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

[MathMan]

My biggest let down was the efficiency. Not the name, not the price. It was the efficiency. And its not a let down cause i wanted to buy it. It is kinda worrisome. Perhaps its nothing to think on but i find myself concerned with it when looking at the bigger picture. No one wants to see AMD fall far behind.

The efficiency of a full Tonga should be quite a bit better than Tahiti. Dave Baumann has hinted that it doesn't have a much DP resources as Tahiti. Let's speculate that it also doesn't have ECC logic. That must free up quite some area for other stuff (geometry processing, larger caches, still 384 bits.)
If Tonga reduced is also doing well against Tahiti, then the full version will blow it out of the water. So it's really more efficient. Probably as efficient as GK104 (which was really quite good in that respect.)

The problem is: the efficiency of gm107 is off the charts. If that carries over to gm204, that's where the problem starts.

Current rumors have a GTX980 at equal or above GTX780. (Larger than a 780Ti remains to be seen.) With a 256-bit bus no less. That's a different class.

Perhaps we are looking too deep and the situation doesnt get bleak. It can go anyway from here.

The real question is, as always, pricing. GTX750Ti is quite expensive for consumers. Nvidia chose margins over volume. They will probably do that same for gm204. AMD probably hopes they do. But no matter what price it will get, it's going to push AMD further into bargain bin territory, something they're very much used to in the CPU world.

 

[MathMan]

I disagree with everything here especially the bolded bit, how can you be confident? You do know that AMD also builds small core apus via tsmc?

The GPU revenue of Nvidia is 3 times larger than AMD. Nvidia also dominates I the high-end, very large die segment, so that acts like multiplier. Nvidia only produces at TSMC, AMD also at GF. There is no way AMD has larger volume at TSMC just because of (smaller!) APUs.

 

[MathMan]

Today on Bloomberg: "The CEO of the Sunnyvale, California-based company said in an interview today that AMD still has to get through 18 months of old technology before the new chips are available."

Ouch. My bet is that he was not only talking about CPUs.

 

[FatherMurphy]

I bet it will destroy GK104 in absolute performance. At the cost of extra DRAMs. I bet it will be destroyed by GM204 despite the latter having 2 DRAMs less. And if gm107 is an indication, I bet gm204 will destroy Tonga as well in perf/mm2. And, finally, gm204 will have better perf/W as well, so the perf/$ for cooling will be better for gm204 as well.

While I have no doubt whatsoever that GM204 will be competing against Hawaii, and not Tonga, in performance, I don't think GM204 will necessarily have equivalent or better "perf/mm2" as Tonga or Hawaii.

Recall that GM107 is a much larger chip than GK107 in order to accommodate the re-worked Maxwell SMMs. In effect, Nvidia (at least on 28nm) has decided to sacrifice die space (mm2) in order to produce a more efficient chip. Therefore, although GM107 is far more efficient and a much better performer than GK107, Nvidia had to make a big chip.

It goes to show that, often, performance/mm2 is a misleading statistic. In fact, it is mostly irrelevant except in situations where the chip is as big as it can be (e.g. GK110) or unless the performance/mm2 is so bad that you have to make an enormous chip to compete with a much small chip in performance, which cripples one's ability to price competitively.

 

[MathMan]

If you don't want to compare gm107 against gk107 in terms of perf/mm2, compare it against gk106...

 

[FatherMurphy]

I'm saying I don't particularly care about performance/mm2. It's an irrelevant metric except in the two situations I described above, neither of which applies here.

Even if you do compare them, again, I don't believe Maxwell achieves the same performance per mm2 as Kepler. I haven't run the numbers, but the 214mm2 GK106 is much faster than than 150mm2 GM107.

That is not a knock on Maxwell. It's a design decision which I think was well made. Basically, lose a bit of performance/mm2 to double perf/w.

Bringing this back to Tonga, Tonga didn't really move the perf/w much, only slightly moved the perf/mm2, and did not move the perf/$$ at all.

 

[MathMan]

I'm saying I don't particularly care about performance/mm2.

