wccftechAMD Pirate Islands : R9 300 Series Alleged Specifications Detailed

[RussianSensation]

yeah we are. RS has already decided the winner of the next gen GPUs. its the GTX 880. hurray to the new king. :biggrin:

I didn't say GTX880 would beat R9 300 series, but Maxwell's 20nm 500mm2+ flagship should beat AMD's R9 300 series 20nm flagship. Maxwell is a brand new architecture that NV has worked on for a long time. GCN 2.0 should only be an improvement from 1.0/1.1. Not only that but NV has already beaten AMD with 8800GTX/285/480/580/780Ti. Obviously I am not talking about price/performance. Maxwell already shows a major improvement in performance/watt on the same 28nm node.

Wait, so in his fairy tale world only nv is capable of improving their uarchs? Who would have guessed...

I didn't say AMD won't improve GCN 2.0 from 1.0/1.1. I am saying that NV already shows a real world improvement of 35% from Kepler with 750Ti and that's only on 28nm. When AMD moved from 4870 to 5870 to 6970, there wasn't much improvement in IPC. Mostly rasterization/geometry was improved while the rest came from more TMUs, SPs, ROPs and memory bandwidth. GCN 1.1 also has no real improvement in IPC I can see from 7970 series, just more units. If you look at R9 290X specs, it scales more or less in line with the increase in SPs, TMUs. 660 is just 20% ahead of 750Ti despite having specs that exceed it by 50-66% across the board! Maxwell has shown a significant increase in IPC because 750Ti performs far above its paper specs as compared to Kepler.

I think if AMD wants to compete with Maxwell, it needs to go bigger yet again (beyond 440mm2) or has to implement some new tricks like stackable DRAM.

 

[realibrad]

The only thing forcing Nvidia to keep on their toes is AMD, so it is good for everyone if AMD is strong - however not too strong, because they also need a strong competitor in Nvidia.

I personally wouldn't mind a third entry to the dGPU market. Intel is pushing ahead in the iGPU market and it'd be pretty awesome if they want for dGPU market even if I think it is beyond unlikely, it'd be pretty cool.

I was with you for the first part, but then you lost me when you said you dont want 1 company to be too strong. I think you have the cart before the horse. AMD is not weak and therefore it has a weaker product, it has a weaker product and there for its weaker to Nvidia. Competition does not happen because companies are equal, it happens when companies try and out do each other with something better. You want AMD and Nvidia making the best thing they can make, and right now Nvidia does not have to. What you really want is the best product you can get, and making 1 company weaker would not give you that.

 

[blackened23]

The 680 was $700, and the best thing in their arsenal for months. How is that midrange?



A midrange card by definition sits in the middle of a range of prices. How exactly did the 680 sit in the middle of anything?

The heck are you talking about?

The GTX 680 launched at 499$ 2 months after the 7970 which launched at 550$.

The 680 was faster, more efficient, and cheaper. I really have NO IDEA what you're talking about here. GTX 680 was also the single GPU crown winner, again, go read some 680 launch reviews. I have no idea where you're getting this 700$ figure from.

This will help:

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review

titled: "NVIDIA GeForce GTX 680 Review: Retaking The Performance Crown"

It also has a price chart which you may find useful.

 

[Gloomy]

in what universe does a 561mm die beating a 438mm die by less than the size difference count as an architectural win

serious question

 

[rtsurfer]

1) 780Ti OC is 20% faster than R9 290X OC. (HardOCP) Based on how much more dense the transistors are inside AMD chips, they produce a lot more heat which is why AMD's 438mm2 chip consumes similar amounts of power to NV's 561mm2. Therefore, if AMD makes a chip larger than 438mm2 and clocks it high, it should have less overclocking headroom than NV's.

Explain this.

If a AMD makes a large chip, the density will go down, so it should clock higher.

Unless AMD makes a larger & more denser chip than Hawaii, which is unlikely.

 

[RussianSensation]

performed better. yeah. clocked better. not really. HD 4890 clocked at 850 mhz stock. sapphire toxic was 960 mhz stock and easily hit 1 Ghz. I don't remember any GTX 285 hitting 1 Ghz.

