Why Tonga and Antigua fail?

[Carnage1986]

Full unlocked(in another theory it's not full unlocked) Tonga(Antigua XT) released couple days ago and it suprasses GTX 960(actually R9 380 also defeats it in most situations) but there's no huge difference within two GPUs. This is not bad for users, because the card has good value but there's a very sad story for AMD. With 5 billion transistors and 359 mm^2 die size, we are entitled to expect more performance. (GTX 960 has 227 mm^2 die size)

How can we analyze this failure? What's the reason/reasons of this?

There's a theory about it. It's based on this block diagram:

The theory is: They had a problem in manufacturing process (suprisingly huge yield drop or another thing) and they cut down 16 ROPs and 2x64 bit memory controller. As far as i concerned this is a noteworthy claim. Because they know that additional 128 bit bus/16 ROP could dramatically increase the performance and why not they didn't try to design it with that way in the first place? And die size and the transistor count support this idea.

What do you think about this? What's wrong with this GPU?

 

[el etro]

Tonga is just Tahiti with GCN1.2 treatment, nothing more. By a engineering point of view, it was never supposed to show a big efficiency gain over GCN1.0/GCN1.1.

 

[LTC8K6]

GCN1.2 is much faster at some things, though.

It's way faster at video decoding / encoding, for example.

http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/4
http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/3
http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/2

 

[Elixer]

What do you think about this? What's wrong with this GPU?

The biggest thing wrong with the 380X is the price.

The way I look at it, they couldn't make it better than a 390 or 390X, without calling it 395.
We know there was speculation about a bigger bus, but, unless AMD has a few more tricks up its sleeve...this is the end of the line for this family of GPUs.
.

 

[ultima_trev]

As already mentioned, price.

I also am puzzled on why AMD chose to clock Tonga with such a conservative frequency. Both the 380/380X should have been easily, EASILY able to handle +100/500 core/memory.

I wouldn't be surprised if a 380X clocked at 1070/6200 came within spitting distance of the reference R9 290s. If AMD would have released it at those clocks at no more than $210 they would have had a real winner on their hands.

Perhaps they don't want to cannibalize R9 390 sales?

 

[Silverforce11]

Because Tonga is just Tahiti with some extra features that aren't related to performance.

Sure, improve tessellation is nice, but only useful if that is the bottleneck. Which it isn't in most games.

The extra ACEs are nice, going from 2 to 8 with deeper queues, but they are idling in DX11 and doing nothing. So that's added transistors that do not contribute to current performance.

In effect, it's a trade-off design, because its basically the same uarch as in the PS4, they just scaled it up. But without the API to take advantage of the hardware, it's not going to show its full potential. Why did they do it this way?

Cash strapped, cannot do R&D on console uarch and PC uarch, cannot split PC into gaming focused vs HPC focused, so they aimed at one uarch that can do it all, but it's running crippled in DX11 and they know it.

This is my take on it after reading into the history of the uarch designs.

Imagine when GCN of Tonga-like was designed, for the PS4, 2009. AMD gambled that Mantle would pave the way to DX12 or Vulkan much earlier, and not 2016. Right? Imagine, they would have planned the transition to be around 2014 with DX12/Vulkan or next-gen games out by then, GCN would have excelled.

This is what happens in DX12, in an engine that favors NV heavily and is actually sponsored by NV, Unreal Engine 4 from Epic.

380X would be quite high above the 7970.

 

[3DVagabond]

Because Tonga is just Tahiti with some extra features that aren't related to performance.

Sure, improve tessellation is nice, but only useful if that is the bottleneck. Which it isn't in most games.

The extra ACEs are nice, going from 2 to 8 with deeper queues, but they are idling in DX11 and doing nothing. So that's added transistors that do not contribute to current performance.

In effect, it's a trade-off design, because its basically the same uarch as in the PS4, they just scaled it up. But without the API to take advantage of the hardware, it's not going to show its full potential. Why did they do it this way?

Cash strapped, cannot do R&D on console uarch and PC uarch, cannot split PC into gaming focused vs HPC focused, so they aimed at one uarch that can do it all, but it's running crippled in DX11 and they know it.

This is my take on it after reading into the history of the uarch designs.

