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

[RussianSensation]

No, but Tahiti has lots of redundancies that at this point can be stripped because of the maturity of 28nm and you can still get good yields without them (as opposed to where tahiti launched when 28nm was still in it's early stages).

AMD had more than a year to increase performance on the same node from 5870 to 6970. In that time we got:

1) Die size increase from 334mm2 to 389mm2 (performance increased 14-16% at 1200p/1600p).* :thumbsdown:

* AMD claims a 10% improvement in performance per square millimeter over the prior VLIW5 design.

2) Real world pixel performance improved 16% and texture fill-rate 15%.

3) Most of the increase came from upgraded geometry engines. In Unigine Heaven extreme tessellation, 6970 scored 58% more.

4) Power consumption increased at least 28-30W.

I think AMD can without much problem increase performance 17-22% to match a stock 780 but who in their right mind would buy a stock 780 now? I can't imagine someone who spends $550+ on a videocard will not do research and see that $650-660 after-market 780s are boosting to 1.1Ghz out of the box and have more overclocking headroom. And really NV can lower pricing $50 and completely neutralize the 9970. The 9970 has to be faster than a reference 780 or have a ton of overclocking headroom. Otherwise NV drops the price $50 and 9970 is irrelevant again. And for AMD to increase performance a lot more, they have to increase ROPs which makes the chip larger.

 

[PPB]

And yet both of those were on a 40nm mature process. 4770 having done the pipecleaning job, 5870 was already pushed to it's limits, only so a modification in the vliw composition could give you a perf increase (pointed mostly at crappy tess perf).

No we have a silicon that wasn't touched since it first came out (7970) which was when the architecture debuted on a new process. There were hints that tahiti was overprovisioned since it launched, one of it being it's bigass memory bus, something rather odd considering ATI/AMD's philosophy for GPU designs.

At a new node, you either go a small die for pipecleaning and leaving big dies for later, or you go mid-big dies with redundancies and overprovisioning to make up for bad yields, or you go with absolute big dies and deal with horribly bad yields.

Lastly, you can just design a die with no overprovisioning and scarse redundancies, but you will have to wait up till the new process gives you acceptable yields. That was the case with GK104. Instead of going really wide with the bus, paying a die size penalty but having your product launch at the same time as your competition.

That waiting gave NV the time to adjust stock clocks so their products across most price tiers would be just a tad faster than AMD's 7xxx series. It was really well played, but still it would be naive to think some portions of not-GK110 kepler were already pushed to the limit (namely, its mem bus and the need to hit mad clocks not no be BW starved).

 

[ShintaiDK]

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.

 

[SiliconWars]

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.

What? The 4770 was a great little card and it most certainly was a pipecleaner. What's more, it let AMD know what the issues with 40nm were (via's), something that they used to their huge advantage with Cypress while Nvidia toiled with Fermi (Nvidia didn't know about the via issue).

 

[raghu78]

I can't imagine someone who spends $550+ on a videocard will not do research and see that $650-660 after-market 780s are boosting to 1.1Ghz out of the box and have more overclocking headroom. And really NV can lower pricing $50 and completely neutralize the 9970. The 9970 has to be faster than a reference 780 or have a ton of overclocking headroom. Otherwise NV drops the price $50 and 9970 is irrelevant again. And for AMD to increase performance a lot more, they have to increase ROPs which makes the chip larger.

RS let me say this you are plainly exaggerating GTX 780. you were bashing the GTX 780 a few weeks back. and here you are now believing its now a phenomenal piece of hardware. first of all not every GTX 780 clocks 1.3 ghz. if you believe that you are plain unrealistic. the average is 1150 - 1200 mhz. you believe every HOF GTX 780 card will clock 1.3 ghz like the cherry picked review sample hardocp got. moreover Galaxy is a sponsor of hardocp. so no surprises there.

stock HD 7970 Ghz - 1x
HD 7970(1.3 Ghz) - 1.2x
stock GTX 780 - 1.2x
stock GTX Titan - 1.33x
GTX 780(1.3 ghz) - 1.6x

if a HD 9970 at stock speeds of 975 mhz beats a titan and get close to 1.4x of stock HD 7970 Ghz, how the heck is a GTX 780 going to make any difference at all. roughly 300 mhz of OC headroom and a 25% perf gain. 1.4 x 1.25 = 1.75x. also when you overclock perf increases are not linear as the extra power and voltage allow better than linear scaling because at stock the card was being throttled. so a HD 9970 could easily gain 0.4x , same as GTX 780, thus hitting 1.8x.

so the GTX 780 does not do anything at all to change things. also the GTX 780 requires roughly 150 mhz extra clocks to match Titan. so GTX Titan at 1.25 Ghz needs 1.4 Ghz GTX 780 to match. so I don't understand this hype around GTX 780.

and why do you think 1.25 - 1.3 Ghz is not achievable on a HD 9970 with a powerful PCB and voltage control. Titan has been shown to hit 1.25 ghz with custom BIOS, wider voltage control and increased power target. its not a miracle that the best samples of all these cards can clock in the 1.25 - 1.3 ghz as they are all made at TSMC 28nm.