I didn't mean clocked better, but overclocked better (i.e., in percentage terms). For instance, both the 480 & 580 could often do 20% on air, while 5870/6970 very rarely achieved that. 780/780Ti have far higher overclocking room % wise vs. 290/290X.

Also the HD 7970 was as good if not a better overclocker than GTX 680. Tahiti was known for its overclocking potential with watercooling. yeah it did require more power than GTX 680.

I didn't mention that match up, partly because 680 was not a large monolith flagship from NV. What I predict is AMD's flagship 20nm won't overclock as well as NV's big-daddy Maxwell on 20nm in % terms.

1. We don't know what GCN 2.0 brings in terms of architectural improvements
2. AMD looks likely to bring HBM to their next gen GPUs. They would be a generation ahead just as they were with GDDR5 on HD 4870.
3. We are already seeing Mantle show better frametimes in BF4 due to it being a better API than DX.

1. True but I don't believe it can match Maxwell's 35%. Why? Because R9 290X has no material improvement in IPC from 7970 despite being released almost 2 years after. If AMD couldn't even make a dent in IPC in 2 years, what makes me believe they will suddenly produce a 35% increase? Not only is Maxwell more efficient than Kepler on 28nm node, but it does so with far less memory bandwidth. That's amazing!

2. If true, that would be a wildcard.

3. With DX12, I think Mantle will matter less. The key for AMD is to team up with major developers to optimize games for their GPUs/architectures. Games like Witcher 3, the Division, Dragon Age: Inquisition, etc. I think AMD also needs to improve its performance in Blizzard games and FTP. I know those are not very GPU demanding but the perception in those titles carries to other games for the average Joe. My friends who play World of Warcraft and Lost Planet 2 see that NV is faster and continue to think NV is faster "in almost all popular games". It's just the perception.

Where I think AMD might have a chance is 4K and multi-monitor as their cards have a lot of ROPs.

 

[Rvenger]

If I see anymore bickering I will be closing this thread and handing out warning/infractions. Please stop NOW.

-Moderator Rvenger

 

[RussianSensation]

in what universe does a 561mm die beating a 438mm die by less than the size difference count as an architectural win

serious question

You missed the entire point of the comparison. No one is disputing that GCN 1.1 R9 290X is more efficient per mm2 than Kepler 561mm2 from a performance point of view. Now imagine a hypothetical 530-560mm2 Maxwell chip that has 35% more performance from IPC alone. How is GCN 2.0 going to compete with that?

Explain this.

If a AMD makes a large chip, the density will go down, so it should clock higher.

Unless AMD makes a larger & more denser chip than Hawaii, which is unlikely.

They went from 354mm2 (365?) with Tahiti XT to 438mm2 with Hawaii and overclocking is worse. 7970 is the exception in AMD's line. 3870/4870/4890/6970 and R9 290X are all poorer overclockers compared to NV. Even now R9 290X trades blows with 780Ti but once both are overclocked, the 780Ti pulls away by a 15-20% since 780Ti overclocks better in % terms from reference/base clock.

Before, AMD played price/performance or bitcoin mining or scrypt mining card or unlocking card. Right now the price difference between R9 290/290X and 780/780Ti is not enough to go with AMD considering NV overclocks much better. That's why AMD is going to keep losing market share until they either reclaim the performance crown, add more game bundles, beat NV to launch with 20nm and/or undercut them by a larger %. by the time Maxwell rolls, scrypt and bitcoin mining will be dead which will leave AMD with no advantages at all that 4000/5000/6000/7000/R9 200 series had. I guess we'll have to revisit if my prediction was correct but based on previous 4-5 generations, NV's flagship overclocked better on air than AMD's, excluding 7970 vs. 680.

Maybe AMD will put all out all the stops and release a $700 CLLC R9 390X with uber high clocks and a bucketload of shaders. That would be quite something!

 

[sushiwarrior]

.

 

[caswow]

actually this thread turned into a "amd is too dumb to match maxwell we dont even give them a chance".