Imagine when GCN of Tonga-like was designed, for the PS4, 2009. AMD gambled that Mantle would pave the way to DX12 or Vulkan much earlier, and not 2016. Right? Imagine, they would have planned the transition to be around 2014 with DX12/Vulkan or next-gen games out by then, GCN would have excelled.

This is what happens in DX12, in an engine that favors NV heavily and is actually sponsored by NV, Unreal Engine 4 from Epic.

380X would be quite high above the 7970.

You might just want to delete the bold. Because the rest of the post contradicts the opening line. Which is just painful

No doubt that AMD has difficulty moving the market. Why do you suppose we don't have any DX12 games yet? I think it's because AMD isn't paying anyone to make them and nVidia is still paying for DX11. nVidia won't be able to hold back DX12 games forever, but I wouldn't be surprised if they don't have any DX12 performance deficits worked out by the time they do come along.

That said, Tonga is not simply Tahiti+ It's a lot better a chip.

 

[SPBHM]

The theory is: They had a problem in manufacturing process (suprisingly huge yield drop or another thing) and they cut down 16 ROPs and 2x64 bit memory controller. As far as i concerned this is a noteworthy claim. Because they know that additional 128 bit bus/16 ROP could dramatically increase the performance and why not they didn't try to design it with that way in the first place? And die size and the transistor count support this idea.

What do you think about this? What's wrong with this GPU?

considering every single "Tonga" based GPU had 256bits, like the recent 380X and the mobile versions with 2048sps I think it's safe to assume that it actually only had 256bits memory support and the notion that it had 384bits capability is not correct, it makes perfect sense with the memory compression stuff, to reduce the bandwidth and save some $.

also even Tahiti only had 32ROPs

the 285 didn't have a chance because the old AMD cards which were faster were to cheap at the time, also it had less memory than the 280, and the new features added were not very useful for most, it didn't bring anything new in terms of performance, price, power usage...

even now, you look at the 380X and the 7970 from 4 years ago, and it's not so interesting.

 

[NTMBK]

What makes you think Tonga was a failure? That's the chip that won them the iMac contract.

 

[3DVagabond]

What makes you think Tonga was a failure? That's the chip that won them the iMac contract.

Often People will have preconceived ideas of what a product should be. Something ends up not fitting their expectations and it failed in their eyes.

 

[Silverforce11]

For those who doubted that Tonga was designed as a 384bit uarch:

http://www.pcper.com/news/Graphics-...a-384-bit-Memory-Bus-Not-Enabled-Any-Products

It's basically Tahiti with some newer features that may or may not enhance performance in games depending on what the bottleneck is.

Because of its memory compression, they can get away with only enabling 256 bit. This means cheaper PCB to produce and vram is 2/4 not 3/6 which will drive up costs for a product that is meant to be very competitive on perf/$.

 

[LTC8K6]

The company never found the right price/performance combination.

WTF? That's kinda stupid, given the pricing they have now.

So, they could possibly have had an r9-285X based card with 6GB of ram?

They could now have a 380X card like that?

 

[SPBHM]

the fact that they had 0 products with it enabled, making it just a waste of die space made me doubt it really was there, and people were claiming the 380x was a fully enabled GPU.

the full GPU would still have 32ROPs/384bits like Tahiti (which kept 32 enabled on the 7870 LE/256bits cards)

it makes sense with the memory compression thing to have a lower bus considering the benefits to cost, the biggest loss for the 285 was having less memory, not really less bandwidth I think.

WTF? That's kinda stupid, given the pricing they have now.

So, they could possibly have had an r9-285X based card with 6GB of ram?

They could now have a 380X card like that?

they could but the cost increase would probably be a lot more significant than any performance benefit, and it would still be far from a 970 or 390 anyway in performance,

 

[LTC8K6]

6Gb of ram is cheaper than 8gb of ram, though. And it would still be plenty. And crossfire. And with the much faster encoding/decoding of Tonga, and...

This could have been back in 2014, too.

They should have released a line of cards in 2014 based on Tonga, imo.

I think they'd be in a better position now.

 

[Silverforce11]

Why the heck do they need 6gb vram for Tonga class GPU?

I mean 4 is good, 2 is too little.
3 is enough, but 4 looks better..

