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

[96Firebird]

Because the site i criticised measure the whole plateform, a bigger GPU will stress the CPU more than a little one, they are measuring the efficency of a whole plateform not of a GPU in isolation, for this you must isolate the GPU power comsumption and relate it to its fps and this is what hardware.fr did, they measure the power through the pci connector as well as through 12V rails that feed the cards, hence they can extract numbers that have a true technical meaning, their numbers will be the same whatever the plateform.

Why do we care about an isolated video card power consumption? I'm running the card on a system, not standalone.

 

[PPB]

Good times. Reception for barts was not the best, but one or two monts later both cards became top sellers.

Really good multi-GPU scaling was Bart's appeal to the mid range gamer.

Buy one now, then buy another one knowing you have 80%+ scaling assured in CFX.

I feel AMD is taking blame for being the real innovator into adding a crapload of Hardware features that cost die space and thus are forced into more cramped designs. I for one cant wait dGPU HSA and better OpenCL support on Tonga. Waiting for someone to make it work with Vray-RT like W9100 was proven to, too.

 

[el etro]

(Seronx Mode on): Iceland is probably a Pitcairn-sized chip made by the same arch as Tonga.

 

[RussianSensation]

Probably because it wasn't called a 665Ti. See what I did there?

No, I don't see. Had AMD called this 285X, you'd have a point. Yet in AMD's line 285 succeeds 280 and 285X succeeds 280X. X overrides 5, meaning 280X is faster than 285 non-X. I don't agree with the naming convention but 650Ti vs. 660Ti vs. 665Ti is not the same comparison since you are keeping the Ti in all cases.

You are a pure bang for buck guy. That is it.

You mean I am consistent? In your case you pick and choose which metric matters like a PR artist:

650Ti - performance/watt
670 - performance/$ (while ignoring this for HD5000/6000/7000 series for most of their useful lives)
760 - performance /$ (see above)
460 - massive overclocking potential (while ignoring 5850/7850/7950/7970 overclocking and 6950 unlocking)
750Ti - performance/watt (but ignored early GCN cards trumping Fermi in performance/watt for 6-9 months)
780Ti - ultimate performance (yet ignored R9 295X trumping Titan Zor HD6990 trumping 590)

I could go through every card NV sells and find reasons why it's better than an AMD's card. However, I try to cut through the AMD and NV marketing BS (like recommending 280/280X over 285) or 270/270X over GTX750Ti (because performance/watt metrics change entirely in the context of the overall system power usage). :thumbsup:

You said there are no compelling reasons for 285 to exist, I gave you plenty of them even though 285 isn't as impressive as Maxwell.

 

[crisium]

look at this(From hardware.fr):

Pretty good "IPC" improvements thanks to the new memory techniques.

And great post, Russian. 285 really only looks bad because a) AMD already has such good value its competing against itself and b) power efficiency is disappointing. Compared to Nvidia's offerings, it fits right where it needs to be. Same price, better performance. The Nvidia tax is there for those who want G Sync and are afraid of AMD drivers, but otherwise the 285 stands strong against Nvidia (it's just the 280x and 280 are so cheap). And it seems you are right that it was good experience engineering this chip.

I enjoyed your post last page, but one thing I have to absolutely agree with on Keysplayer is that the naming scheme is bollocks. It's now definitely inconsistent since the 265 beats the 260X but now the 280X beats the 285 which is opposite. And when you have to get into things such as "the tens matters more than the X, but the X matters more than the 5" you KNOW you have a bad naming convention. I miss xx50 and xx70. If the 280 were called the 8750 and the 280x were called the 8770, then the 285 could simply be the 8760. But trying to fit a card that performs inbetween a 280 and 280X? You either give it a higher/lower number and have it be somewhat misleading (as AMD chose to do) or you'd have to give it a whole new arbitrary moniker like 280Z and then you'd be as silly as Nvidia during the the 8800 series (GS<GTS640<GT<GTS512<GTX). God save the old AMD naming scheme.

 

[FatherMurphy]

I agree, very good post, Russian. We'll know next week (it sounds like) whether AMD has correctly positioned these Tonga cards against the upcoming 2nd Gen Maxwell cards.

