Full AMD Polaris 10 GPU has 2304 Stream Processors

[dogen1]

Ashes of the Singularity.
Hitman.

The 1070 outperforms the 480 in those games by quite a bit. Has anyone done power tests on either game? I don't think I've seen any. The 480 would have to be majorly undercutting the 1070 in power use for you to be right.

 

[96Firebird]

For Polaris to be more efficient than Pascal in Ashes of the Signularity, Polaris would have to be faster than the GTX 1070. I don't see that happening...

http://www.guru3d.com/articles_pages/amd_radeon_r9_rx_480_8gb_review,13.html

Unless you think Polaris uses significantly less power than the GTX 1070 in that game, which I doubt.

Edit - Beat!

 

[3DVagabond]

Not more than Geforce GTX 1070 is. 12.5% extra memory bandwidth provides 6% better performance according to ComputerBase. This would still put it behind the Nano in their charts.

https://www.computerbase.de/2016-06...diagramm-performancerating-speicherbandbreite

There's a huge gap to overcome to even match Geforce GTX 1080, let alone GP102 VGAs rumoured to launch in a few months.

Do we know the efficiency (perf/W) of the GP102? We shouldn't assume it's the same as GP104.

 

[William Gaatjes]

You are not reading slides correctly. A GCN instruction runs for 4 cycles, on each cycle p 16 vector lanes are processed, for a total of 64 lanes. This is the smallest unit of computation on GCN and it's called a wavefront. Wavelets are completely different objects

An NVIDIA SM smallest unit of computation is called warp and it's 32 wide. This makes things a bit better for NVIDIA with code that has flow control but I'd be surprised if it brings to the table a lot of performance. NVIDIA advantage on this front is likely to be quite small, perhaps irrelevant in most cases.

I mixed up the word wavelet with something else. Forget about that.
But indeed, i thought that the simd units in gcn were separate functional units that could run independently. But they are and can not. I assume that for utilization efficiency and die space and clocking, they are coupled with each other. Calling it a simt (single instruction multiple threading ) unit automatically explains it much better. Those 4 (16 wide) vector units are closely connected to each other. If i understand it correctly, they all have to execute the same instruction as well for those 4 cycles in lockstep. After 4 cycles, all 4 vector units have finished the instruction and another instruction can be done. And that is how 64 work items are done. If i am wrong, anybody feel free to correct me.

 

[William Gaatjes]

If i am not mistaken, gcn has the scalar unit for flow control. The 4 (16 wide) vector ALU's do not perform flow control instructions only computations.

 

[ocre]

In the land of team-based discussion and fanboyism everything is black or white.

I think it's perfectly fair, and most likely, that both the process and the architecture have failings and successes. Clearly the GF process is dense, and yielding fairly well. Clearly, TSMC have much, much more experience manufacturing modern GPUs. I very much doubt GF's process for GPUs is to TSMC's level. I doubt anyone in the entire industry is as good at GPUs as TSMC. I also think its perfectly likely that nVidia sent more engineers for more time to sit and work with the TSMC engineers due to their proportionately larger R&D budget. It isn't either or.

There is a lot of back and forth about the foundries. Strangely the sentiment before Polaris launched was completely at odds to what many are saying now. The tsmc 16f was supposedly a disadvantage for nvidia, there was list being posted all over on the many ways the node nvidia was stuck with was inferior. I remember reading some gloating about how Nvidia was stuck helpless as they would never be able to use glofo and that they burnt bridges with their Samsung lawsuit.
The reason I am bringing this up? The irony amazes me.

The chatter now is completely flipped, its now being insisted that the tsmc process is not only an Nvidia advantage but now the nodes alone are responsible for the vastly different polaris vs pascal.

There is no way to ignore that nvidia and amd used different nodes from different foundries. But given that AMD seems to have a lot of full polaris chips, the node is yielded good...or it seems to be.
You cant blame everything on the node, the foundry does not design chips. The node offers what it offers and the foundry has to produce as many useful chips per wafer as possible. The chip design should not be downplayed, it is not some small part of the outcome, its actually on the top here. The node produces chips, their characteristics are a product of their design.