As a consumer, you shouldn't care. It doesn't impact you. For a company, it's very important.

Even if you do compare them, again, I don't believe Maxwell achieves the same performance per mm2 as Kepler. I haven't run the numbers, but the 214mm2 GK106 is much faster than than 150mm2 GM107.

I randomly took the BF4 bench results of AnandTech's GTX750Ti review.
GTX660 (gk106): 55.4fps/214mm2 = 0.261fps/mm2
GTX750Ti (gm107): 48.9/150 = 0.326 fps/mm2

That makes gm107 25% more perf/mm2 efficiënt than gk106.

USD different benchmarks if you like, I expect similar numbers.

That is not a knock on Maxwell. It's a design decision which I think was well made. Basically, lose a bit of performance/mm2 to double perf/w.

That's the threat to AMD: Nvidia found a way to improve perf/w and perf/mm2 in one architecture, and they don't seem to have an answer.

Bringing this back to Tonga, Tonga didn't really move the perf/w much, only slightly moved the perf/mm2, and did not move the perf/$$ at all.

Agreed. It's a necessary improvement, but I doubt it's sufficient.

 

[Cloudfire777]

Defending and buying old architectures when companies are lowering the prices of some cards, is undoubtly giving the companies signals that its allright to postpone great leaps in GPU technology and rather play with the price instead of working for progress in new architectures that will eventually lead to much greater cards. Not just performance and cost, but also heat and power draw.

Atleast buy cards of a new architecture when the alternative is there

 

[sontin]

While I have no doubt whatsoever that GM204 will be competing against Hawaii, and not Tonga, in performance, I don't think GM204 will necessarily have equivalent or better "perf/mm2" as Tonga or Hawaii.

GM204 will have much better perf/mm^2 than Tonga. Even Kepler is better in this regard with 296mm^2.

Recall that GM107 is a much larger chip than GK107 in order to accommodate the re-worked Maxwell SMMs. In effect, Nvidia (at least on 28nm) has decided to sacrifice die space (mm2) in order to produce a more efficient chip. Therefore, although GM107 is far more efficient and a much better performer than GK107, Nvidia had to make a big chip.

It goes to show that, often, performance/mm2 is a misleading statistic. In fact, it is mostly irrelevant except in situations where the chip is as big as it can be (e.g. GK110) or unless the performance/mm2 is so bad that you have to make an enormous chip to compete with a much small chip in performance, which cripples one's ability to price competitively.

That is not true at all. nVidia was able to put 640 Cores (1GPC, 5SMM, 2MB L2 Cache!, ...) into 60W. This was possible through the improvements in energy efficient. And with this they could build a larger die to put everything into it.

Except GM200 no Maxwell GPU will be size limited. Everyone of them will be limited by the maximum power allowed by nVidia.

 

[Attic]

What's up with the Anandtech review? Bad card, no card?

 

[MathMan]

What's up with the Anandtech review? Bad card, no card?

Life happened. Career changes etc. Just give them some slack.

 

[RaistlinZ]

What's up with the Anandtech review? Bad card, no card?

Pointless card.

 

[HurleyBird]

Life happened.

While it's true that Anandtech is going through a time of transition, it's disconcerting that these delayed video card reviews are becoming the norm instead of the exception.

 

[monstercameron]

they already lost the race to first, does it really matter now?

 

[itsmydamnation]

is it known for sure where Tonga is manufactured? i see internet links saying Global foundries , could explain some of the high power numbers. taking it easy on the first large GPU to be manufactured there.

 

[metalliax]

is it known for sure where Tonga is manufactured? i see internet links saying Global foundries , could explain some of the high power numbers. taking it easy on the first large GPU to be manufactured there.

It says "Made in Taiwan" which implies TSMC made it. Bummer, as I was hoping this would be a GloFo chip too. I believe it is on the 28nm HPM process, while the Tahiti chip was made on the 28 HP process. The HPM process should be better at lower frequency though, so hopefully the mobile part (m295x) ends up being more efficient.