I can't help but think that AMD would have preferred to be competing in the sub-$300 range with a much smaller chip, however. Anything to help their bottom line and margins. Not that I think Nvidia will, but it could, at least based on the die size of GK104, compete dollar for dollar against Tonga with a GK104. If the 770 was re-branded as 860 (or 960), for $250-ish dollars, you'd have a very tight competition there (tighter than currently exists, where, based on pure performance/$, AMD appears to have the upper hand).

I suspect there is a good chance that this chip gets directly die-shrunk ASAP (as I equally suspect, like many others here, that this chip was originally destined for 20nm).

Edit: I guess my point is that 20nm needs to hurry the eff up.

 

[Keysplayr]

No, I don't see. Had AMD called this 285X, you'd have a point. Yet in AMD's line 285 succeeds 280 and 285X succeeds 280X. X overrides 5, meaning 280X is faster than 285 non-X. I don't agree with the naming convention but 650Ti vs. 660Ti vs. 665Ti is not the same comparison since you are keeping the Ti in all cases.



You mean I am consistent? In your case you pick and choose which metric matters like a PR artist:

650Ti - performance/watt
670 - performance/$ (while ignoring this for HD5000/6000/7000 series for most of their useful lives)
760 - performance /$ (see above)
460 - massive overclocking potential (while ignoring 5850/7850/7950/7970 overclocking and 6950 unlocking)
750Ti - performance/watt (but ignored early GCN cards trumping Fermi in performance/watt for 6-9 months)
780Ti - ultimate performance (yet ignored R9 295X trumping Titan Zor HD6990 trumping 590)

I could go through every card NV sells and find reasons why it's better than an AMD's card. However, I try to cut through the AMD and NV marketing BS (like recommending 280/280X over 285) or 270/270X over GTX750Ti (because performance/watt metrics change entirely in the context of the overall system power usage). :thumbsup:

You said there are no compelling reasons for 285 to exist, I gave you plenty of them even though 285 isn't as impressive as Maxwell.

They are compelling to you. So enjoy. What can I say?
It was hyped too much and a letdown. And 280X is quite a bit faster I'm seeing. Maybe it should have just been called 280XL.

 

[Enigmoid]

Because the site i criticised measure the whole plateform, a bigger GPU will stress the CPU more than a little one, they are measuring the efficency of a whole plateform not of a GPU in isolation, for this you must isolate the GPU power comsumption and relate it to its fps and this is what hardware.fr did, they measure the power through the pci connector as well as through 12V rails that feed the cards, hence they can extract numbers that have a true technical meaning, their numbers will be the same whatever the plateform.
Actualy he cant and he s on the bad side for the reasons above,
with Hfr protocol you cant twist the numbers by changing plateforms or CPUs.

Yes you are right. However every GPU needs a CPU and if your drivers are not as good the CPU may be using more energy than it needs to, thus driving system efficiency down.

I'll give you some.

1) Performance. Even if R9 285 isn't as good from a performance/$ as 280/280X, it's still better than the $250 760. Sure, some 760s are going for $230 but it'll be really quick before 285 matches the 760 in price and well 285 is faster overall from nearly every review online:

http://www.computerbase.de/2014-09/amd-radeon-r9-285-test-benchmarks/5/

More importantly, most of the time it's expected that NV/AMD card at a similar price level could be trading blows +/-15% in some games. However, in some games, 760 completely bombs against the 285. It'll take a card that's sometimes 10-15% slower at a similar price but no way when it's 30% and 40% slower in big games like this.

R9 280 had an MSRP of $279 but market price quickly dropped to $250. AMD priced 285 at $250 so they could clear 280 asap. Once 280's stock runs out, 285 will drop below $250.

True, but the 760 with rebates can be found for $215 for a good aftermarket edition.

http://www.ncixus.com/products/?usa...=02G-P4-2765-KR&manufacture=eVGA&promoid=1011

A 280 costs about $200

http://www.newegg.com/Product/Produc...D=3938566&amp;SID=

And a R280X about $220

http://www.ncixus.com/products/?usa...n=CGAX-R928X7&manufacture=Club3D&promoid=1011

IMO this seems just like the R2xx series launch. The rebrand/relaunch launched at higher prices. (AMD's attempt to drive prices and profits up?).