Add the cheaper PCB and no perf gains for extra bandwidth it doesn't need.. :/

 

[Ansau]

I wouldn't categorize Tonga as a failure, as we should do the same for all GCN architectures released.
First of all, it is not a good idea to measure the success of a chip comparing size and transistor count with the performance, since a part of any chip is not related to the rendering process, but it's part of external features the gpu has. Tonga has a lot of "other" things that if used, the improvement over Tahiti are quite substantial.

Then, the differences of hardware between Tonga XT and Pro are just about 14% of shaders, TMUs and ROPs. But the final rendering involves more parts of the gpus, so it's normal to see improvements of just about 7-10% between both gpus.
Here I think AMD went short with the cut, but this comes from Tahiti, and it should have had 1536 shaders instead of 1792, so both gpus would be more separated (280 vs 280x performance differences came basically because the 280x was clocked much higher than 280).

People underestimate the improvements from Tonga over Tahiti. If we look at the games released this year, the r9 285/380 matches the performance of r9 280x in quite a lot of them (if we discount the frequency differences, fighting 280x when it's clocked at 1000/1500 base is a bit unfair).
But this doesn't finish here, because if games uses the benefits of Tonga, then it smashes Tahiti full chip even with a cut Tonga, like in Witcher 3, Civilization: Beyond Earth, Thief or Project Cars.

 

[LTC8K6]

Why the heck do they need 6gb vram for Tonga class GPU?

I mean 4 is good, 2 is too little.
3 is enough, but 4 looks better..

Add the cheaper PCB and no perf gains for extra bandwidth it doesn't need.. :/

384bit = 3gb or 6gb, I think.

I may be wrong.

 

[Nachtmaer]

I think another reason for Tonga not having a native 256-bit bus (or fully enabled 384-bit) might be that AMD started with Tahiti's layout, added the necessary features and found out that 256-bit was enough to feed Tonga thanks to the color compression. It's possible that spending more time and money on completely redesigning the layout and going through all the testing would end up costing them more than keeping it as is and just disable part of the memory controller (for reasons already mentioned in this thread), even when the chip is slightly bigger than it could've been. Of course this is just an uneducated guess as any other.

 

[Elixer]

The biggest thing wrong with the 380X is the price.

The way I look at it, they couldn't make it better than a 390 or 390X, without calling it 395.
We know there was speculation about a bigger bus, but, unless AMD has a few more tricks up its sleeve...this is the end of the line for this family of GPUs.
.

And that trick up AMD's sleeve would have been the 384 bit bus.
That means, we will never see a fully enabled tonga.
(Well, unless they rehash it again in 2016).

 

[Silverforce11]

I think another reason for Tonga not having a native 256-bit bus (or fully enabled 384-bit) might be that AMD started with Tahiti's layout, added the necessary features and found out that 256-bit was enough to feed Tonga thanks to the color compression. It's possible that spending more time and money on completely redesigning the layout and going through all the testing would end up costing them more than keeping it as is and just disable part of the memory controller (for reasons already mentioned in this thread), even when the chip is slightly bigger than it could've been. Of course this is just an uneducated guess as any other.

That makes a lot of sense. It's the cheapest approach to use what they already have and bolt on more.

That's why Fury is the way it is, take Tonga and bolt on more shaders. It becomes bottlenecked at lower resolutions, front end cannot keep shaders fed. Likewise for DX11, it's running gimped.

 

[RussianSensation]

considering every single "Tonga" based GPU had 256bits, like the recent 380X and the mobile versions with 2048sps I think it's safe to assume that it actually only had 256bits memory support and the notion that it had 384bits capability is not correct

The Chipworks die layout showed 6x64-bit controllers, and finally AMD officially confirmed that Tonga is indeed a 384-bit chip, with 2x64-bit controllers purposely locked out.

"AMD Confirms 384-bit bus available on Tonga, just not enabled on any product, Including 380X. Did not find a perfect perf / $ slot."

http://www.computerbase.de/2015-09/...t/4/#diagramm-rating-2560-1440-hohe-qualitaet

"AMD's Raja Koduri confirmed Tonga's 384-bit bus tonight, and our own Ryan Shrout broke the news on Twitter."
http://www.pcper.com/news/Graphics-...a-384-bit-Memory-Bus-Not-Enabled-Any-Products

That makes a lot of sense. It's the cheapest approach to use what they already have and bolt on more.

