(8-17-13) AMD 9970 28nm vs. 20nm Poll

[wlee15]

I wouldnt call the RV740 a pipecleaner. It was so disasterous that the product got replaced with 55nm. RV740 was a good example on what happens when the process node is not even ready.

There were no further new 55nm GPU from ATI after the launch of the HD 4770. The last 55nm GPU was the HD 4890 which launched approximately 1 month before the HD 4770.

 

[ShintaiDK]

There were no further new 55nm GPU from ATI after the launch of the HD 4770. The last 55nm GPU was the HD 4890 which launched approximately 1 month before the HD 4770.

HD4860 and HD4730 launched after the HD4770. They simply couldnt produce the HD4770 and had to replace it pricewise with a 55nm GPU.

 

[Vesku]

I don't think the idea that AMD has potentially gone for a big die with the 9970 can be easily dismissed, at least until AMD reveals the actual specs. AMD wants to make gains in the HPC market.

 

[sushiwarrior]

go ahead and write it down folks:

28nm
2560 sp's
160 TMU's
40 ROP's
1000 GPU clock
6000 MHZ RAM clock
3 Gig DDR5
384 bit memory bus

1/8 would not write down

 

[RussianSensation]

during the 28nm gen AMD has kept the GPU crown for the same time as Nvidia. also 28nm has another 9 - 12 months of lifecycle.

I am talking about AMD's best foot forward vs. NV's on the same node. GTX280 > 4870, GTX285 > 4890, GTX480 > 5870, GTX580 > 6970, GTX780 > 7970GE. In each generation, NV is always on top 15% at least. GTX680 vs. 7970 doesn't really count since it's more of a flagship delay by NV than NV's inability to compete. In fact, with 780, NV crushed AMD by an embarrassing amount.

yes AMD has not replied to GK110 yet. but thats what they are doing with Hawaii. in Oct we will know if Nvidia is still on top for the rest of 28nm gen.

I'll give them the benefit of the doubt and see what they can come up with 9970 but I bet 780 OC will beat 9970 OC by 15% still.

you are not able to see AMD gain 35 - 40% because of very rigid views that AMD will not go big die.

I am not saying AMD will not increase die size, but I am not sure I believe they'll make a chip with 470-480mm2 right now.

you are also of the opinion that AMD will not improve perf/sq mm over Tahiti. that I feel is ridiculous. Tahiti is the most inefficient GCN chip because of an unbalanced GPU and is the easiest to improve.

Nope, I never said that at all in relation to Tahiti. I am saying that while they will improve performance/watt and perf/mm2 over Tahiti, I don't think 9970 OC will match 780 OC. My view is they'll go bigger chip but not big enough that they'll be able to get away with only 975mhz clocks. I believe their clocks will be higher which will leave a lot less room for overclocking. I am betting that they will have very aggressive Turbo similar to what NV now has with Kepler. With NV since they made such a large 561mm2 chip, they could afford to clock it in mid-800s/low-900s. I do not believe AMD will have this die size luxury. As a result, they'll have to settle between a balance of die size and clock speed.

remember sp are very high density and take only a 1/3rd of total die space. so 2048 sp is roughly 110 sq mm. add another 37.5% sp for roughly 41 sq mm. throw in doubling of front end to improve perf/sp and increasing ROPs by 50% for another 55 - 60 sq mm.its a chip which would be 450 - 480 sq mm.

For 5 AMD generations we got hype - 2900XT, 3870, 4870, 5870, 6970. None of those chips lived up to the hype, always beaten by NV. Why would I bet that this generation suddenly AMD will beat NV's best card if they haven't done it for 5 generations in a row?

also AMD has been steadily moving away from sweet spot and going in the direction of larger GPUs. also by your logic there can be no surprises at all. so you believe GK110 is unbeatable. say it out and I will hold you to it when we revisit the situation in mid Oct.

6970 was 389mm2, but 7970 is only 365mm2. They haven't been moving up. Yes, in my opinion GK110 overclocked will beat 9970 overclocked. I'll say a 1.3Ghz 780 will beat 9970 overclocked even, not even the full blown 2880 SP GK110.