 

[RussianSensation]

By different memory config I don't mean HBM necessarily. I said BIG die. 550mm^2 doesn't hold a candle to it.

If AMD builds a 500-550mm2 big die which neither ATI nor AMD would dare to manufacture, without Quadro and Tesla to amortize the manufacturing costs, how are they going to be able to sell such chips at 'reasonable' prices? I guess AMD will move up to $699-749 price bracket for R9 390X? Perhaps since AMD sees that people are buying 780Ti for $699, then why not target the same price level?

Why do I get a feeling that between all the NV vs. AMD wars, both of those companies just have a goal of making as much $ as possible from us? Take a look at this

4870X2 = $549
5970 = $599
6990 = $699
7990 = $999
295X2 = $1499

What's wrong with this picture? The price of AMD's dual-GPU flagship tripled since August 2008. AMD used to be about price/performance but for them to release a $699+ GPU, they better have the performance crown. :hmm:

actually this thread turned into a "amd is too dumb to match maxwell we dont even give them a chance".

Ummm...no. We are just discussing the possible match up of Maxwell vs. GCN 2.0. GCN was the newer architecture and it still couldn't beat the best Kepler chip GK110 780Ti. With Maxwell, NV is bringing 35% IPC to start with and 2x the performance/watt on 28nm. How is GCN 2.0 going to be able to overcome both of those points when Maxwell jumps to 20nm and maintains 500mm2 large die? Only sushiwarrior hinted that AMD will aim to increase the die size even further. Everyone else is downplaying my comments.

Imo, on paper AMD is massively disadvantaged for the next round. GCN was the new kid on the block when Kepler was just Fermi 2.0. Now, Maxwell is the new kid and 750Ti is kicking major ass if we ignore its price, and it's just on 28nm.

Another way to look at it, if Maxwell is 35% more efficient, then a 365-380mm2 28nm Maxwell chip should at least equal GTX780Ti. With a move to 20nm and the same 35%, a 20nm 550mm2 Maxwell should be a monster!

 

[Gloomy]

By different memory config I don't mean HBM necessarily. I said BIG die. 550mm^2 doesn't hold a candle to it.

2,000 dollar GPU confirmed D:

 

[DiogoDX]

By different memory config I don't mean HBM necessarily. I said BIG die. 550mm^2 doesn't hold a candle to it.

But the cost would be prohibitive since 20nm will be more expensive. Maybe Glofo 20nm FD-SOI?

But I have my doubts if AMD will move a big die to Glofo new process. Would be a bold and risc move.

 

[sushiwarrior]

If AMD builds a 500-550mm2 big die which neither ATI nor AMD would dare to manufacture, without Quadro and Tesla to amortize the manufacturing costs, how are they going to be able to sell such chips at 'reasonable' prices? I guess AMD will move up to $699-749 price bracket for R9 390X? Perhaps since AMD sees that people are buying 780Ti for $699, then why not target the same price level?

.

 

[SlowSpyder]

It seems like you are saying that the increase in IPC is due to the node shrinking and thus being able to add more to the same size chip. If Nvidia gets an increase of 35% because the number of transistors goes up per size unit, why would AMD not also get the same increase when they shrink down to 20nm?

Both AMD and Nvidia should get equal benefits of shrinking in terms of adding more transistors right?

Check out the review from Anand on the Sapphire Radeon R9 290 Tri-X OC. Look at the difference in clock rate, fps and temp. You will find that when you look at the cards you see that clock per clock the fps are within 10% or less diff. Then you look at the power draw not a huge diff. The efficiency difference is not all that great when you properly cool a 290. The main diff is that Maxwell can OC better because the larger chip means lower heat density. That comes at the cost of being more expensive to produce.

I don't think it is quite that simple. I wouldn't try and draw too many conclusions on what to expect out of 20nm Maxwell based off of what the current Maxwell GPU does on a very mature 28nm node.