Okay examples: 40nm tsmc.
Go back in time. Amd launched the 5870. Wow, what an amassing improvement, in every single way, over the 55nm 4800series. The tsmc 40nm node was freaking amassing. Right?
Well Nvidia got very different results with their GTX 480. Had they launched on different foundries, I guess we could keep using this argument but they did not. Also, nvidia had to redesign their Fermi on a transistor level and the next launch of a very similarly stuctured gf110 had very different success. The 580 was higher clocked, more active cores, while using less power. The fermi blunder had nvidia totally rethink their strategy. Putting a newly created engineering department together that would focus on nodes and dies shrinks for their architectures.

Then 28nm- it should be obvious that nvidia was getting higher clock speeds while also being more efficient. They were using smaller and less beefed up chips to compete. This was 1000% by design. This was the direction nvidia took. Their are Maxwell chips that can clock 1500-1600 mhz. This was due to chip design.
I am not suggesting AMD was better or worse, just stating that there is clearly a different focus last gen. Nvidia had extracted higher clocks and lower consumption from the 28nm node. Amd had good designs, dont get me wrong. They had different routes, amd used hbm which greatly reduced their power consumption from their wide bus for example. It was cutting edge, the path they took.

I am not talking about just the layout, uarc. The lay out and core design is how the chip handle data on a large scale. Those blocks are made up of billions of transistors, tiny components that make up the chip that the foundry will produce. Any chip can be improved on a transistor level, it is painstaking but can be well worth it.
If we look at intel, their node shrinks have offered such minor improvements. It is baffling when you consider that these tiny inch gains are also the result of architectural changes and the smaller and smaller nodes. It's really clear that intel has gained very small amounts since 32nm.
In that light, the 480 is a major improvement. Massive compared to what little intel gains. Amd has gained a tremendous amount in that light.

The node shrinks today are no where close to what we had in the past. These shrinks are inches and the big gains are only to be had from architectural and chip level design. We seen nvidia talk about this and they have proven this with out a doubt with maxwell. I would expect that nvidia keep focussing on the things that have given them such success in the past. Watching their 1080 reveal, they spent the first segment talking amount the advancement in transistor level voltage spikes,in comparison to maxwell. They talk about their drive for efficiency and speed, this surely would be a focus, it was with maxwell. Its clear to me, they would keep making efforts and strides in the areas that would pay of big time. They were the ones who said years ago, we cant depend on nodes for results.

Surely, nvidia had to take the Samsung 14nm glofo node seriously. They would have to have. I believe they probably did spend a fortune on getting the best efficiency and highest clocks possible. Its naive to think otherwise. Why would they just rely on the node? 14nm offers a lot.

Just looking at the tonga and fury clocks, the 480 is a nice step up with a good reduction in power consumption.
Looking at maxwell boost clocks vs gcn last gen, it seems that nvidia has found ways to achieve higher clocks when on the same node. Nvidia also claims that pascal was engineered for speed, I think that focus had something to do with the result. The gap from 480 to 1070 in mhz, it seems this gap started last gen when both were on the same node

 

[renderstate]

There is a lot of back and forth about the foundries. Strangely the sentiment before Polaris launched was completely at odds to what many are saying now. The tsmc 16f was supposedly a disadvantage for nvidia, there was list being posted all over on the many ways the node nvidia was stuck with was inferior. I remember reading some gloating about how Nvidia was stuck helpless as they would never be able to use glofo and that they burnt bridges with their Samsung lawsuit.

The reason I am bringing this up? The irony amazes me.



The chatter now is completely flipped, its now being insisted that the tsmc process is not only an Nvidia advantage but now the nodes alone are responsible for the vastly different polaris vs pascal.



There is no way to ignore that nvidia and amd used different nodes from different foundries. But given that AMD seems to have a lot of full polaris chips, the node is yielded good...or it seems to be.

You cant blame everything on the node, the foundry does not design chips. The node offers what it offers and the foundry has to produce as many useful chips per wafer as possible. The chip design should not be downplayed, it is not some small part of the outcome, its actually on the top here. The node produces chips, their characteristics are a product of their design.



Okay examples: 40nm tsmc.

Go back in time. Amd launched the 5870. Wow, what an amassing improvement, in every single way, over the 55nm 4800series. The tsmc 40nm node was freaking amassing. Right?