 

[sushiwarrior]

*snip*

I feel like almost all of what you just said hinges off the fact that

A. Nvidia usually releases their cards AFTER AMD, hence they have a few months advantage
B. Nvidia's flagships are almost universally more expensive than AMD's
C. Nvidia's flagships are also usually bigger, use more power, etc.

So essentially Nvidia has more time, more die size, more power budget, more cost budget, more of everything essentially. If Nvidia somehow STILL didn't have the fastest card, then I wouldn't even know what to believe!

 

[RussianSensation]

So essentially Nvidia has more time, more die size, more power budget, more cost budget, more of everything essentially. If Nvidia somehow STILL didn't have the fastest card, then I wouldn't even know what to believe!

I don't deny that. I keep buying AMD cards because they keep delivering the best price/performance + overclocking every generation so far in my price range. If AMD beat NV for flagship GPU performance, NV should be embarrassed considering that they can afford to make 550-560mm2 GPUs courtesy of Quadro/Tesla lines. If AMD could make a 550-560mm2 Tahiti style chip, NV would need to find a cave to crawl into. Luckily for NV, AMD has been unable to make an efficient and powerful chip bigger than 410mm2 for 6-7 years now. If AMD could somehow make a 550-560mm2 GCN next chip on 28nm, NV would need Alien technology to compete.

If you look at it, the transistors AMD and NV use must have different properties. NV's 780 in HOF form can hit 1300mhz on 1.20V despite 561mm2 size. AMD's 7970 cannot hit 1300mhz even on 1.30V. When both are overclocked, despite 7970's 365mm2 die size, the power consumption between them is very similar. AMD's transistor density and transistor composition seems inferior to NV at the moment when it comes to their relationship to power consumption. I haven't dug deeper, but it's even possible that NV turns off some DP transistors to save power consumption in games but those transistors are emitting power on the AMD side. There discrepancy between clock speeds vs. die size vs. power consumption show how grossly power inefficient Tahiti is vs. GK110.

 

[tviceman]

This thread is full of ridiculous un-grounded hope. Hope that AMD can catch Nvidia in the single crown arena when comparing the fastest GPU's when the releases are through, hope that AMD will make a die bigger than the R600 (which was a huge, huge flop), hope that it will be out this fall on a node process that isn't ready yet, and hope that little red piggies will take to the sky and shoot down the horrible green goblin JHH minions.

AMD's R&D has been cut making a big die GPU (the biggest GPU AMD/ATI has ever made according to some predictions in this thread) 20 months into a node process a whismical fantasy and Roy Read made it a point when he was hired to make sure existing nodes are fully utilized before moving on to new ones. This is the first time since the 8800 series where Nvidia's lead is large enough they can rest on their laurels, releasing cut down chips as their series flagships and charging ultra premiums for it at the same time. Hawaii is going to come out, deliver 780 performance for $100 less, force some price cuts on the Geforce 700 series cards, and then Nvidia will respond 8-10 weeks later with a 14-15 SMX GTX 785 for $699.

Anyone that already has hd7870 / gtx660 or better will find nothing worthwhile with these new cards even after a small / moderate price war is initiated. 20nm can't come soon enough.

 

[tviceman]

If AMD could somehow make a 550-560mm2 GCN next chip on 28nm, NV would need Alien technology to compete.

Funny I've thought the same thing before but I don't really believe that is the case. It's just not comparable in that regard. Considering GK104, if Nvidia had built it to be the same size at Tahiti by adding another memory controller and SMX, then it would have completely obliterated Tahiti and we'd be saying the same thing about AMD needing alien technology to keep up with GK104 blah blah.

Also just imagine the power draw of Tahiti scaled up to 550mm^2. Already at the 365ish mm^2 it is now, power draw per mm^2 compared to GK110 is absolutely awful. And now I see you said that in your post with your ninja edit. Hahaha.

 

[raghu78]

I am talking about AMD's best foot forward vs. NV's on the same node. GTX280 > 4870, GTX285 > 4890, GTX480 > 5870, GTX580 > 6970, GTX780 > 7970GE. In each generation, NV is always on top 15% at least. GTX680 vs. 7970 doesn't really count since it's more of a flagship delay by NV than NV's inability to compete. In fact, with 780, NV crushed AMD by an embarrassing amount.

firstly right from HD 4870 upto HD 6970 , Nvidia had a much larger die. the difference was massive at times. so it would be embarassing to nvidia if they did not beat AMD by atleast 15%. also times have changed and strategy most definitely. AMD realizes the dGPU market is shrinking and they need more of high margin revenue to sustain R&D , which means they need to take Nvidia head on at the high end.

