[Ars] AMD confirms high-end Polaris GPU will be released in 2016

[R0H1T]

It would be pathetic because it would mean AMD did nothing to improve architectural efficiency beyond what the FINFET process node shrink brings. It would mean AMD have done nothing from 2012 to 2016 to improve perf/sp perf/watt, perf/sq mm purely from an architectural standpoint. It would also mean GP106 will murder it. So if AMD are so dumb as to not anticipate Nvidia improving Maxwell further and try to compete with that then they might as well shut down their business.

Indeed the Polaris 11 has to be close to Fury non X, if not better, along with Polaris 10 being closer to the 380x. The GCN4 uarch improvements, 14nm FinFet & HBM2 or 20nm GDDR5(x?) respectively will give them some neat boost, notwithstanding the DX12 gains.

 

[JDG1980]

You're not being accurate or realistic. After I just showed you that netting an 80% increase in transistor density, and doubling perf/mm2 would massively outstrip what that last node leap accomplished, you answer in disbelief. AMD has said time and again 2x perf/w improvement. Look at Pitcairn; double the performance and keep the power envelope the same. That gets you 390 performance at Pitcairn power levels, 2x perf/w.

I'm going by what was reported in the Polaris demo shown to the press. The Polaris 10 card was put up against GTX 950. Now, the 950 is about equivalent in performance to AMD's R9 270X. (TechPowerUp's newest performance summary charts have them neck-and-neck at 1080p.) The reference R9 270X, when tested by TPU, averaged 111W during gaming; there is no reference GTX 950, but the ones TPU tested average about 95W. Polaris 10, in a gaming scenario, was pulling 86W at the wall, compared to 140W for GTX 950. That would mean the Polaris 10 card couldn't be pulling much more than 40W. So we're talking close to three times the power efficiency of 28nm GCN - unless you count the Fury Nano, which they might be.

Regarding the last node, there were a lot of issues with Tahiti which make it a poor comparison. It was the first chip not only of a new node, but a new architecture; this meant very conservative design. Pitcairn came only a couple months later, and AMD had already made the memory controllers far smaller and more efficient and increased the transistor density from 12.25 million to 13.21 million transistors per square millimeter.

Also, the move to the 14nm Samsung-based process is a bit more than a full node shrink. 40nm->28nm reduced the gate pitch by about 30% and the metal pitch by 25%. 28nm->14nm reduces the gate pitch by about 34% and the metal pitch by about 29%. Not a huge difference but every little bit adds up. We have seen with the A9 chips that the Samsung process is denser than TSMC.

40-50% you say? Can I quote the bolded part in my signature so we can revisit all these crazy claims when reality sets in?

I could be wrong. I'd be reluctant to bet more than ten bucks on any of the speculations I'm making here. AMD has dropped the ball before, and they could disappoint again. But they really do seem to be going all-in on FinFET. And they have a pretty good history of pulling off node shrinks effectively and with few hitches.

Let me ask you: what effect do you think FinFET will have on GPU clock speeds?

 

[MrTeal]

Indeed the Polaris 11 has to be close to Fury non X, if not better, along with Polaris 10 being closer to the 380x. The GCN4 uarch improvements, 14nm FinFet & HBM2 or 20nm GDDR5(x?) respectively will give them some neat boost, notwithstanding the DX12 gains.

The expectations for Polaris are pretty out of line with what what we've seen. I would love to see P10 hit that kind of performance, especially at the 75W figure that's been tossed around. I'll be looking for a new laptop in the next year, 380X performance in that envelope would be great.

I think it's telling though that AMD has thus far compared P10 to a 950, and has said they're competing in that space. A 380X is ~50% faster than a 950 at 1080p, and while there might be performance on tap to get to or pull ahead of a 960, the 380X is a much bigger target. If Ryan Smith was correct and P10 is smaller than 120mm², hitting 380X levels would be more than 3x increase in perf/mm².

 

[AtenRa]

I would say Polaris 11 could come close to 20-30% faster than Hawaii.