Well Nvidia got very different results with their GTX 480. Had they launched on different foundries, I guess we could keep using this argument but they did not. Also, nvidia had to redesign their Fermi on a transistor level and the next launch of a very similarly stuctured gf110 had very different success. The 580 was higher clocked, more active cores, while using less power. The fermi blunder had nvidia totally rethink their strategy. Putting a newly created engineering department together that would focus on nodes and dies shrinks for their architectures.



Then 28nm- it should be obvious that nvidia was getting higher clock speeds while also being more efficient. They were using smaller and less beefed up chips to compete. This was 1000% by design. This was the direction nvidia took. Their are Maxwell chips that can clock 1500-1600 mhz. This was due to chip design.

I am not suggesting AMD was better or worse, just stating that there is clearly a different focus last gen. Nvidia had extracted higher clocks and lower consumption from the 28nm node. Amd had good designs, dont get me wrong. They had different routes, amd used hbm which greatly reduced their power consumption from their wide bus for example. It was cutting edge, the path they took.



I am not talking about just the layout, uarc. The lay out and core design is how the chip handle data on a large scale. Those blocks are made up of billions of transistors, tiny components that make up the chip that the foundry will produce. Any chip can be improved on a transistor level, it is painstaking but can be well worth it.

If we look at intel, their node shrinks have offered such minor improvements. It is baffling when you consider that these tiny inch gains are also the result of architectural changes and the smaller and smaller nodes. It's really clear that intel has gained very small amounts since 32nm.

In that light, the 480 is a major improvement. Massive compared to what little intel gains. Amd has gained a tremendous amount in that light.



The node shrinks today are no where close to what we had in the past. These shrinks are inches and the big gains are only to be had from architectural and chip level design. We seen nvidia talk about this and they have proven this with out a doubt with maxwell. I would expect that nvidia keep focussing on the things that have given them such success in the past. Watching their 1080 reveal, they spent the first segment talking amount the advancement in transistor level voltage spikes,in comparison to maxwell. They talk about their drive for efficiency and speed, this surely would be a focus, it was with maxwell. Its clear to me, they would keep making efforts and strides in the areas that would pay of big time. They were the ones who said years ago, we cant depend on nodes for results.



Surely, nvidia had to take the Samsung 14nm glofo node seriously. They would have to have. I believe they probably did spend a fortune on getting the best efficiency and highest clocks possible. Its naive to think otherwise. Why would they just rely on the node? 14nm offers a lot.



Just looking at the tonga and fury clocks, the 480 is a nice step up with a good reduction in power consumption.

Looking at maxwell boost clocks vs gcn last gen, it seems that nvidia has found ways to achieve higher clocks when on the same node. Nvidia also claims that pascal was engineered for speed, I think that focus had something to do with the result. The gap from 480 to 1070 in mhz, it seems this gap started last gen when both were on the same node

Best post I read on this forum in a long time!

 

[Sweepr]

 

[Headfoot]

There is a lot of back and forth about the foundries. Strangely the sentiment before Polaris launched was completely at odds to what many are saying now. The tsmc 16f was supposedly a disadvantage for nvidia, there was list being posted all over on the many ways the node nvidia was stuck with was inferior. I remember reading some gloating about how Nvidia was stuck helpless as they would never be able to use glofo and that they burnt bridges with their Samsung lawsuit.
The reason I am bringing this up? The irony amazes me.

The chatter now is completely flipped, its now being insisted that the tsmc process is not only an Nvidia advantage but now the nodes alone are responsible for the vastly different polaris vs pascal.

There is no way to ignore that nvidia and amd used different nodes from different foundries. But given that AMD seems to have a lot of full polaris chips, the node is yielded good...or it seems to be.
You cant blame everything on the node, the foundry does not design chips. The node offers what it offers and the foundry has to produce as many useful chips per wafer as possible. The chip design should not be downplayed, it is not some small part of the outcome, its actually on the top here. The node produces chips, their characteristics are a product of their design.