A price drop is needed bring any improvement on the perf/$ front.

3) Strengthening track record of high transistor density - an engineering exercise for the future.

Tahiti = 11.8 million / mm2
Hawaii = 14.16 million / mm2
Tonga = 13.93 million / mm2

vs.

780TI = 12.66 million / mm2
750Ti = 12.64 million / mm2

Since AMD hasn't been able to compete with NV on die size, they have to cram more transistors into a smaller space to make up for this disadvantage. More experience under your belt hitting 14 million / mm2 gives more confidence that perhaps the next gen design you could shoot for 15-16 million / mm2.

Tahiti was rushed resulting in poorer efficiency and a larger die.
Pitcarin is 2800M transistors and 212 mm^2. Works out to 13.21 M/mm^2. Oland is 1040 and 90 mm^2 (11.55 M/mm^2). Transistors/mm^2 is also almost always in favour of a larger die.

GK107 : 11.01
GK106 : 11.49
GK104 : 12.04

This has to do with non-duplicated resources such as the video engine, etc. I expect to see larger maxwell dies with significantly higher transistors/mm^2. Maxwell seems very bandwidth efficient (850m DDR3 is about 70-85% the performance of the 860m despite being clocked lower and having 32 GB/s vs. 80 GB/sec bandwidth).

GK110 is 12.88 M/mm^2 (7.1B, 551 mm^2), largely because of the greater DP (also why Tahiti sucked). Hawaii is significantly denser despite the greater DP (likely the new memory bus).

4) Improving memory bandwidth efficiency by 40%.

7950 has 36% more memory bandwidth (240 vs. 176) but it's slower. While this benefit doesn't mean much for 280/280X user, it means AMD reduced the need for redundant memory bandwidth that plagued 280/280X/R9 290X series by utilizing the available bandwidth more effectively. This could mean a 384-bit bus on 390X to reduce power usage, OR less memory bandwidth bottlenecks of GDDR5 before GDDR6/HBM drops OR less need to use more expensive and power hungry GDDR5.

Nope can't say that. 580 had the same bandwidth as the 680 and the 680 was significantly faster. Was the 580 simply bandwidth bound or did it simply have extra bandwidth?

Look at the 750 TI vs. the 265 and the 260X and bandwidth efficiency is completely differenct.

5) More than doubling the geometry performance on the same node.

Alright, so 285 doesn't produce the magical 2x performance/watt of Kepler but none-the-less AMD more or less doubles Tahiti's geometry performance and trumps R9 290X by 70%! Since tessellation has been a weak spot for AMD, this is a big breakthrough without even discussing what GCN 2.0 will bring over GCN 1.2. OK so we know GCN 1.2 won't match Maxwell since it just catches up to Kepler but with AMD finally catching up to NV in tessellation, it forces NV to innovate and double tessellation with Maxwell. You know that magical phrase - competition forces innovation. :thumbsup:

This is good.

6) Architectural timeframe context. Perhaps AMD cannot afford or doesn't want to take a risk of designing a brand new architecture like GCN for 3-5 years and then have it fail (Bulldozer anyone). First of all, it's very expensive and AMD doesn't have the financial resources of Intel/NV. Second of all, it's too risky. What they keep doing is improving GCN step-by-step as GCN is solid enough to last them another 2-3 years. They doubled tessellation performance per mm2, they increased memory bandwidth efficiency by 40%*, and they did this in about a year since GCN 1.1 launched. Not bad consider Maxwell has been in development for 3-4 years! They have another 6-12 months to refine GCN 1.2 even more to improve performance/Stream processor and perhaps wait long enough for 20nm to close the performance/watt gap. Since NV is repeating bifurcation of its next generation (GK104-> GK110 and now surely GM204 / GM210), maybe AMD thinks it's better to have incremental changes to the architecture

* - As claimed by AMD

There are even more points:

- new pre-scaler and upgraded display controllers
- upgraded instruction set for compute
http://videocardz.com/51462/amd-officially-announces-radeon-r9-285-never-settle-space-edition

^^^ These minor improvements don't translate to more performance in games but they allow AMD to test and slowly implement new features that deal with 4K and 8K compatibility, as well as focus on improving GPGPU performance for its HSA initiative and FirePro line.