Im now starting to believe that Polaris 11 could replace Fiji and Hawaii at the $300-600 price range.

More like,
$550-650 for full Polaris 11 , it will replace Fiji (Fury X ) , same or little better performance at DX-12 for half the TDP.
$450-550 Polaris 11 (less Shaders) to replace Fury and NANO ,

We could also see a $350-400 SKU with less memory (4-6GB HBM2), less bandwidth and cut off Shaders. That could replace R9 390/X , same or better performance at half the TDP. This could come later in Q3-4.

Polaris 10 could replace R9 380x ($250-300) down to R7 370. Almost the same performance or better in DX-12 than R9 380X at half the TDP.

 

[R0H1T]

The expectations for Polaris are pretty out of line with what what we've seen. I would love to see P10 hit that kind of performance, especially at the 75W figure that's been tossed around. I'll be looking for a new laptop in the next year, 380X performance in that envelope would be great.

I'm not expecting it to be 75W, if it is then it'd be something like the Nano i.e. power starved &/or geared towards SFF. If anything, looking at the recent past where AMD have generously overvolted their GPU's, the TDP should be closer to 100W & that's where I'd say the 380x perf level (prediction) comes from. So it could very well be two different cards, from the same GPU, akin to the Fury X & Nano.

I think it's telling though that AMD has thus far compared P10 to a 950, and has said they're competing in that space. A 380X is ~50% faster than a 950 at 1080p, and while there might be performance on tap to get to or pull ahead of a 960, the 380X is a much bigger target. If Ryan Smith was correct and P10 is smaller than 120mm², hitting 380X levels would be more than 3x increase in perf/mm².

It wouldn't have to come to that, since we're basing this assumption just on the 75W TDP, even a 20~30W headroom would give it better performance albeit at ~2.5x level of (slightly) lower efficiency.

 

[The Alias]

I would say Polaris 11 could come close to 20-30% faster than Hawaii.

Im now starting to believe that Polaris 11 could replace Fiji and Hawaii at the $300-600 price range.

More like,
$550-650 for full Polaris 11 , it will replace Fiji (Fury X ) , same or little better performance at DX-12 for half the TDP.
$450-550 Polaris 11 (less Shaders) to replace Fury and NANO ,

We could also see a $350-400 SKU with less memory (4-6GB HBM2), less bandwidth and cut off Shaders. That could replace R9 390/X , same or better performance at half the TDP. This could come later in Q3-4.

Polaris 10 could replace R9 380x ($250-300) down to R7 370. Almost the same performance or better in DX-12 than R9 380X at half the TDP.

Is there another chip that I'm missing? If not this is extremely disappointing considering the performance bar wouldn't be moving up at all

 

[JDG1980]

If Ryan Smith was correct and P10 is smaller than 120mm², hitting 380X levels would be more than 3x increase in perf/mm².

Don't forget that Tonga is one of AMD's most inefficient chips in terms of perf/mm^2. Pitcairn and Hawaii are far better designs.

Also, the Polaris 10 card matched against GTX 950 in the demo was running at very low wattage - probably about 40W, considering the whole system was only pulling 86W at the wall. Even if the card comes without a PCIe connector, you can put a power limit of 75W on the desktop version with no problem. That means much higher clocks. The laptop version may only be on par with a GTX 950 in order to save battery life and reduce chassis heat.

 

[raghu78]

AMD did show up for the fight, they just lost.

For the Maxwell generation AMD did not show up for the fight. Its well reflected in their market share loss. Never before in the long history of ATI or AMD have we seen such low market share (20%) as they have now. Its very evident AMD did not have any new products except Fury X (which disappointed against GM200 because AMD did not put in any major architectural improvements). Nvidia laughed all the way to the bank once AMD failed to show up for the fight.

 

[ShintaiDK]

I would say Polaris 11 could come close to 20-30% faster than Hawaii.

Im now starting to believe that Polaris 11 could replace Fiji and Hawaii at the $300-600 price range.