Okay examples: 40nm tsmc.
Go back in time. Amd launched the 5870. Wow, what an amassing improvement, in every single way, over the 55nm 4800series. The tsmc 40nm node was freaking amassing. Right?
Well Nvidia got very different results with their GTX 480. Had they launched on different foundries, I guess we could keep using this argument but they did not. Also, nvidia had to redesign their Fermi on a transistor level and the next launch of a very similarly stuctured gf110 had very different success. The 580 was higher clocked, more active cores, while using less power. The fermi blunder had nvidia totally rethink their strategy. Putting a newly created engineering department together that would focus on nodes and dies shrinks for their architectures.

Then 28nm- it should be obvious that nvidia was getting higher clock speeds while also being more efficient. They were using smaller and less beefed up chips to compete. This was 1000% by design. This was the direction nvidia took. Their are Maxwell chips that can clock 1500-1600 mhz. This was due to chip design.
I am not suggesting AMD was better or worse, just stating that there is clearly a different focus last gen. Nvidia had extracted higher clocks and lower consumption from the 28nm node. Amd had good designs, dont get me wrong. They had different routes, amd used hbm which greatly reduced their power consumption from their wide bus for example. It was cutting edge, the path they took.

I am not talking about just the layout, uarc. The lay out and core design is how the chip handle data on a large scale. Those blocks are made up of billions of transistors, tiny components that make up the chip that the foundry will produce. Any chip can be improved on a transistor level, it is painstaking but can be well worth it.
If we look at intel, their node shrinks have offered such minor improvements. It is baffling when you consider that these tiny inch gains are also the result of architectural changes and the smaller and smaller nodes. It's really clear that intel has gained very small amounts since 32nm.
In that light, the 480 is a major improvement. Massive compared to what little intel gains. Amd has gained a tremendous amount in that light.

The node shrinks today are no where close to what we had in the past. These shrinks are inches and the big gains are only to be had from architectural and chip level design. We seen nvidia talk about this and they have proven this with out a doubt with maxwell. I would expect that nvidia keep focussing on the things that have given them such success in the past. Watching their 1080 reveal, they spent the first segment talking amount the advancement in transistor level voltage spikes,in comparison to maxwell. They talk about their drive for efficiency and speed, this surely would be a focus, it was with maxwell. Its clear to me, they would keep making efforts and strides in the areas that would pay of big time. They were the ones who said years ago, we cant depend on nodes for results.

Surely, nvidia had to take the Samsung 14nm glofo node seriously. They would have to have. I believe they probably did spend a fortune on getting the best efficiency and highest clocks possible. Its naive to think otherwise. Why would they just rely on the node? 14nm offers a lot.

Just looking at the tonga and fury clocks, the 480 is a nice step up with a good reduction in power consumption.
Looking at maxwell boost clocks vs gcn last gen, it seems that nvidia has found ways to achieve higher clocks when on the same node. Nvidia also claims that pascal was engineered for speed, I think that focus had something to do with the result. The gap from 480 to 1070 in mhz, it seems this gap started last gen when both were on the same node

Good post. The process tide raises all boats -- but I have to agree it seems like process doesn't get you anywhere near as far as it used to. Architecture and micro optimization is taking a larger and larger role. The pre-release hype about Glofo magically overtaking TMSC in process was ridiculous. They'd only ever dropped the ball in the past so expecting a miracle was astounding wishful thinking. Even with the Samsung process, that would lead me to believe they'd be good at making lots of low power small dies since the process was used primarily to make phone SOCs. And here we see a small die, apparently yield/price optimized process which is exactly in line with what you'd expect from a Samsung derived phone SOC process. But the people building hype train are all about emotion and not evidence.

 

[krumme]

I think its a miracle gf can turn out 14nm finfts gpu at about same time as tsmc going by their prior less than stellar record. Huge surprise to me. And I don't really care if it's 10% better worse whatever, just as long as they can crank out those gpu in huge numbers.

That said I haven't noticed any body hyping gf process. We already know the basic differences with the apple a9.

Who are/were those people hyping it? Beats me. Seems like a strawman to me.

Probably you can find one or two who thinks 14 is better than 16 but that is not "the general sentiment" of this board.

 

[Dresdenboy]

Cross posted from the Zen thread:

Why is it the process's fault? Why is it never the fault of the company that actually designed the product?