Probably lack of resources to design a lean 256-bit Tonga with HDMI 2.0. The card would use even more power with the full 384-bit bus but probably get very little gain from higher memory bandwidth. During the ATI era, this would have never happened since ATI would have had the funds/resources to design a lean 256-bit chip instead of paying higher manufacturing costs of Tonga while fusing/locking parts of those transistors off. Whatever $ you save on the chip design you lose with manufacturing costs over time.

 

[RussianSensation]

Often People will have preconceived ideas of what a product should be. Something ends up not fitting their expectations and it failed in their eyes.

It is a failure from a marketing and execution point of view though.

Look at it, right now R9 380 4GB costs $160, thus offering the best price performance of any GPU in the $140-200 range besides R9 280X. So even with class leading performance and price/performance, it's still failing. Why? Poor marketing and perception of AMD cards being hot and loud.

To this date people are not buying the R9 380 despite it pummeling the competition. Why is that?

AMD also should have offered 285 4GB and the fully unlocked Tonga about 10-12 months ago. Execution failure and not understanding how to gain competitive advantage over the weak 960 2GB. Think about it, AMD killed the 285 2GB on day 1 because both 280 and 280X were better. Look how 2GB cards are doing in modern AAA games?

Had AMD released 285 4GB, it would have been much easier to make a recommendation against the 280 and surely against the crippled x60 2GB. But now, this generation is coming to and end and the competitor has marketed its way to > 80% market share with no good cards below $250.

 

[3DVagabond]

It is a failure from a marketing and execution point of view though.

Look at it, right now R9 380 4GB costs $160, thus offering the best price performance of any GPU in the $140-200 range besides R9 280X. So even with class leading performance and price/performance, it's still failing. Why? Poor marketing and perception of AMD cards being hot and loud.

To this date people are not buying the R9 380 despite it pummeling the competition. Why is that?

AMD also should have offered 285 4GB and the fully unlocked Tonga about 10-12 months ago. Execution failure and not understanding how to gain competitive advantage over the weak 960 2GB. Think about it, AMD killed the 285 2GB on day 1 because both 280 and 280X were better. Look how 2GB cards are doing in modern AAA games?

Had AMD released 285 4GB, it would have been much easier to make a recommendation against the 280 and surely against the crippled x60 2GB. But now, this generation is coming to and end and the competitor has marketed its way to > 80% market share with no good cards below $250.

You know as well as I do that there's no good reason for most of nVidia's cards to outsell AMD. The cards aren't failing. It's AMD's marketing that is. Again, they are a bunch of engineers, not entrepreneurs. They know how to make hardware and write software, but they couldn't sell water in the desert.

 

[AtenRa]

We know Tonga has 384bit Memory (6x 64) over a year now.

Also, 2x 64bit controllers of Tonga are not larger than 14-16mm2 when the entire die size is 366mm2. So its not that much die size to really matter on a very mature 28nm process.

 

[master_shake_]

You know as well as I do that there's no good reason for most of nVidia's cards to outsell AMD. The cards aren't failing. It's AMD's marketing that is. Again, they are a bunch of engineers, not entrepreneurs. They know how to make hardware and write software, but they couldn't sell water in the desert.

amd's marketing is crap that's true.

but i don't think this is helping.

I recommend to not link any GPU review from that site for many years to come until they show objectivity or the site's editor changes.

First, look at the performance difference between R9 280X/R9 290 and GTX960. Now, please take your time to read the conclusion in their GTX960 review on this page. What card is being recommended in this review and which cards are being neglected/ignored for price/performance and VRAM?

Second, take your time to read the conclusion in their R9 380X review. What cards are being recommended in the review now?

Have you noticed something that strikes you as incredibly odd between those reviews yet? :sneaky:

When the author of a professional review site goes out of his way to ignore VRAM limits, ignores price/performance of competing AMD cards in an NV review but then goes out of his way to recommend gamers spend more for superior price/performance cards like the GTX970, do you think this blatant inconsistency is a quality of a professional objective reviewer?

Would a professional objective reviewer remove Dirt Showdown because it highly favours one brand but then use Project CARS in reviews that highly favours the other brand (only to remove it later due to pressure from readers)? Would a professional objective reviewer spend 1.5 years using nV-developed FCAT to demolish HD7000 cards for frame times but ignore poor frame times of Fermi, and early launch Maxwell SLI frame times issues? I'll let you contemplate that.

RS was completely right.