 

[RussianSensation]

These AMD doom and gloom threads are too funny. NV was 6 months late with Fermi 470/480, 9 months late with 460, and competed with inferior GTX200 series all that time = weathered the storm, gaining market share with Fermi from 2010 to 2012. NV was 6-9 months late with 650-660Ti roll-outs and managed alright. AMD was nearly 6 months late with R9 290 series and it didn't go bankrupt now. Look at where we are at now: NV doesn't have a single card worth buying from $170-$1000 level and people are calling end of the world for AMD. At every price level mentioned, AMD has price/performance and/or top performance locked in. Are people rushing to the store to buy AMD? Nope, not really.

What's more interesting, nearly every single poster who is making these doom and gloom posts is either an NV owner or only buys NV. Look, don't complain that NV keeps raising prices and keeps them high when every single generation since HD4000 series AMD beat NV on price/performance at nearly every price level and what did you do to tell NV that you aren't going to accept premiums? Nothing. Exactly.

I mean let's get the details out - for the last 3 consecutive generations, you could buy almost 2 of AMD's 2nd best cards and smash NV's flagship to pieces:

$600 Unlocked 6950s vs. $500-550 GTX580
$600 7950s vs. $500 GTX680
$750 R9 290s vs. $650-700 780Ti

In all of these cases, the dual AMD setup would beat NV's by 40-70%. So please get off the whole NV can keep prices high and we are doomed since AMD can't compete nonsense. First of all, you have no one but yourself to blame if you think a 10% faster 980 for $500-550 is somehow a good value by simply accepting it as such. How about skip that and get the GM210 or wait until the value proposition for upgrading is there?

It's the consumers who allow these high prices to remain because they have this constant need to upgrade. Nothing wrong with that but then if you always want the latest tech, you almost have to pay the early adopter premium. Sometimes NV hits a home run like $399 670, an other times it milks like $649 280 or 780. It often makes sense to buy when both AMD and NV have both lineups out and are viciously competing because that's when the best game bundles and price reductions happen on newer tech.

Secondly, even when AMD offered unbeatable price/performance and enthusiast overclocking features and dual bioses for flashing and unlocking, people still kept and keep buying NV. It's clear that AMD's performance and price/performance are not the issue here for most consumers especially since 290s dominate 780Tis at 4K and multi-monitor gaming and yet high end enthusiasts still buy 780Tis.

So until AMD can change brand perception, driver perception or bring some killer features, even IF they brought a card that beat GM204 in every metric, people would still buy NV. And finally, as has been stated by many informed members, the critical battlefield is mobile dGPU space since most growth is happening there. In that regard, AMD is far behind in design wins, vendor relations and Enduro/Optimus.

 

[Silverforce11]

It's clear that AMD's performance and price/performance are not the issue here for most consumers especially since 290s dominate 780Tis at 4K and multi-monitor gaming and yet high end enthusiasts still buy 780Tis.

As I've said, the reason the R285 is bad isn't because its not competitive with NV's offerings, it is, very much so.

It's bad because its not competitive with AMD's own offerings, 7950/70, R280/X etc.

NV don't need to compete on perf/$ because they never had to. Plenty of loyal NV fans don't CARE what AMD has to offer, period. Sadly, they never appreciate that if AMD is competitive, its good news for them, because it forces NV to drop prices a little bit so now they can get their NV cards cheaper. A logical person would want competition. A [insert whatever you want] wants AMD to die.

 

[3DVagabond]

As I've said, the reason the R285 is bad isn't because its not competitive with NV's offerings, it is, very much so.

It's bad because its not competitive with AMD's own offerings, 7950/70, R280/X etc.

NV don't need to compete on perf/$ because they never had to. Plenty of loyal NV fans don't CARE what AMD has to offer, period. Sadly, they never appreciate that if AMD is competitive, its good news for them, because it forces NV to drop prices a little bit so now they can get their NV cards cheaper. A logical person would want competition. A [insert whatever you want] wants AMD to die.