Anyway, Polaris got a big jump in perf/watt over Tonga, no doubt process related. But AMD's architecture was so behind NVIDIA's to begin with that expecting to hit the perf/watt levels that NVIDIA did with Pascal on 16FF+ (which is a superior process, but not by THAT much) was just unrealistic wishful thinking.

I think, you're right on the money here. Polaris seems to be an evolutionary step in the GCN family. Given AMD's R&D situation, FinFET costs, and cost targets (mainstream markets), which seems to have driven them to a high volume die strategy both for GPUs and CPUs, it would be reasonable to expect them to also keep design costs for lower margin products in check. This means: improved GCN and not a completely reworked GCN (e.g. faster clockable pipeline -> lower FO4/stage) for Polaris.

So what AMD could get out of Polaris was limited from day one. With a smaller die than Nvidia's DT Pascal GPUs (to keep costs down), the only option to get more performance out of it is ... upping the clock frequency into the less efficient range.

Here is my view on this:

Well, problem is, there are a lot of games that are not DX12, and a lot of other cards in the line-up for both companies. Anyway, 1060 competes with 480, not 470, and indications are very strong that perf/watt will strongly favor 1060. Your scenario, is a cherry picked, best case scenario for AMD. Maybe we should examine all cards in each lineup, in a wide variety of games.

I didn't check, but if DX12 is easier to use, offers more performance, is in more widespread use (Win10 adoption thanks to the upgrade offer and XBox One, etc.), DX11 (where AMD's drivers lack anyway, regardless if it's a FinFET GPU or not) will fade away over time. This might also be an effect of game project schedules.

This is getting off topic in the cpu forums, but nobody argues that the 480 offers a lot of performance for the price. Again though, it was overhyped by AMD, and especially by some in these forums.

It clearly failed to meet the performance per watt that people expected, and is basically over-volted in order to reach acceptable frequencies, with rapidly increasing power use with minimal overclocking. AMD is covering their butts now by saying "Oh, we will get 2.8x performance per watt with Polaris, but not *this* polaris. Wait for the 470." So there is a lot of criticism, rightly so IMO.

I think the whole back and forth in opinions here and elsewhere is a to be expected result of people with a more or less complete picture of semiconductors discussing a complex topic with only part of needed information at hand.

E.g. some information that might have gone missing might be this one:

Many people seem to be so easily triggered, that they begin to run before knowing the direction... This is true both for sceptics and "hypesters".

 

[Dresdenboy]

Cross quoted from there:

Wow, I haven't actually seen that slide. I wonder why people reference the 2.8x to the 480?

Because AMD puts out confusing slides? Slides that have to be deciphered and footnoted, and interpreted.

Because of this.

Top left "RX 480 Built on 14nm FinFET"
Middle of the screen "UP TO 2.8X" PPW with a tiny footnote sign.

If they want to play the footnote game... they get the PR and credibility to match.

Personally I see nothing confusing about this: they lied. As I may have said before on the VC&G forum, from now on AMD better back any claim they make with a live demo: zero credibility policy from me.

If I see this right, at that time, RX 470 and RX 460 weren't announced. So they couldn't speak about them. I'm also sure, that at a lower power tune setting, RX 480 might get closer to that "up to" point, while the RX 470 is better configured to reach it.

 

[krumme]

Amd is not only starved for ressources on designing a proper high freq design but eg also tons of marketing project.

But they brought themselves in the situation by not cutting product portfolio more. Lisa haven't been strong enough here. A part of the problem is many of those project started years ago but continuing crap product is just the fallacy of sunk cost.

Lots of meaningless learning product like all server product with a57 as the most prime example and also the huge die fiji vega hbm2 nonsense. Product designed by engineers without any business understanding. Sorry.

 

[Arachnotronic]

Amd is not only starved for ressources on designing a proper high freq design but eg also tons of marketing project.

But they brought themselves in the situation by not cutting product portfolio more. Lisa haven't been strong enough here. A part of the problem is many of those project started years ago but continuing crap product is just the fallacy of sunk cost.

Lots of meaningless learning product like all server product with a57 as the most prime example and also the huge die fiji vega hbm2 nonsense. Product designed by engineers without any business understanding. Sorry.

This is a really good post. :thumbsup:

 

[dark zero]

AMD failed to deliver what the market wanted, despite going years in advance. Sadly this paid off hard for them.

