AMD's Tonga - R9 285 (Specs) and R9 285X (Partial Specs)

[Haserath]

From AMD side it seems R9 300 mid-range won't be out for at least 6 months then. It would be an interesting strategy if NV's first wave of Maxwell is 28nm but AMD waits to launch R300 on 20nm since they know they can't deliver 2x perf/watt of Maxwell on 28n. Essentially we would have NV competing on far superior architecture from perf/watt but AMD will compensate with 20nm. This time it's really hard to get any read on 300 series since there have been no leaks at all. The question is why didn't AMD just lower 280/280X prices by $50 and added the same game bundle? For gamers a 3GB 280X for $249 MSRP would have probably been more preferable than a 285X 2GB for $279. Maybe AMD can reposition the 285X as a $279 4GB card.

To get the $ improvements. More density, less memory, no Cfx connector, lower board cost.

They don't like selling their 360mm 3GB cards at $200-300 while Nvidia has a <300mm 2GB card for $300+.

Who knows if AMD even knew about Maxwell's performance characteristics before they started designing this card(how well do these company's gather intel on the other(pun somewhat intended))? They knew the timeframe 20nm was going to be available and they knew about GK104, so they designed this.

 

[Silverforce11]

Hopefully Nvidia doesn't get too cocky and end up raising their prices because of this. Hate to see a $1500 Titan II and a $650-700 GTX 880 lol.

You can hope. But given their recent actions, if there's no competition from AMD, you are going to pay and pay dearly.

 

[jpiniero]

2014
Sorted by performance:
GTX 750 Ti
R9 280
R9 285
R9 280X
R9 285X
R9 290X/GTX 870
GTX 780Ti
GTX 880

Probably more like:

750 Ti (60W)
860 (100W)
285 (190W)
285X (240W)
870 (140W)
290X (290W)
880 (190W)

 

[FatherMurphy]

Doubt there will be 50W difference between GTX 880 and 870. There certainly wasn't that much between 680 and 670.

 

[exar333]

To get the $ improvements. More density, less memory, no Cfx connector, lower board cost.

They don't like selling their 360mm 3GB cards at $200-300 while Nvidia has a <300mm 2GB card for $300+.

Who knows if AMD even knew about Maxwell's performance characteristics before they started designing this card(how well do these company's gather intel on the other(pun somewhat intended))? They knew the timeframe 20nm was going to be available and they knew about GK104, so they designed this.

Pretty much this...

 

[ocre]

That's the problem with over hyping cards. People were actually claiming Tonga was gonn improve performance per watt by 2 and this was amd's response to nvidia and their maxwell.

I am always against out of control prehype.

About the 285, I cannot see how it will be faster than the r9 280. Across an average, I can't see it. The price is strange, there are gtx760 OC models for cheaper now. So I am not sure on the price. But AMD might not have a huge stock and surely they will be able to produce these chips cheaper. Maybe they will try to have some major deals for the holidays, compete on killer prices. I don't know.

 

[3DVagabond]

That's the problem with over hyping cards. People were actually claiming Tonga was gonn improve performance per watt by 2 and this was amd's response to nvidia and their maxwell.

I am always against out of control prehype.

About the 285, I cannot see how it will be faster than the r9 280. Across an average, I can't see it. The price is strange, there are gtx760 OC models for cheaper now. So I am not sure on the price. But AMD might not have a huge stock and surely they will be able to produce these chips cheaper. Maybe they will try to have some major deals for the holidays, compete on killer prices. I don't know.

Who was saying this? Because I can find someone that'll say just about anything if I look hard enough. Was this any kind of official statement or even reasonable industry leak? Or was it some guy on a forum somewhere?

 

[Borealis7]

the 285 could be interesting in an XFire setup for 500$ which could be stronger than a single 500$ card. but then again, you'd have to settle (pun) with Microstutter.

EDIT: 250$ seems a bit much for this card IMO.

 

[boxleitnerb]

I distinctly remember AMD saying in some presentation (about a year ago?) that they found a way to double perf/W each generation. Now I don't know if they specifically clarified what they view(ed) as a new "generation", if that includes node shrinks or not and if it is only applicable to specific products (like APUs).
I'm trying to find this document.

 

[ShintaiDK]

I distinctly remember AMD saying in some presentation (about a year ago?) that they found a way to double perf/W each generation. Now I don't know if they specifically clarified what they view(ed) as a new "generation", if that includes node shrinks or not and if it is only applicable to specific products (like APUs).
I'm trying to find this document.

Most likely node shrink. But that fun is long over.

 

[monstercameron]

the 285 could be interesting in an XFire setup for 500$ which could be stronger than a single 500$ card. but then again, you'd have to settle (pun) with Microstutter.

EDIT: 250$ seems a bit much for this card IMO.

in dx9 titles, this card has a xdma block.

 

[boxleitnerb]

Most likely node shrink. But that fun is long over.