 

[VulgarDisplay]

I honestly don't think we will see any 20 nm gpu's until this time next year and that amd is releasing a refresh on 28nm because of this.

 

[Zanovar]

I think RS is just bored,wants a new card.

 

[raghu78]

I honestly don't think we will see any 20 nm gpu's until this time next year and that amd is releasing a refresh on 28nm because of this.

yeah. I am quite sure AMD made the decision to go 28nm after a lot of careful analysis of available information. the earliest 20nm cards will hit retail in June or July 2014. but 20nm supply won't be significant until late Q3 2014. AMD will have their 20nm Pirate Islands flagship ready by late Q3 2014 to catch the back to school and holiday 2014 sales. hawaii takes care of 2013 holiday sales. Nvidia's GPU division's record sales and margins are going to come under pressure over the next 12 - 15 months.

 

[Fastx]

1. 20nm node is not a few months away. Why this myth keeps being perpetuated on our forums

I could list more but here you go imo.
(just having some fun with you in sort of a timeline and yes I know some of the below is circumstantial, but enough circumstantial info could possible mean something?

Last year, a report revealed that AMD’s Volcanic Islands series would be the first to make use of a 20nm process and would arrive sometime in 2014. However, in a recent interview with AMD’s Jim Keller and Chekib Akrout, Rage3D has revealed that the Volcanic Islands GPUs would launch in 2013.

The next generation of graphics cards – Volcanic Islands – is coming this year and shaping up nicely. When you name products after places, it leads to interesting thoughts about where to hold events surrounding that namesake product. Typically places that are cheap to get travel to internationally for a worldwide congregation are preferred, so I’m off to renew my passport on the off chance I’m headed to Reykjavik later this summer. But I sure as hell wouldn’t complain about being sent to Honolulu, either.” Rage3D This could be huge news if true for AMD fans and could possibly provide a hint at the late launch of the much anticipated Radeon HD 8000 series GPUs. It was hinted by various techsites that the launch wasn’t planned to be late but it was delayed instead. The new VI (Volcanic Islands) GPU’s would be manufactured on a 20nm Gate-Last process through either TSMC or Common Platform Alliance.

4-13

While TSMC has four "flavors" of its 28nm process, there is one 20nm process, 20SoC. "20nm planar HKMG [high-k metal gate] technology has already passed risk production with a very high yield and we are preparing for a very steep ramp in two GIGAFABsTM," Sun said.

8-13

“On 20nm [process technology], we see little competition. The risk production has started in the first quarter [2013] and the volume production will start in early 2014 next year […] the equipment [is] already being installed, the equipment [is] streaming in and [is] being installed. […] The volume production will start in early 2014,” said Morris Chang, the head of TSMC.

Update 8/7/13 @ 4pm: AMD has clarified that it is the Hawaii/Volcanic Islands tech day in September, not the launch. The launch will officially happen in Q4 some time, the dates in this article are our speculation.

Interesting?^

I hope they do get 20nm out this year even if it is a low volume(risk production) release till next year, it would not be the first time

Rage3D veteran poster. ^ fwiw

Radeon HD 9900, 9800 and 9700 could appear in one or two month’s time. This makes sense, because NVIDIA Maxwell architecture is coming sooner than expected (can’t share any details yet),

Kind of interesting on Maxwell depending on how one looks at this if this is going to be 28nm or will it turn out to be 20nm?

I'm still confused at 20nm... TSMC + release date says absolutely no, but the specs increase + TDP say absolutely yes...

Also kind of interesting comment. ^

 

[CakeMonster]

Interesting? Not really that interesting considering several of those contradict each other. But I guess its fun to speculate and dream about what might come, and that's probably why we're all here in the first place.

 

[bystander36]

It would seem when they talk about yields and successes with the 20nm process, they are just talking about samples, and not production yields.

 

[BallaTheFeared]

^ proof AMD is going 20nm

Uber top secret!

 

[Fastx]

It would seem when they talk about yields and successes with the 20nm process, they are just talking about samples, and not production yields.

Bystander just so you know these passed risk production wafers below can be used ( bought/purchased from TSMC) for retail cards till volume hits.
4-13

While TSMC has four "flavors" of its 28nm process, there is one 20nm process, 20SoC. "20nm planar HKMG [high-k metal gate] technology has already passed risk production with a very high yield and we are preparing for a very steep ramp in two GIGAFABsTM," Sun said.