Well once they goes away, the 1st error nVIDIA does, kills for real the GPU market and won't resurrect for some years. Similar thing with x86 CPU. That's why Microsoft is trying to going to ARM, since that market can still be saved and still has more variety (and considering that nVIDIA and Intel left that market...)

Also... interesting to say, nVIDIA didn't managed to shrink perfectly their GPU uarch to mobile devices. Wondering if Pascal can do it, but considering that they threw away all the ARM project they had....

 

[JDG1980]

But they brought themselves in the situation by not cutting product portfolio more. Lisa haven't been strong enough here. A part of the problem is many of those project started years ago but continuing crap product is just the fallacy of sunk cost.

Lots of meaningless learning product like all server product with a57 as the most prime example and also the huge die fiji vega hbm2 nonsense. Product designed by engineers without any business understanding. Sorry.

I agree that there were definitely some white-elephant projects. In general, none of AMD's ARM-based projects make much sense from what I can tell. The purchase of SeaMicro was a complete waste, Seattle was a dud, and K12 should have been cut long ago. There have been multiple attempts, from many different companies, to get ARM into the server market to compete with Intel; so far all have largely failed. Apple might be able to do it if they tried, but probably no one else has the resources.

On the other hand, I don't think AMD can simply give up on the high end of GPUs. Profit margins are higher (especially if AMD can make some inroads into the lucrative workstation and HPC markets), and the R&D costs aren't all that massive because the same architecture is usually used across a full line of GPUs. It does appear in retrospect that AMD jumped the gun on HBM; they would have been better off putting more effort into memory compression technology, and using GDDR5X as a stopgap. In the long run, though, HBM2 is absolutely essential to fulfilling the APU dream.

 

[krumme]

Amd have tons of valuable ip. Thousands of extremeskill people with experience. Massive methology and process platforms.

Amd today is in a far better position than 12 years ago. Yet they don't turn a profit and beeing tied to that concrete block of gf via wsa is only one factor here.

In the grander sceme they are a small player. But so is nv. Nv have adapted technology at a very conservative way. Perhaps with pascal they should have gone for some more radical changes because of dx12 hitting hard now but even when perhaps beeing a bit off in a nearly 20 years history they are always incredible spot on. Looking at amd it's nearly opposite. Always 2-3 years to early with tech. Hbm2 could easily be introduced with Vega as a first wave if they absolutely needed something for that market for whatever reason that is still a mystery.
From a business perspective they try to enter professional market where they have no leverage.
They are smearing a thin layer of r&d over way to big portfolio.
Why on earth do they continue like they did 4 years ago with more or less same product portfolio?
A guess: Rory failed to axe the heads preventing progress here and Lisa is to much intelligent tech head to do so.

 

[ocre]

That said I haven't noticed any body hyping gf process. We already know the basic differences with the apple a9.

Who are/were those people hyping it? Beats me. Seems like a strawman to me.

Specifically.. forum chatter, i spoke of. I mostly read in my time between these days. forum chatter I speak of is not a specific jab at anyone or ATF at all. It is not directed towards this one place or threads here. There was a lot of chatter before polaris and pascal launched. These things were wrote and there was belief that nvidia would be at a disadvantage.

I wont deteriorate the conversation by re posting, totally not the direction i want to go in... it is not any attempt to gloat or "strawman".

Perhaps read my post, its just my thoughts and POV. A contribution...an attempt on sharing thoughts and ideas on technology. no need to go backwards now

Amd have tons of valuable ip. Thousands of extremeskill people with experience. Massive methology and process platforms.

Amd today is in a far better position than 12 years ago. Yet they don't turn a profit and beeing tied to that concrete block of gf via wsa is only one factor here.

In the grander sceme they are a small player. But so is nv. Nv have adapted technology at a very conservative way. Perhaps with pascal they should have gone for some more radical changes because of dx12 hitting hard now but even when perhaps beeing a bit off in a nearly 20 years history they are always incredible spot on. Looking at amd it's nearly opposite. Always 2-3 years to early with tech. Hbm2 could easily be introduced with Vega as a first wave if they absolutely needed something for that market for whatever reason that is still a mystery.
From a business perspective they try to enter professional market where they have no leverage.
They are smearing a thin layer of r&d over way to big portfolio.
Why on earth do they continue like they did 4 years ago with more or less same product portfolio?
A guess: Rory failed to axe the heads preventing progress here and Lisa is to much intelligent tech head to do so.