These are nice. Seems a lot like the hype of something like AVX 2.0 though.

So overall, for gaming only, 285 doesn't set the world on fire and fails hard on performance/watt but AMD has done a lot of other things here that allows them to focus more on performance/watt and other gaming aspects with R9 300 series having at least addressed memory bandwidth and tessellation issues with GCN 1.2.

Thing is though, between $200 R9 280, $250 R9 285, $260-280 R9 280X and $325 R9 290, $450 R9 290X, AMD has the entire $200-500 segment locked in for anyone but the most hardcore NV loyalist. And that means NV will have to respond with aggressive improvement in performance, performance/watt and performance/$$ to keep their market share.

AMD's desktop GPU competitiveness is nothing like its CPU competitiveness. That's a win-win for us gamers since NV can't stand still.

I agree. Personally after the debacle that AMD on mobile was (7000 series and enduro) I would pay a small premium for Nvidia though nothing large.

Tonga is OK. Its nothing to really get excited about.

 

[HurleyBird]

OK so we know GCN 1.2 won't match Maxwell since it just catches up to Kepler but with AMD finally catching up to NV in tessellation.

Well, they're catching up in unrealistic and/or poorly implemented scenarios where polygons are subdivided to the size of a pinhead. IIRC they were already ahead in more moderate and realistic scenarios where tessellation is used intelligently. Tonga likely extends this lead.

 

[Jack77]

Well, in my simple view, the 285 is just a 370X released early, so that they can already benefit from cost savings (smaller memory bus, less memory...).

Additional features must for now make up for the lack of performance improvement. But it will be a great 370X

 

[sontin]

AMD's desktop GPU competitiveness is nothing like its CPU competitiveness. That's a win-win for us gamers since NV can't stand still.

AMD hasnt been able to compete with GM107 on a perf/watt basis and now they have released a card which is equal to a 30 months old GTX680.

So, what is AMD's answer to GM204? A 300w 290X? :hmm:

 

[3DVagabond]

They are compelling to you. So enjoy. What can I say?
It was hyped too much and a letdown. And 280X is quite a bit faster I'm seeing. Maybe it should have just been called 280XL.

I must have missed all this hype. We had one benchmark released a few days before launch. No leaks, no hype, that I recall.

This card brings AMD's latest feature set to the $250 price point. Simple enough?

 

[parvadomus]

AMD hasnt been able to compete with GM107 on a perf/watt basis and now they have released a card which is equal to a 30 months old GTX680.

So, what is AMD's answer to GM204? A 300w 290X? :hmm:

Im not impressed by maxwell. And it looks like AMD is not either, they are throwing rare dies that had nothing to do with power efficiency.
Maybe when NV actually do something other than a low end chip that is handily beaten by an 3 year old die, the competition might start.

 

[Keysplayr]

I must have missed all this hype. We had one benchmark released a few days before launch. No leaks, no hype, that I recall.

This card brings AMD's latest feature set to the $250 price point. Simple enough?

Yes. You must have. The nomenclature alone would be enough.

 

[sontin]

Im not impressed by maxwell. And it looks like AMD is not either, they are throwing rare dies that had nothing to do with power efficiency.
Maybe when NV actually do something other than a low end chip that is handily beaten by an 3 year old die, the competition might start.

Yeah, how could anybody be impressed by Maxwell:

http://www.hardware.fr/articles/926-7/consommation-efficacite-energetique.html

 

[boxleitnerb]

Energy efficiency is everything since nowadays where you hardly can increase the TDP of any given performance segment (much) further, efficiency = performance. Unless you want a 300, 350, 400W GPU...where would that end? 250-300W is enough for a highend GPU imo. And imo that stands for years to come.

 

[parvadomus]

.

Already saw that long time ago, and Im sure AMD did also, even before you did, they are still releasing whatever they want, and wasting silicon in "inefficient dies". They must be really impressed :awe:

 

[Abwx]

Yeah, how could anybody be impressed by Maxwell:

http://www.hardware.fr/articles/926-7/consommation-efficacite-energetique.html

There is two graphs on this page...the one you didnt post :

 

[sontin]

And you think this looks better? They are still losing to Kepler.
Heck nVidia doesnt even need to improve their perf/watt ratio they only need to make GK104 bigger...