I am not saying or arguing a 20nm is coming, the above info in post #52 was mainly directed to RS quoted question. But at the same time a circumstantial argument (imo) could be made for a possible FS 20nm depending how people look at the info that is out now. Imo I think a 20nm is still possible at this time for the AMD Flagship with what ever sku is used for the FS. I would also not be that suprised if we see a 20nm should that turn out to be the case imo.

 

[tviceman]

That's not realistic, unless they drop DP. 7970 is a 365mm2 chip. You can't only add 35mm2 and suddenly match 780s unless your chip runs at 1400mhz. I think 430-440mm2 is what they'd need bare minimum. You are also giving AMD a lot more credit than they deserve. You think NV needs to make a chip 40% larger to match AMD's 400mm2 in performance? Do you think NV put placebo transistors inside their GK110?

Well I think we all know that GK110 has transistors dedicated to some of the compute functions that is exclusive to Kepler (dynamic Parallelism, Hyper-Q, Grid Management Unit, etc.) and also to the extra cache the chip has.

I think AMD has room to improve upon Tahiti's efficiency to increase it's performance per watt. And as I already said, I do not see them going much beyond 400mm^2, maybe 410 absolute tops. I think it will end up under 400mm^2.

 

[raghu78]

Well I think we all know that GK110 has transistors dedicated to some of the compute functions that is exclusive to Kepler (dynamic Parallelism, Hyper-Q, Grid Management Unit, etc.) and also to the extra cache the chip has.

I think AMD has room to improve upon Tahiti's efficiency to increase it's performance per watt. And as I already said, I do not see them going much beyond 400mm^2, maybe 410 absolute tops. I think it will end up under 400mm^2.

no. 400 - 410 sq mm is not enough to increase performance significantly. Given the 21 month gap and a mature TSMC 28nm node I am quite sure AMD has gone for a chip which delivers 35 - 40% more perf. releasing a chip 5 - 8 months after the competition (GK110) and not being able to match it would be a huge failure. die size of 450 sq mm - 480 sq mm. thats more realistic.

 

[tviceman]

no. 400 - 410 sq mm is not enough to increase performance significantly. Given the 21 month gap and a mature TSMC 28nm node I am quite sure AMD has gone for a chip which delivers 35 - 40% more perf. releasing a chip 5 - 8 months after the competition (GK110) and not being able to match it would be a huge failure. die size of 450 sq mm - 480 sq mm. thats more realistic.

Your head is in the clouds man. You are completely dreaming. AMD is not going to release a chip that big.

 

[3DVagabond]

no. 400 - 410 sq mm is not enough to increase performance significantly. Given the 21 month gap and a mature TSMC 28nm node I am quite sure AMD has gone for a chip which delivers 35 - 40% more perf. releasing a chip 5 - 8 months after the competition (GK110) and not being able to match it would be a huge failure. die size of 450 sq mm - 480 sq mm. thats more realistic.

I agree that if Hawaii wasn't going to be able to compete with GK110 releasing it now would be a dumb move.

 

[RussianSensation]

I agree that if Hawaii wasn't going to be able to compete with GK110 releasing it now would be a dumb move.

There is always another possibility - 15-17% faster than 7970GE for $499 (or come in within 5% of a reference 780 for $150 less). That would save it from directly competing with $650-690 GTX780s, while handily beating $450-450 770 4GB cards. This is kinda what NV did when they dropped 760 completely disrupting the pricing on 7950/7970. Right now there is no card from AMD or NV that is worth buying between $450-$600. I would even say there is no card in the market worth buying between $350 and $650. Right now in US one would have to be the most hardcore NV fanboy to pay $400-460 for 770 cards when 7970 1Ghz/GEs are going for $290-350. AMD can make 770 completely irrelevant between 7970 and 9950/9970 lines without ever directly addressing the 780.

 

[3DVagabond]

There is always another possibility - 15-17% faster than 7970GE for $499 (or come in within 5% of a reference 780 for $150 less). That would save it from directly competing with $650-690 GTX780s, while handily beating $450-450 770 4GB cards. This is kinda what NV did when they dropped 760 completely disrupting the pricing on 7950/7970. Right now there is no card from AMD or NV that is worth buying between $450-$600. I would even say there is no card in the market worth buying between $350 and $650. Right now in US one would have to be the most hardcore NV fanboy to pay $400-460 for 770 cards when 7970 1Ghz/GEs are going for $290-350. AMD can make 770 completely irrelevant between 7970 and 9950/9970 lines without ever directly addressing the 780.

Why wait this long to do that though? I would have released that before the 780 dropped. Not now, and certainly without all of the fanfare and hoopla.