I doubt that. The chip is simply not big enough. I would guess on a 3072SP part. And I doubt GCN 1.3 offers any meaningful benefits.

 

[maddie]

Two very recent statements from AMD.

Koduri:
Yes. We have two versions of these FinFET GPUs. Both are extremely power efficient. This is Polaris 10 and that’s Polaris 11.

RoyTaylor:
f you look at the minimum spec for either the Oculus or the HTC, and then you look at how many units of the minimum spec have been sold since their launch, so I'm talking about the Radeon 290 or GeForce GTX 970, according to Jon Peddie Research, the total install base of those parts or better is 7.5 million units. So we're going to have to make it possible to run good quality VR at a much lower price. And I'm confident with Polaris we're going to have a big impact to help that.

We can safely say that polaris 10 as demoed will not be the good quality VR GPU. If you interpret the 2 versions statement to mean 2 GPU only, then the affordable VR GPU [290/970 class] must be Polaris 11. Cheaper than R9 390.

This leaves nothing higher in the performance and price range.
Is this realistic?

If however, we interpret the 2 versions as different designs, one with GDDR5 and the other with HBM, then we can start to see a much wider product offering. Polaris 10 (small) as the demoed GPU, Polaris 10 (large) as the affordable VR, possibly 390 performance at 380 prices.

Then we move to Polaris 11 and HBM. This is the one for which no info is known, probably very deliberately.

 

[Tuna-Fish]

I doubt that. The chip is simply not big enough. I would guess on a 3072SP part.

3072 sounds, if anything, too big to me if it's not a HBM design. PHYs don't shrink, and Pitcairn already used the denser, slower GDDR interface. Even with 2:1 scaling for logic, 3072 would get really tight.

However, you have to remember clocks too. Transistor performance increases significantly more between 28nm -> 14nm FF than it did between 40 -> 28. I expect that we will see substantial clock speed increases on the desktop parts that get nice fat TDPs.

And I doubt GCN 1.3 offers any meaningful benefits.

You know, by their own words they only changed every relevant piece of the design. Clearly there can be no benefits...

 

[Glo.]

I will again pop back the post on SemiWiki: https://www.semiwiki.com/forum/content/3884-who-will-lead-10nm.html

GloFo/Samsung Process is 2.2 times denser than 28nm TSMC.

Lets think about it for a second. 600 mm2 Die from TSMC like Fiji would be... 250mm2 on 14 nm. It is extremely close to that 232 mm2 die size.

Whats more, if reality turns out to be true with what AMD says about the architecture, it is cleaned up, revamped and redesigned same GCN architecture. It would also affect the density of the chip. Coincidence? Of course, apart from my educated guesses/guesses I do not have anything to prove it.

But think about it. What better way for AMD would be bringing VR performance to Pitcairn price/performance bracket? They can also cut it to 3072 GCN cores, and still have VR GPU.

I will not talk about performance/SIMD because we do not know anything apart from words from Mahigan on that side, but we may be really surprised...

Imagine a 3072 GCN4 core GPU with 199$ price point matching in performance Titan X...

 

[Erenhardt]

I expect die sizes round what we saw with 28nm introduction. P10 around 200-250mm2 and GDDR5X and P11 with 350-400mm2 + HBM.

The next big thing will probably be 2xP11 on single interposer acting as one GPU.

 

[Glo.]

P10, from what people reported, who have seen the die, is around 120mm2.

 

[ShintaiDK]

You know, by their own words they only changed every relevant piece of the design. Clearly there can be no benefits...

Yes they did that every time. But the same uarch is the same uarch.

They also said the performance/watt is 2x. A direct shrink will pretty much do that.

 

[Leadbox]

Yes they did that every time. But the same uarch is the same uarch.

They also said the performance/watt is 2x. A direct shrink will pretty much do that.

They also said it was a 70-30 split, I think you're just caught up on the GCN generational naming scheme in drawing your conclusions.

 

[jpiniero]

Imagine a 3072 GCN4 core GPU with 199$ price point matching in performance Titan X...