Found it:
http://www.pcgameshardware.de/AMD-R...mit-Dr-John-Gustafson-von-AMD-Teil-2-1065308/

Original sentence:

Und ohne zu viel verraten zu wollen: Wir haben Mittel und Wege gefunden, um den Wirkungsgrad unserer zukünftigen Chipgenerationen pro Watt zu verdoppeln.

My translation (native German speaker):
Without wanting to reveal too much:
We have found ways and means to double the efficiency (per Watt) of our future chip generations/families.

The context of this interview was clearly on GPUs. I don't know how one can make such a general statement since there are always hurdles and new requirements that cannot be foreseen. How many GPU generations did he mean? The next one, two, ten? What workloads?

Anyway, remembering that interview and that statement I really had high expectations for Tonga (and Hawaii for that matter). It was 14 months ago, but I don't know if AMD had just started implementing these ways and means or if they were already well underway by the time of this interview.

 

[3DVagabond]

Found it:
http://www.pcgameshardware.de/AMD-R...mit-Dr-John-Gustafson-von-AMD-Teil-2-1065308/

Original sentence:


My translation (native German speaker):
Without wanting to reveal too much:
We have found ways and means to double the efficiency (per Watt) of our future chip generations/families.

The context of this interview was clearly on GPUs. I don't know how one can make such a general statement since there are always hurdles and new requirements that cannot be foreseen. How many GPU generations did he mean? The next one, two, ten? What workloads?

Anyway, remembering that interview and that statement I really had high expectations for Tonga (and Hawaii for that matter). It was 14 months ago, but I don't know if AMD had just started implementing these ways and means or if they were already well underway by the time of this interview.

It's hard to say as they are currently mixing families/generations together and without a node shrink. Isn't that a fairly typical scenario though when you take shrinks and generational improvements into acct.?

 

[TrulyUncouth]

Found it:
http://www.pcgameshardware.de/AMD-R...mit-Dr-John-Gustafson-von-AMD-Teil-2-1065308/

Original sentence:


My translation (native German speaker):
Without wanting to reveal too much:
We have found ways and means to double the efficiency (per Watt) of our future chip generations/families.

The context of this interview was clearly on GPUs. I don't know how one can make such a general statement since there are always hurdles and new requirements that cannot be foreseen. How many GPU generations did he mean? The next one, two, ten? What workloads?

Anyway, remembering that interview and that statement I really had high expectations for Tonga (and Hawaii for that matter). It was 14 months ago, but I don't know if AMD had just started implementing these ways and means or if they were already well underway by the time of this interview.

Throw in HBM along with 20/16nm and I am sure they will hit 2x perf/watt. I'll be very interested to see how these actually perform and how much they cost in retail. Surely if the performance isn't there they will drop price until it is.

 

[3DVagabond]

Throw in HBM along with 20/16nm and I am sure they will hit 2x perf/watt. I'll be very interested to see how these actually perform and how much they cost in retail. Surely if the performance isn't there they will drop price until it is.

There are rumors that GM200 will come ~Oct. I hope AMD will have something else bigger coming soon too. Without it who knows what pricing will do for both companies. Little chips like this aren't making too much of a statement towards AMD's performance capabilities.

 

[boxleitnerb]

It's hard to say as they are currently mixing families/generations together and without a node shrink. Isn't that a fairly typical scenario though when you take shrinks and generational improvements into acct.?

Yes it is, but then such emphasis in that statement about new "ways and means" would make no sense, would it?

GM200 won't be here in October, I assure you of that. Tape out was supposedly in June, so the very earliest date I see is December and only a paper launch if any launch 2014 at all.

 

[RussianSensation]

I distinctly remember AMD saying in some presentation (about a year ago?) that they found a way to double perf/W each generation. Now I don't know if they specifically clarified what they view(ed) as a new "generation", if that includes node shrinks or not and if it is only applicable to specific products (like APUs).
I'm trying to find this document.

Has nothing to do with discrete desktop GPUs. The article you posted failed to clarify the context. Those statements were primarily related to APU gains through HSA and related to programs for specific CPU workloads that benefit from parallel nature of GPGPUs. Even the fact that they are discussing 25X gains in 6 years is obvious this has nothing to do with gaming graphics:
http://mobile.extremetech.com/lates...-efficiency-improvement-in-the-next-six-years

Any claims about Tonga delivering some 50-100% perf/watt are all online fanboy fabrication. AMD or no other credible source ever claimed such numbers. And everyone else who claimed Tonga was AMD's response to Maxwell also just made it up of thin air. AMD never intended to use GCN 1.1 against Maxwell. The competitor is GCN 2.0/300 series. Now if NV beats 300 series to launch, that's AMD's problem. As a product generation, 285/285X are a year late since AMD should have launched them alongside 260X/290/290X.

Furthermore, it completely defies logic how a GCN 1.1 Tonga could provide 2X the performance/watt, yet it's the same architecture as 260X/290 series. You can't have some GCN 1.1 SKU with some gargantuan leap in perf/watt on the desktop but another GCN 1.1 SKU is barely better than 7870 GCN 1.0. It doesn't work like that. As ocre mentioned, almost everything about Tonga was fabricated hype based on lack of understanding of its architecture and specs.