Of course they realize it. That's why they blame AMD for nVidia's high prices (but pay them anyway).

 

[MathMan]

Of course they realize it. That's why they blame AMD for nVidia's high prices (but pay them anyway).

The scent of straw man is strong around here.

There are many reasons people buy a product for a certain brand. They have good experience with it in the past. They have particular features that they like. They think the brand is plain and simple cooler than the other. (Let's face it, AMD is a real champion at being uncool, just compare the design of their websites.) For many, cost is not the only factor. A $50 price difference for otherwise equivalent cards is, what, one dinner for two at a slightly upper scale restaurant? Nothing worth worrying about...

 

[Erenhardt]

The scent of straw man is strong around here.

There are many reasons people buy a product for a certain brand. They have good experience with it in the past. They have particular features that they like. They think the brand is plain and simple cooler than the other. (Let's face it, AMD is a real champion at being uncool, just compare the design of their websites.) For many, cost is not the only factor. A $50 price difference for otherwise equivalent cards is, what, one dinner for two at a slightly upper scale restaurant? Nothing worth worrying about...

LOL. But 20Watts more on hawaii is breaking the bank.:whiste:

 

[3DVagabond]

The scent of straw man is strong around here.

There are many reasons people buy a product for a certain brand. They have good experience with it in the past. They have particular features that they like. They think the brand is plain and simple cooler than the other. (Let's face it, AMD is a real champion at being uncool, just compare the design of their websites.) For many, cost is not the only factor. A $50 price difference for otherwise equivalent cards is, what, one dinner for two at a slightly upper scale restaurant? Nothing worth worrying about...

That was not what I was commenting on. Do you know why many people said that the Titan was $1K? It was because AMD didn't have a competitive card. It wasn't because nVidia just decided that their customers would be willing to pay $1000 for it. Which of course was the real reason, and they were right. My response was specifically to Silverforce11's post, "Sadly, they never appreciate that if AMD is competitive, its good news for them, because it forces NV to drop prices a little bit so now they can get their NV cards cheaper." I wasn't discussing people's willingness to pay more for nVidia.

 

[yannigr]

they already lost the race to first, does it really matter now?

When you are the first page that many open when searching for hardware reviews, yes it matters. Being second, third or tenth doesn't really matter. Not having exclusivity, first or tenth it could be just a matter of minutes or a couple hours. 99% of the readers wouldn't notice. But no review at all it is a problem, because you are sending your readers to the other sites.

I wouldn't ask myself what AMD is doing because they do remind me the public sector here in Greece. Slow and inefficient in many occasions. With an AMD center button in the front page of a huge site like Anandtech, an AMD center link that doesn't lead to a review of their latest product when someone clicks on it, it is a huge fail for the marketing department. More examples of AMD. They took them 6 months to realize that they had a mistake in their R7 260 page saying that 260 wasn't supporting trueaudio. I had filed 3-4 times the form there in that 6 months period before seeing them fixing it. After over a year they where selling it, you couldn't find 760K in the cpu comparison page. I send them a very angry form, they finally added it. Of course if you go at the Athlon II processor page you are not going to find 700 or 800 series Athlon II. They only list AM3 and FM1. It's a joke.

 

[n0x1ous]

It's not too far from the truth. The 285 came at a bad time for us - Anand's retirement, labor day weekend, and a business trip I'm in the middle of. We've never missed a review like this before (and we won't again), but it was exceptional circumstances.

You'll see it tomorrow.

Still haven't seen this unless I am blind......

 

[MathMan]

LOL. But 20Watts more on hawaii is breaking the bank.:whiste:

I doubt that there are many customers who care about the power consumption per se when buying a card. But that doesn't mean it's not important:
- the GPU maker may have an inferior thermal engineering team that design sucky coolers. In that case, the extra Watts are also extra noise. (100% hypothetical, of course.) People do care about noise.
- the high-end card will reach their limit sooner, reducing absolute performance. People care about that too.
- higher W means higher cost cooler, means less profit. Companies care about profit.