 

[raghu78]

There is always another possibility - 15-17% faster than 7970GE for $499 (or come in within 5% of a reference 780 for $150 less). That would save it from directly competing with $650-690 GTX780s, while handily beating $450-450 770 4GB cards. This is kinda what NV did when they dropped 760 completely disrupting the pricing on 7950/7970. Right now there is no card from AMD or NV that is worth buying between $450-$600. I would even say there is no card in the market worth buying between $350 and $650. Right now in US one would have to be the most hardcore NV fanboy to pay $400-460 for 770 cards when 7970 1Ghz/GEs are going for $290-350. AMD can make 770 completely irrelevant between 7970 and 9950/9970 lines without ever directly addressing the 780.

no. being slower than GTX 780 is a damn stupid release. GTX 780 is missing one complete GPC = 3 SMX. Nvidia is now capable of releasing a 2688 sp or even 2880 sp GTX 785. fully enabled GK110s are already shipping in Quadros. so by Q4 its going to be quite easy for Nvidia to push out a GTX 785. what will AMD do then ? Nvidia will push GTX 770 to 350, GTX 780 to 500 and GTX 785 to USD 650. so you see AMD is screwed unless they compete head on with a fully enabled GK110. Hawaii needs to beat GK110 for perf/sq mm and perf/watt for AMD to be successful. the competiton is very well entrenched.

unless Hawaii XT is 35 - 40% faster than HD 7970 Ghz , matching or edging Titan and HD 9950 matches a GTX 780 , Nvidia will have no problem just pushing current cards down in price and releasing newer cards at the top.

 

[SiliconWars]

Your head is in the clouds man. You are completely dreaming. AMD is not going to release a chip that big.

No you're hoping more like. It's clearly a big chip, that's why there are 3 SKU's coming off the Hawaii die from the start.

 

[RussianSensation]

unless Hawaii XT is 35 - 40% faster than HD 7970 Ghz , matching or edging Titan and HD 9950 matches a GTX 780 , Nvidia will have no problem just pushing current cards down in price and releasing newer cards at the top.Hawaii needs to beat GK110 for perf/sq mm and perf/watt for AMD to be successful.

AMD has lost HD2900, 3870, 4870/4890, 5870/6970 and 7970/7970GE generations on the high-end. What makes you think AMD has any chance of competing against a fully unlocked 2880 GK110 aka GTX785? Unless AMD introduces 20nm, I just can't see how a stock 9970 will end up 40% faster than 7970GE. The hype for this card is now building up to 6970 levels. Also, all of these NV competitors are coming in clocks in the high 800s-low 900s range and scale to 1.2-1.3Ghz with overclocks. In my eyes AMD has lost this generation on the high end. They should focus on delivering better game bundles and price/performance until Pirate Islands and invest all their resources because Maxwell should be a beast.

 

[hawtdawg]

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

 

[raghu78]

AMD has lost HD2900, 3870, 4870/4890, 5870/6970 and 7970/7970GE generations on the high-end. What makes you think AMD has any chance of competing against a fully unlocked 2880 GK110 aka GTX785? Unless AMD introduces 20nm, I just can't see how a stock 9970 will end up 40% faster than 7970GE.

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.

Jan 2012 - Mar 2012 HD 7970
Apr 2012 - Jun 2012 GTX 680
July 2012 - Feb 2013 HD 7970 ghz
Mar 2013 - Sep 2013 GK110

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.

you are not able to see AMD gain 35 - 40% because of very rigid views that AMD will not go big die. 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. 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. there are process improvements which allow AMD to reduce redundancies to a much lesser level compared to Tahiti. AMD can pack transistors closer on mature 28nm process and transistor density goes up. its almost a given that Hawaii is not an incremental advance like HD 5870 - > HD 6970.

the time gap between 4870 and 5870 was 15 months. 5870 and 6970 was 15 months. 6970 and 7970 was 13 months. and here for the first time we see a gap of 21 months. I don't think we can expect a meagre 15 - 20% improvement. 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.

 

[SiliconWars]

AMD has lost HD2900, 3870, 4870/4890, 5870/6970 and 7970/7970GE generations on the high-end. What makes you think AMD has any chance of competing against a fully unlocked 2880 GK110 aka GTX785?

Lost? The only one of them that was even remotely close to high-end size was the 2900.

Pretty sure the 7970GE beat the 680 for longer (8 months) than the 7970 "lost" to it (3 months). Let me guess though, you're counting Titan's 6 months of beating the 7970GE right?

Unless AMD introduces 20nm, I just can't see how a stock 9970 will end up 40% faster than 7970GE.

How did Titan end up 50% faster than a 680? Was it magic 20nm pixie dust or is it because it is 87% bigger, with twice the transistor count?