. AMD was more restrictive with voltage and alot of newer cards come voltage locked.

http://www.overclock.net/t/1394623/b...guide-to-7950s

1200 on water isn't that impressive for any card this generation. Saying the 7950 is underclocked and saying all them reach 1200 with water isn't that impressive, if this was air I would be more impressed. The base 7950 just doesn't perform that well at completely stock clocks and they need overclocking for sure. I agree that stock 7950's and 7970's are underclocked. But not so much their ghz(1050MHZ) or boost editions(950mhz) which are becoming more and more the norm.


I am talking about the 7970 ghz edition, which I have been using as a baseline as far as performance goes all thread. Some 7970 ghz editions in reviews get like 100mhz out of the cards and that's it.

[PedantOne]

Tahiti at 1,05 GHz is on maximal TDP, imagine +20 percent bigger chip with higher clocks = 30 percent above Tahiti TDP = First One GPU card with TDP of dual card? - Impossible.