 

[AtenRa]

Performance per watt alone means nothing. You really need to have a context, something to compare like performance, because you can have the same performance per watt at 100W TDP and at 500W TDP.

 

[boxleitnerb]

Performance per watt alone means nothing. You really need to have a context, something to compare like performance, because you can have the same performance per watt at 100W TDP and at 500W TDP.

Hello Capt. Obvious, trying to downplay an Nvidia advantage...how surprising. Better perf/W means you can ultimately build faster GPUs than your competition - unless of course the competition makes its card more noisy or more expensive with water cooling or such. But at a certain point the vast majority of buyers won't accept an ever increasing TDP. I'm glad we have hit a wall with that one. Wattage has barely increased since GTX 480 and that is a very good thing.

 

[Erenhardt]

(Seronx Mode on): Iceland is probably a Pitcairn-sized chip made by the same arch as Tonga.

Sushi said its going to be big... like double that. Hawaii's bigger brother
Probably watercooled (link to the asetek multi M$ deal for watercooling IO for GFX cards that I can't find now)

285 is very smart move. Just link two dots. AMD ground breaking free sync is not going to work with 280/X - a go to card for mainstream $200 market. Who would buy 280/X after this announcement? Here comes AMD with cheaper to produce, faster card that supports FS and a lot more latest tech goodies which make up for a small premium over their equivalent parts (280 and 280X).

Now the king of the mainstream market is back properly refreshed. Not like it was with 7970 and 280x (rename), but a full feature list update. This should be the 280/X, not a straight rebrand we had.

 

[AtenRa]

Hello Capt. Obvious, trying to downplay an Nvidia advantage...how surprising. Better perf/W means you can ultimately build faster GPUs than your competition - unless of course the competition makes its card more noisy or more expensive with water cooling or such. But at a certain point the vast majority of buyers won't accept an ever increasing TDP. I'm glad we have hit a wall with that one. Wattage has barely increased since GTX 480 and that is a very good thing.

GPU A produces 30fps by using 100Watts, performance per watt is 0.3

GPU B produces 70fps by using 250Watts, performance per watt is 0.28

GPU B has lower performance per watt but it is more than 2x faster

 

[FatherMurphy]

285 is very smart move. Just link two dots. AMD ground breaking free sync is not going to work with 280/X - a go to card for mainstream $200 market. Who would buy 280/X after this announcement? Here comes AMD with cheaper to produce, faster card that supports FS and a lot more latest tech goodies which make up for a small premium over their equivalent parts (280 and 280X).

Now the king of the mainstream market is back properly refreshed. Not like it was with 7970 and 280x (rename), but a full feature list update. This should be the 280/X, not a straight rebrand we had.

I just don't know about all that.

First, the Tonga die is only negligibly smaller, so it is not cheaper to produce. I guarantee you that AMD does not want to be selling a 350mm2+ die for $220-$300 for another year.

Second, there are indications that the die might actually have a 384-bit bus (128-bit of which was disabled for 285) that will be enabled for the 285x (see Scott Wasson's comments in his review); if that is the case, 285x will certainly not be cheaper to produce (i.e. because it will have the same or more memory as the current 280 series).

Third, although the feature set is updated somewhat, the value of such features for this particular chip are dubious. Sure there is the TrueAudio... but what games are going to use that? FreeSync isn't projected for true commercial viability (i.e. monitors for sale) until the 1st part of 2015. By that point, AMD sure as hell better be thinking about replacing Tonga at least with a shrunk down version in order move the performance meter forward.

Fourth, this chip doesn't move the performance or performance/$ metrics forward at all. To call it a refresh in the usual sense of the word is somewhat misleading. It brings an updated feature set, some features of which are not going to be useful a while (perhaps for the majority of this chip's life cycle).

I don't think Tonga is smart or dumb. I think it's all AMD can bring to the table right now, even though this chip only separates itself from its 2.5 year old predecessor through relatively minor feature set advances (though XDMA is cool). I expect AMD has much better, more progressive technology for next year, but Tonga is simply a tourniquet to stop the bleeding until then.

 

[el etro]

Sushi said its going to be big... like double that. Hawaii's bigger brother

whoa, very good to know.