 

[boxleitnerb]

Has nothing to do with discrete desktop GPUs. The article you posted failed to clarify the context. Those statements were primarily related to APU gains through HSA and related to programs for specific CPU workloads that benefit from parallel nature of GPGPUs:
http://mobile.extremetech.com/lates...-efficiency-improvement-in-the-next-six-years

Any claims about Tonga delivering some 50-100% perf/watt are all online fanboy fabrication. AMD or no other credible source ever claimed such numbers. And everyone else who claimed Tonga was AMD's response to Maxwell also just made it up of thin air. AMD never intended to use GCN 1.1 against Maxwell. The competitor is GCN 2.0/300 series. Now if NV beats 300 series to launch, that's AMD's problem. As a product generation, 285/285X are a year late since AMD should have launched them alongside 290/290X.

Are you sure? The article I posted is way older than the new one with 25x perf/W increase. Maybe those two aren't directly related? We cannot be sure I think.

And I don't understand why AMD wouldn't position Tonga and co. against Maxwell. If Nvidia can achieve such an perf/W improvement on 28nm, why wouldn't AMD be able to do the same? Wouldn't that be a sensible goal, especially in light of how costly/late 20nm would be (I'm sure AMD and NV knew about this very early, i.e. when designing the first Maxwell and Tonga respectively).

I especially don't understand what AMD should be able to do with GCN 2.0 what they cannot do with GCN 1.1. Their strategy was outlined as bringing (incremental) improvements to the market as they are ready and not wait years until they can improve a whole lineup -> see Bonaire and Hawaii.

It would only make sense if
a) AMD was not able to increase the efficiency to Maxwell levels on 28 nm at all or at least in the current timeframe
b) the financial cost of doing so was deemed too great and unnecessary at this point

 

[caswow]

so because the mini maxwell is somewhat of efficient everybody thinks it will scale 1:1 up to the mid and highend cards? because thats the base of all the speculations here. and i cant recall any realworld reviews of the new card so whats the deal? crying before even seeing anything is the new thing right now or what?

 

[Erenhardt]

so because the mini maxwell is somewhat of efficient everybody thinks it will scale 1:1 up to the mid and highend cards? because thats the base of all the speculations here. and i cant recall any realworld reviews of the new card so whats the deal? crying before even seeing anything is the new thing right now or what?

Shhhh! I hope this thread gets snapshotted. Will be fun to revisit those posts after the reviews.

gtx750 ti with 640 shaders?! Booo I say. We had 768 shader gtx650ti released 2 years ago. Much dissapoint. Nv is loosing grasp.

and gloom

 

[Insomniator]

Cards a joke, and its not like Nvidia has brought anything great to the mid-high end market either (because they haven't needed to I assume). The 750ti is nice, but where the hell is the rest of the lineup?

We're still dealing with rebrands and or redesigns that are slightly better or worse, for basically the same price and power. Have games caught up though? Is there anything more harsh on a 7950 now than 2 years ago? Watchdogs? Or are we waiting for Witcher 3 and Unreal 4?

 

[boxleitnerb]

You have to understand that it doesn't necessarily make financial sense to release too many GPUs in a short time period. Nvidia surely wanted to sell the produced GK110s (as to their AIBs). There is a balance between saving money with making a smaller GPU (GM204 vs GK110), selling at the same price point and the R&D costs associated with that new GPU.

Imagine theoretically that an IHV could increase efficiency per mm2 each 6 months by 50%. But this development were to cost $250 million each time. Would it be worth to release these new GPUs every 6 months or would you make losses in the process (or get an even zero which would be just as bad since then you wouldn't have money to invest in new R&D)?

 

[redzo]

so because the mini maxwell is somewhat of efficient everybody thinks it will scale 1:1 up to the mid and highend cards? because thats the base of all the speculations here. and i cant recall any realworld reviews of the new card so whats the deal? crying before even seeing anything is the new thing right now or what?

GPU architectures scale pretty good. This more of a fact and less speculation. The rest is speculation.

 

[FatherMurphy]

Agreed. The Kepler architecture scaled nearly 1:1 from GK107 to GK110 in efficiency, I believe. On an uber-mature 28nm process, I see no reason to doubt that the scaling would be any worse from GM107 to GM204.

 

[Erenhardt]

Agreed. The Kepler architecture scaled nearly 1:1 from GK107 to GK110 in efficiency, I believe. On an uber-mature 28nm process, I see no reason to doubt that the scaling would be any worse from GM107 to GM204.

We are not funding here new Anand religion, your believes are not in place.
Do you have any scientific proof of that? Like say: both chips manufactured with the same tools, in the same foundy, by the same people, in the same time, running at the same frequency and voltage scale 1:1 per cuda core in performance and power consumption? If yes, then feel free to open a new thread with your findings!