Look at the transistor density on both AMD and Nvidia's GPU's in the midrange / performance parts vs. their flagship GPU's. Density isn't the same across the board with just more or less silicon. And without knowing what AMD can do with GCN 2.0 or what 20nm Maxwell would do, I think it is a bit premature to try and draw any real hard conclusions, that is all I am saying.

AMD and Nvidia's top end parts also have different thermal limits. We have no idea what those limits will be on 20nm. What if AMD's GPU doesn't throttle until 100C but Nvidia's GPU throttles at 88C?

I'm not saying what I think will happen, because I really don't know... just that with both companies bringing new or updated architectures to a brand new manufacturing process, I wouldn't try and draw too many conclusions based off of what is available today.

*edit - Sadly, I think the only thing we can count on for sure with the 20nm parts is that AMD and Nvidia both know that we'll pay way more than $500 for a top end card, and I would expect next gen pricing to reflect that.

 

[Skurge]

If AMD builds a 500-550mm2 big die which neither ATI nor AMD would dare to manufacture, without Quadro and Tesla to amortize the manufacturing costs, how are they going to be able to sell such chips at 'reasonable' prices? I guess AMD will move up to $699-749 price bracket for R9 390X? Perhaps since AMD sees that people are buying 780Ti for $699, then why not target the same price level?

Why do I get a feeling that between all the NV vs. AMD wars, both of those companies just have a goal of making as much $ as possible from us? Take a look at this

4870X2 = $549
5970 = $599
6990 = $699
7990 = $999
295X2 = $1499

What's wrong with this picture? The price of AMD's dual-GPU flagship tripled since August 2008. AMD used to be about price/performance but for them to release a $699+ GPU, they better have the performance crown. :hmm:



Ummm...no. We are just discussing the possible match up of Maxwell vs. GCN 2.0. GCN was the newer architecture and it still couldn't beat the best Kepler chip GK110 780Ti. With Maxwell, NV is bringing 35% IPC to start with and 2x the performance/watt on 28nm. How is GCN 2.0 going to be able to overcome both of those points when Maxwell jumps to 20nm and maintains 500mm2 large die? Only sushiwarrior hinted that AMD will aim to increase the die size even further. Everyone else is downplaying my comments.

Imo, on paper AMD is massively disadvantaged for the next round. GCN was the new kid on the block when Kepler was just Fermi 2.0. Now, Maxwell is the new kid and 750Ti is kicking major ass if we ignore its price, and it's just on 28nm.

Another way to look at it, if Maxwell is 35% more efficient, then a 365-380mm2 28nm Maxwell chip should at least equal GTX780Ti. With a move to 20nm and the same 35%, a 20nm 550mm2 Maxwell should be a monster!

I would believe sushiwarrior. Rember this post from October?

http://forums.anandtech.com/showpost.php?p=35572519&postcount=171

 

[Gloomy]

Hold the phone-- if this is legit info, this means the chip is more than a twinkle in AMD's eye.

So AMD is leading with a big die?

 

[krumme]

Fiji will be something new and different, with respects to memory config. Big die.

Damn !!!

 

[wand3r3r]

Fiji will be something new and different, with respects to memory config. Big die.

By different memory config I don't mean HBM necessarily. I said BIG die. 550mm^2 doesn't hold a candle to it.

What I'm trying to decipher from this is if there is a 390x in the works which is considerably larger then 550mm^2, or if the 390x will be so powerful that it is similar to a larger die.

I would believe sushiwarrior. Rember this post from October?

http://forums.anandtech.com/showpost.php?p=35572519&postcount=171

Cat 13.6 doesn't support hawaii, fake, only cat 13.20 and above

Vesuvius, it exists.

Stock watercooling, Asetek design, I have used the card myself.

Oh, I didn't catch that one! I've seen several very very "prophetic" er... hints from him/her before which have been accurate. Crazy, that was last October!

 

[Kippa]

Just out of curiosity as a general guess, will the next gen 20nm cards be beefy enough to support 4K gaming with the details in the game dialled high? If not I can see myself skipping the first batch of 20nm cards and waiting for a more powerful and mature card that can support 4K with details high.