HD 4870 - 256 sq mm GTX 280 - 576 sq mm
HD 4890 - 282 sq mm GTX 285 - 470 sq mm
HD 5870 - 334 sq mm GTX 480 - 529 sq mm
HD 6970 - 389 sq mm GTX 580 - 520 sq mm

the closest gen was HD 6970 / GTX 580 in die size. but still Nvidia had a 33% larger die. also with Fermi Nvidia realized that massive 500+ sq mm on a raw process are a recipe for disaster. yields are a problem as it is at the leading edge node at the start of its lifecycle. let me say this the days of a 500+ sq mm chip on a new process are gone. so stop calling it delay.you think Nvidia would have not released a GK100 if they could. they would have charged the same ridiculous USD 1000 or even more. but guess what Nvidia just could not deliver a GK100.

I'll give them the benefit of the doubt and see what they can come up with 9970 but I bet 780 OC will beat 9970 OC by 15% still.

thats absolute crap. so you believe only a 780 can clock to 1.3 ghz or you believe clock for clock GTX 780 is 15% faster than HD 9970. remember GTX 780 is boosting to 1006 mhz out of the box.

http://www.anandtech.com/show/6973/nvidia-geforce-gtx-780-review/3

a stock HD 9970 running at 925 mhz and boosting to 975 mhz which is 35 - 40% faster than HD 7970 Ghz will beat any golden GTX 780 provided you compare a golden HD 9970 OC vs golden GTX 780 OC. you cannot beat a chip which is faster clock for clock unless you are guaranteed to have a much higher clock ceiling. since these are all TSMC 28nm chips they all are capable of hitting 1.3 ghz. its plain silicon binning and the lottery. if a 7 billion transistor chip can hit 1.3 ghz so can a 5.5 - 6 billion transistor chip. remember Tahiti, Pitcairn,Cape Verde ,Bonaire, GK110, GK104, GK106 have all now been shown to hit 1.25 - 1.3 ghz. what makes you believe Hawaii cannot.

I am not saying AMD will not increase die size, but I am not sure I believe they'll make a chip with 470-480mm2 right now.

so we willl have to wait for the facts in mid oct.

Nope, I never said that at all in relation to Tahiti. I am saying that while they will improve performance/watt and perf/mm2 over Tahiti, I don't think 9970 OC will match 780 OC. My view is they'll go bigger chip but not big enough that they'll be able to get away with only 975mhz clocks. I believe their clocks will be higher which will leave a lot less room for overclocking. I am betting that they will have very aggressive Turbo similar to what NV now has with Kepler. With NV since they made such a large 561mm2 chip, they could afford to clock it in mid-800s/low-900s. I do not believe AMD will have this die size luxury. As a result, they'll have to settle between a balance of die size and clock speed.

to remind you GK110 is not a very efficient chip from a perf / sq mm perspective. thats because whatever compute features they stripped off from GK104 to make it lean and efficient they had to put it back and some more for GK110 to make it a compute monster. GK104 - 294 sq mm. GK110 - 561 sqmm. thats a 90% larger die. Titan has 75% more sp than GTX 770. performance is 35% more than GTX 770.

In fact Pitcairn (212 sq mm) to Tahiti(365 sq mm) was a similar 35 - 40% improvement for a roughly 72% larger die size. Tahiti has borne a lot of the luggage of a compute chip - wide memory bus, double precision , etc while not having enough raw sp and other front end resources (ACE, geometry engines,rasterizer) / back end resources (ROPs) to drive perf to a good perf / sq mm level. AMD will show with Hawaii vs GK110 as they did with Pitcairn vs GK106 that they have better perf/sqmm and similar perf/watt.

For 5 AMD generations we got hype - 2900XT, 3870, 4870, 5870, 6970. None of those chips lived up to the hype, always beaten by NV. Why would I bet that this generation suddenly AMD will beat NV's best card if they haven't done it for 5 generations in a row?