I agree, AMD IP is in a league of its own. Talent and Engineers are as well

As for HBM.
HBM cut down consumption significantly which then allowed more compute units moving up from 2816sp to a whopping 4000sp. This was their path to higher performance, from the 390x to fury X putting AMD way up there in the high end. It allowed AMD to compete in the high end.

What do you think AMD should have done different? I mean, in your opinion. Please dont take this as an attack and I absolutely am not trying to argue.
Honestly, hoping to have deeper technology conversations on ideas.
I am sure you see what the plan was. AMDs plan was to use HBM and leverage it to their advantage as their path to compete against maxwell in the high end.

What are your thoughts/opinions? Alternate paths they could have taken? Should they have ceded the high end? Or do you have ideas on different routes they could have taken?

I seriously would like to have a conversion on ideas here

 

[BasementWayne]

What are your thoughts/opinions? Alternate paths they could have taken? Should they have ceded the high end? Or do you have ideas on different routes they could have taken?

I seriously would like to have a conversion on ideas here

You are asking the right questions. So while stuck at 28nm, both AMD and Nvidia needed more memory bandwidth and lower power usage to make more powerful chips.

AMD chose to design an exotic new memory type. Wikipedia says the HBM project started in 2008. AMD obviously felt that 4GB would be enough at the time, but the project drug on for seven years before they could ship it. When it did ship, the memory limitations crippled the card. Of course if AMD could have shipped it in 2012, they have a huge advantage over Nvidia.

Meanwhile, Nvidia adopted a different approach. They added more cache, optimized pathways, and added compression to effectively increase memory bandwidth. Nvidia also invested much more in process technology to lower power usage.

CEOs make dozens of these types of decisions every year. Typically it can be 3-5 years down the road before the payoff is determined. It's easy in retrospect to say AMD had no competency in designing memory. But it takes a great CEO to make those decisions years in advance.

 

[Bacon1]

4GB hasn't been a limitation for Fury cards. Even cranking up SSAA to 4x @ 4k it held its own against the 12 GB Titan X - http://www.hardocp.com/article/2016/02/29/rise_tomb_raider_graphics_features_performance/12

 

[dark zero]

You are asking the right questions. So while stuck at 28nm, both AMD and Nvidia needed more memory bandwidth and lower power usage to make more powerful chips.

AMD chose to design an exotic new memory type. Wikipedia says the HBM project started in 2008. AMD obviously felt that 4GB would be enough at the time, but the project drug on for seven years before they could ship it. When it did ship, the memory limitations crippled the card. Of course if AMD could have shipped it in 2012, they have a huge advantage over Nvidia.

Meanwhile, Nvidia adopted a different approach. They added more cache, optimized pathways, and added compression to effectively increase memory bandwidth. Nvidia also invested much more in process technology to lower power usage.

CEOs make dozens of these types of decisions every year. Typically it can be 3-5 years down the road before the payoff is determined. It's easy in retrospect to say AMD had no competency in designing memory. But it takes a great CEO to make those decisions years in advance.

As far I know it was Intel who helped nVIDIA with the cache the tech with their cross license agreements.

And If I am not wrong, Intel did it because they wanted to buy nVIDIA in their moment. However nVIDIA rejected them. And now Intel is trying to take nVIDIA down due nVILINK and the alliance with IBM Power

And returning to the topic.. the only way we will see Polaris with HBM will be on the Apple Mac Pro.

 

[Headfoot]

As far I know it was Intel who helped nVIDIA with the cache the tech with their cross license agreements.

And If I am not wrong, Intel did it because they wanted to buy nVIDIA in their moment. However nVIDIA rejected them. And now Intel is trying to take nVIDIA down due nVILINK and the alliance with IBM Power

And returning to the topic.. the only way we will see Polaris with HBM will be on the Apple Mac Pro.