 

[parvadomus]

True but I don't believe it can match Maxwell's 35%. Why? Because R9 290X has no material improvement in IPC from 7970 despite being released almost 2 years after. If AMD couldn't even make a dent in IPC in 2 years, what makes me believe they will suddenly produce a 35% increase?

Ofc Hawaii has no material IPC improvement. Its based on the same building blocks as Tahiti, just rearranged for better. This is just like what AMD does with their CPUs and every other die. They reuse the same architecture for different purposes, just like the stars cores went into the first APU, or when they got turbo, etc. This time AMD only added turbo, XDMA, trueaudio..
Im 100% sure the X300 series will be a real architecture change.

 

[VulgarDisplay]

Again, 500-550mm^2 isn't even in the same ballpark AFAIK. It's not that big because they're cramming shaders in though.

I'm guessing some type of high speed cache built into their next gen gpu's which will also apply to their apu's moving forward.

Hot or cold?

 

[VulgarDisplay]

They went from 354mm2 (365?) with Tahiti XT to 438mm2 with Hawaii and overclocking is worse. 7970 is the exception in AMD's line. 3870/4870/4890/6970 and R9 290X are all poorer overclockers compared to NV. Even now R9 290X trades blows with 780Ti but once both are overclocked, the 780Ti pulls away by a 15-20% since 780Ti overclocks better in % terms from reference/base clock.

There is literally no difference between Hawaii and Tahiti core over clocking. The only difference is ram oc'ing.

Both Tahiti and Hawaii are basically guaranteed 1075-1125mhz on the core with good chips allowing 1125-1225mhz and the truly golden chips getting around 1300mhz. The different bus widths make comparing ram oc between them irrelevant.

Hawaii on the stock cooler was an atrocity and gave people the perception that the cards cannot over clock which is false. There is also this assumption that all Tahiti cards were able to overclock to 1350mhz which has given the impression that Hawaii is worse. I would wager the majority of Tahiti GPUs were like mine and could not pass 1150mhz no matter what voltage I added.

 

[Silverforce11]

@RS, you cannot be serious with comparisons of potential max OC as a major factor, look at the 7950, 45-50% OC were possible. Doesn't mean jack, because AMD set the default clocks way too low to position their stack in perf/$ appropriately.

Also, that assumption that the 780ti is 20% faster OC is based on a poor review from [H] with a bad R290X sample that overheated requiring max fan speed (something they specifically mentioned) from the get go when other sites testing the same card did a bigger OC with 50% fan (Guru3d, TPU, HC etc).

Looking at the current reviews with updated drivers from both sides, I've already linked to you Computerbase.de that finds R290X essentially matching 780ti that is allowed 100% fanspeed and power limit increased. It would mean a stock R290X against an OC 780ti in a sense because the R290X clocks don't raise above their default, it just throttles less when its power limit is raised. That's from their large suite of games, but not including BF4 with Mantle. As I've said, if they had included the Mantle result, the R290X would have beaten the 780ti.

So a smaller die is matching or beating a bigger die, and in CF vs SLI, beats it outright at 4K by large margins. That's bad?? Show poor signs of things to come, how is that even a proper interpretation?

There is a reason AMD labeled Hawaii as GCN 1.1 because its a slight change from Tahiti as GCN 1.0, if anything, VI being GCN 2.0 should hint to the observer that it will be a major change.

 

[TreVader]

The heck are you talking about?

The GTX 680 launched at 499$ 2 months after the 7970 which launched at 550$.

The 680 was faster, more efficient, and cheaper. I really have NO IDEA what you're talking about here. GTX 680 was also the single GPU crown winner, again, go read some 680 launch reviews. I have no idea where you're getting this 700$ figure from.

This will help:

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review

titled: "NVIDIA GeForce GTX 680 Review: Retaking The Performance Crown"

It also has a price chart which you may find useful.

I was thinking about the 780 regarding price. More to the point, $500 is midrange now?



Also, the 7970 beats the 680 consistently. The 770 is about equal to the 7970.


Again, please explain how the 680 was midrange when there was only one other card around more expensive, and none from nvidia?