2900 was a disaster. From there AMD has steadily improved and with HD 7970 / HD 7970 ghz for the first time they held the GPU crown. what makes you believe Hawaii will not match or even go one up on GK110.

6970 was 389mm2, but 7970 is only 365mm2. They haven't been moving up. Yes, in my opinion GK110 overclocked will beat 9970 overclocked. I'll say a 1.3Ghz 780 will beat 9970 overclocked even, not even the full blown 2880 SP GK110.

compare 1st gen 40nm chip to 1st gen 28nm . so wrt HD 5870 to HD 7970 yeah they went up. now with Hawaii 28nm they are going to be even more bigger than ever before. most probably their first 450+ sq mm chip. mark your statement down. lets revisit in Oct.

 

[sushiwarrior]

I don't deny that. I keep buying AMD cards because they keep delivering the best price/performance + overclocking every generation so far in my price range. If AMD beat NV for flagship GPU performance, NV should be embarrassed considering that they can afford to make 550-560mm2 GPUs courtesy of Quadro/Tesla lines. If AMD could make a 550-560mm2 Tahiti style chip, NV would need to find a cave to crawl into. Luckily for NV, AMD has been unable to make an efficient and powerful chip bigger than 410mm2 for 6-7 years now. If AMD could somehow make a 550-560mm2 GCN next chip on 28nm, NV would need Alien technology to compete.

If you look at it, the transistors AMD and NV use must have different properties. NV's 780 in HOF form can hit 1300mhz on 1.20V despite 561mm2 size. AMD's 7970 cannot hit 1300mhz even on 1.30V. When both are overclocked, despite 7970's 365mm2 die size, the power consumption between them is very similar. AMD's transistor density and transistor composition seems inferior to NV at the moment when it comes to their relationship to power consumption. I haven't dug deeper, but it's even possible that NV turns off some DP transistors to save power consumption in games but those transistors are emitting power on the AMD side. There discrepancy between clock speeds vs. die size vs. power consumption show how grossly power inefficient Tahiti is vs. GK110.

Blimey, a cherry picked max bin card can hit 1300, and the average 7970 can only hit 1210. Excellent job comparing golden cream of the crop apples to average oranges. A golden 7970 could easily hit 1300MHz given proper cooling, for example the average overclock on HWBOT under water is 1290MHz.

 

[Skurge]

Funny I've thought the same thing before but I don't really believe that is the case. It's just not comparable in that regard. Considering GK104, if Nvidia had built it to be the same size at Tahiti by adding another memory controller and SMX, then it would have completely obliterated Tahiti and we'd be saying the same thing about AMD needing alien technology to keep up with GK104 blah blah.

Also just imagine the power draw of Tahiti scaled up to 550mm^2. Already at the 365ish mm^2 it is now, power draw per mm^2 compared to GK110 is absolutely awful. And now I see you said that in your post with your ninja edit. Hahaha.

Well, simplistically. If you just take Pitcairn and straight up double it you'll end up with a 430mm^2 GPU that would be faster than a Titan. While using about 220W of power at most. AMD stated the made huge improvements to GCN's efficiency moving from Tahiti to Pitcairn. Those lessons have not yet been applied to Tahiti. So I can see Hawaii at 410/430mm^2 beating Titan/780. If they go even bigger and match 550mm^2... Well you get my drift.

That is an optimistic way of looking at it. Tahiti is the Exception, not the rule to 28nm. Every other GPU from either NV or AMD is far more efficient that Tahiti.

 

[raghu78]

Blimey, a cherry picked max bin card can hit 1300, and the average 7970 can only hit 1210. Excellent job comparing golden cream of the crop apples to average oranges. A golden 7970 could easily hit 1300MHz given proper cooling, for example the average overclock on HWBOT under water is 1290MHz.

the most important point is Tahiti was the first GPU chip out of TSMC on 28nm. AMD must have taped out Tahiti in March 2011. AMD's decision to do that I believe cost them. the better option always is to tape out a smaller chip and then from the learning improve the bigger chip. just like HD 4770 to HD 5870. AMD's Tahiti chip design must have been very conservative with gobs of redundancy to improve yields and also AMD's understanding of the process limited as it was their first 28nm GPU tapeout. remember the first Tahiti chips came at 1.175v and clocked 1100 - 1150 mhz with stock voltage. most GTX 780 cards are hitting 1150 - 1200 mhz and not 1.3 Ghz at 1.2v. a cherry picked review sample is not the norm.

also the 28nm process would have been in better shape when GK110 taped out. the GK110 design would have been able to incorporate the 28nm learning from the earlier designs like GK104 and GK106/GK107.