Source? This seems pretty "rumor mill"

 

[Sweepr]

Call of Duty Black Ops 3 - All Fury cards slower than R9 390X
https://www.computerbase.de/2016-06...diagramm-call-of-duty-black-ops-iii-1920-1080

Doom OpenGL @ 4K Nightmare - Fury X slower than R9 390
http://www.pcgamer.com/dooms-nightmare-graphics-image-quality-compared-and-benchmarked

Mirror's Edge Catalyst with Hyper textures:

In addition, Hyper also offers higher resolution textures. However, these are rarely actually visible better. Measurable these, however, are consistently much larger. It requires at the "Hyper-textures" in Full HD at least a graphics card with six gigabytes and from 2,560 × 1,440 then must it be eight gigabytes.

... The Radeon R9 390 easy to lose 25 percent higher and the Radeon R9 Fury X 27 percent. The latter has the problem that the four gigabyte memory is no longer sufficient. The FPS-loss thereby is indeed low, but the game not running smoothly.

www.computerbase.de/2016-06/mirrors...#abschnitt_die_pcversion_und_die_hyperdetails

Rainbow Six Siege @ 4K Ultra - Fury X slower than R9 390

https://nl.hardware.info/reviews/64...met-22-gpus-testresultaten-ultra-hdn3840x2160

4GB already limits Fury cards today and it will only get worse from now on. One might make an exception for RX 480 / 470 and maybe GTX 1060 (3GB) due to their price, but 4GB for a high-end card is unacceptable from now on.

 

[krumme]

Specifically.. forum chatter, i spoke of. I mostly read in my time between these days. forum chatter I speak of is not a specific jab at anyone or ATF at all. It is not directed towards this one place or threads here. There was a lot of chatter before polaris and pascal launched. These things were wrote and there was belief that nvidia would be at a disadvantage.

I wont deteriorate the conversation by re posting, totally not the direction i want to go in... it is not any attempt to gloat or "strawman".

Perhaps read my post, its just my thoughts and POV. A contribution...an attempt on sharing thoughts and ideas on technology. no need to go backwards now



I agree, AMD IP is in a league of its own. Talent and Engineers are as well

As for HBM.
HBM cut down consumption significantly which then allowed more compute units moving up from 2816sp to a whopping 4000sp. This was their path to higher performance, from the 390x to fury X putting AMD way up there in the high end. It allowed AMD to compete in the high end.

What do you think AMD should have done different? I mean, in your opinion. Please dont take this as an attack and I absolutely am not trying to argue.
Honestly, hoping to have deeper technology conversations on ideas.
I am sure you see what the plan was. AMDs plan was to use HBM and leverage it to their advantage as their path to compete against maxwell in the high end.

What are your thoughts/opinions? Alternate paths they could have taken? Should they have ceded the high end? Or do you have ideas on different routes they could have taken?

I seriously would like to have a conversion on ideas here

As other poster said ita far more easy to look at things retrospective.
I would have avoided fiji vega and hbm2 and all the cpu server stuff plus all the professional gpu stuff because its just draining the company of cash and ressources. That would be a cut that would liberate for better sale marketing and eg critical paths for high freq design both for gpu and zen.

Now thats looking back. Looking forward is more difficul as we dont have the info say Lisa have. Foremost i would not bring a product to market where i wouldnt be number one.
The reason is simply thats where the high margins is. I would select for short term profit as profit builds and shapes future. Sorry but that is the tendency today in most markets.

My take on future is:
Get new owners.
Change board.
Get out of gf deal. Play hardball and take the risk.
If zen is not well because eg gf process is not 100 up to it. Wait.
Dump all cpu crap that dont really sell. Take the loss now.
If zen is not distint advantage at desktop or whatever market dont market it there at all.
Drop all b2b gpu business.
Go 100% where amd is strongest. Semicustom on consumer side. Entertainment.

Whats your proposal for the future?

 

[Thala]

4GB already limits Fury cards today and it will only get worse from now on. One might make an exception for RX 480 / 470 and maybe GTX 1060 (3GB) due to their price, but 4GB for a high-end card is unacceptable from now on.

Strongly disagree. All you examples show marginal image IQ improvements at resolutions with inacceptable framerates. So no Fury is not limited by 4GByte if you chose any reasonable settings.
Your examples are artificial and thus purely hypothetical.