 

[SiliconWars]

I am talking about AMD's best foot forward vs. NV's on the same node. GTX280 > 4870

Most sites had that as a win for the 4870, by a distance. A vastly smaller die that was within 10-15%. The 4870x2 also crushed the 280 and basically tied with the 295 while being far cheaper to manufacture.

GTX285 > 4890

10% or so with a much bigger die? This is not a "win", but your belief that it is, is why AMD is going big this time.

GTX480 > 5870

So that was what, 10-15% faster, 6 months later and with near double power draw?

GTX580 > 6970

Another 10-15%, still with a huge die vs AMD's cancelled 32nm chip.

GTX780 > 7970GE.

20% faster, 11 months later, 65% more transistors and 59% more area is a win for you? Are you even serious?

In each generation, NV is always on top 15% at least.

They always have dies that are much bigger, that's the only reason. That was because AMD's strategy was for a smaller die. There is no other reason why. There is no reason to believe that AMD won't beat Titan this time around with a bigger die.

GTX680 vs. 7970 doesn't really count since it's more of a flagship delay by NV than NV's inability to compete.

If Titan is a flagship delay, Hawaii is a flagship delay.

In fact, with 780, NV crushed AMD by an embarrassing amount.

Right, almost a year later and with ~60% more resources, winning by only 20%. That's embarrassing for sure.

I'll give them the benefit of the doubt and see what they can come up with 9970 but I bet 780 OC will beat 9970 OC by 15% still.

I will take that bet.

I am not saying AMD will not increase die size, but I am not sure I believe they'll make a chip with 470-480mm2 right now.

They won't, it'll be over 500mm2.

Nope, I never said that at all in relation to Tahiti. I am saying that while they will improve performance/watt and perf/mm2 over Tahiti, I don't think 9970 OC will match 780 OC. My view is they'll go bigger chip but not big enough that they'll be able to get away with only 975mhz clocks. I believe their clocks will be higher which will leave a lot less room for overclocking. I am betting that they will have very aggressive Turbo similar to what NV now has with Kepler. With NV since they made such a large 561mm2 chip, they could afford to clock it in mid-800s/low-900s. I do not believe AMD will have this die size luxury. As a result, they'll have to settle between a balance of die size and clock speed.

There are 3 desktop Hawaii SKU's, XT, Pro and LE. What is the reason for that if it's not a (really) big die? How many months did we have to wait for Tahiti LE, and even then it was given a 7870 tag.

For 5 AMD generations we got hype - 2900XT, 3870, 4870, 5870, 6970. None of those chips lived up to the hype, always beaten by NV.

What hype did you see about the 3870 and 4870? As for the 5870, it crushed Fermi almost single-handedly. To say it didn't live up to the hype is totally absurd, what planet are you living on? It is probably THE great card of the last 5 years.

Yes the 6970 was disappointing but it was only half the card it was supposed to be because of TSMC.

Why would I bet that this generation suddenly AMD will beat NV's best card if they haven't done it for 5 generations in a row?

Why would you bet on Nvidia releasing their midrange card first on 28nm? Simply put, strategies change by necessity. Nvidia got lucky because AMD got lazy with Tahiti. Do you really believe they'll make that same mistake twice?

6970 was 389mm2, but 7970 is only 365mm2. They haven't been moving up.

You're just not listening. AMD knows that big early chips at TSMC are disasters waiting to happen. Even Nvidia realised that eventually. Now though, there is no reason at all to not go with a very big chip on 28nm. It will be over a year before TSMC's 20nm is even in decent shape (2 years from now we might see a huge chip), Nvidia will go with a midrange 20nm chip first anyway and it won't be able to beat a large 28nm gpu.

 

[SiliconWars]

Also just imagine the power draw of Tahiti scaled up to 550mm^2. Already at the 365ish mm^2 it is now, power draw per mm^2 compared to GK110 is absolutely awful. And now I see you said that in your post with your ninja edit. Hahaha.

Increasing die size normally doesn't increase power by anything like the amount you think it would. The reason for that is simply, higher clocks with less shaders use far more power than lower clocks with more shaders.

There is no magic here. The 7970GE has power characteristics based on being a card that is pushed the limit. Fermi (the 480) was another, the difference with Fermi is that Nvidia had to push a huge chip to the limit to win and that's why it was the joke of the industry on temps and power draw.

 

[3DVagabond]

I am talking about AMD's best foot forward vs. NV's on the same node. GTX280 > 4870, GTX285 > 4890, GTX480 > 5870, GTX580 > 6970, GTX780 > 7970GE. In each generation, NV is always on top 15% at least. GTX680 vs. 7970 doesn't really count since it's more of a flagship delay by NV than NV's inability to compete. In fact, with 780, NV crushed AMD by an embarrassing amount.

When it takes a company 6 months to a year longer to release a faster chip, that's not winning on a node, IMO. It could be claimed that AMD won by being first to market. Whoever is winning at the time someone is ready to buy is what's important.

The most interesting part for me is how price doesn't matter anymore. Most 780's are over $650 right now and people don't care. That's not progress when almost 2 years later prices are still going up on a node. Add to that you still can't buy a Geforce that isn't made with a salvaged GK110 and the price really seems out of whack to me.

 

[el etro]

now with Hawaii 28nm they are going to be even more bigger than ever before. most probably their first 450+ sq mm chip. mark your statement down. lets revisit in Oct.

That's all that we are really waiting for. AMD going big to fight with Nvidia's big chips. Forget the 20nm advantage for now.

I bet ~Pitcairn x2 here(~450mm² on 28nm and 10% perf over titan, maybe matching and eventual full enabled GK110).

 

[ruhtraeel]

IIRC, the 4870 and 5870 were incredible cards, much better than their NV counterparts. 6970, not so much. 7970 is decent.

 

[jpiniero]

That's not progress when almost 2 years later prices are still going up on a node.

It sure looks like prices are going up anyway, esp if nVidia's claim that 20 nm isn't any cheaper per transistor ends up being true.

 

[spat55]

I think it will be just a refresh, like the GTX 7xx series was for Nvidia, we won't be seeing the 20nm till at least mid 2014, maybe even late 2014. That is why I got another HD 7850 for £100 and crossfired.

 

[StrangerGuy]

It sure looks like prices are going up anyway, esp if nVidia's claim that 20 nm isn't any cheaper per transistor ends up being true.

If anything I expect 20nm to be more long in the tooth than 28nm.

 

[champion-7891]

My opinion on this matter is that since new nodes cost more and that cost is increasing, AMD may have decided to take some 28nm parts into the next (2014) generation Radeon xxxx. If per wafer, taking early yields for 20nm into account, AMD can have a larger margin on a 450-500mm 28nm die as compared to lets see, a 350-380mm 20nm die, even taking into account the greater number of die's per the 20nm wafer, I think it would make more sense financially to follow this strategy.

Lets see, a 2560-3072 SP GPU, 420-500mm die on 28nm, 99xx series, will continue on as a rebadge into the next (2014) x8xx series (maybe with slightly higher clocks, DDR6 VRAM?), while a smaller, 350-380mm die on 20nm with 4096SP/256TMU etc will take over as the flagship x9xx series. If they can have about the same margins on both, why not have a mixed 2014 series and use your massive die longer. It would also give AMD time to prepare a truly massive 450-500mm die on the mature/FinFET node with a brand new architecture for 2015. Just my 2 cents

 

[wlee15]

HD4860 and HD4730 launched after the HD4770. They simply couldnt produce the HD4770 and had to replace it pricewise with a 55nm GPU.

Both of them are derived from the existing RV790 and RV770 generation, and if AMD really wanted to replace the HD 4770 with those chips they would have just have called them the HD 4770.

 

[SolMiester]

The 9970 will be a 28nm X2 7850 with hardware frame pacing...to take on 780, I dont think AMD is doing a big die.....

 

[lavaheadache]

The 9970 will be a 28nm X2 7850 with hardware frame pacing...to take on 780, I dont think AMD is doing a big die.....

I would be irate if this is actually what happens