The way AMD is talking about it the full Polaris 10 is going to be in that price range, not Polaris 11. I'll say $599 if it can get close to Fury X performance.

 

[Glo.]

The way AMD is talking about it the full Polaris 10 is going to be in that price range, not Polaris 11. I'll say $599 if it can get close to Fury X performance.

That would be the point IF... AMD themselves said that they have to bring that kind of performance lower in the price/performance bracket.

 

[railven]

Two very recent statements from AMD.


We can safely say that polaris 10 as demoed will not be the good quality VR GPU. If you interpret the 2 versions statement to mean 2 GPU only, then the affordable VR GPU [290/970 class] must be Polaris 11. Cheaper than R9 390.

This leaves nothing higher in the performance and price range.
Is this realistic?

If however, we interpret the 2 versions as different designs, one with GDDR5 and the other with HBM, then we can start to see a much wider product offering. Polaris 10 (small) as the demoed GPU, Polaris 10 (large) as the affordable VR, possibly 390 performance at 380 prices.

Then we move to Polaris 11 and HBM. This is the one for which no info is known, probably very deliberately.

More I think of it:

So we're going to have to make it possible to run good quality VR at a much lower price. And I'm confident with Polaris we're going to have a big impact to help that.

It don't explicitly say that Polaris would be the means to make the "quality VR run at a much lower price" the second statement implies Polaris will help.

Perhaps by replacing the top tier price point and thus allowing AMD to sell older GPUs that can run "quality VR at a much lower price." Polaris would maintain their margins while last gen spreads VR goodness to the masses.

 

[maddie]

Yes they did that every time. But the same uarch is the same uarch.

They also said the performance/watt is 2x. A direct shrink will pretty much do that.

For TSMC at least, this is not true.

http://www.tsmc.com/english/dedicatedFoundry/technology/16nm.htm

TSMC's 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed, around 2 times the density, or 70 percent less power than its 28HPM technology.

 

[Kenmitch]

It don't explicitly say that Polaris would be the means to make the "quality VR run at a much lower price" the second statement implies Polaris will help. Perhaps by replacing the top tier price point and thus allowing AMD to sell older GPUs that can run "quality VR at a much lower price." Polaris would maintain their margins while last gen spreads VR goodness to the masses.

Isn't VR more AFR friendly now? Isn't that the way AMD plans on doing it?

VR it seems will have way more uses than gaming. Around here the focus is strictly on the gaming aspect. To me it seem like less gpu power would be needed to render a static image of say the Smithsonian or a persons heart, etc.

 

[ShintaiDK]

For TSMC at least, this is not true.

http://www.tsmc.com/english/dedicatedFoundry/technology/16nm.htm

You use ISO terms as reference. Its pretty much useless for chip comparison.

If we was to believe that all the time. We would sit with 5Ghz GPUs instead of ~1-1.2Ghz

 

[maddie]

More I think of it:



It don't explicitly say that Polaris would be the means to make the "quality VR run at a much lower price" the second statement implies Polaris will help.

Perhaps by replacing the top tier price point and thus allowing AMD to sell older GPUs that can run "quality VR at a much lower price." Polaris would maintain their margins while last gen spreads VR goodness to the masses.

I suppose this is possible, but can you imagine the marketing gymnastics to sell these separate lines with their hugely divergent power consumptions, especially with AMD strongly stated their power efficiency drive.

As I said, the interpretation of the 2 GPU statement is the key. The popular one leads to marketing, pricing and performance complications. Witness the long discussions on what might be possible within that limitation.

The other interpretation of it meaning 2 classes of GPU allows a much more realistic and simpler scenario. A lot of problems disappear.

Occam's razor applied, in effect.

All my opinion of course.

 

[AtenRa]

I doubt that. The chip is simply not big enough. I would guess on a 3072SP part. And I doubt GCN 1.3 offers any meaningful benefits.

28nm Hawaii/Grenada XT
438mm2 - 6200 Million transistors

~14,15 Transistors/mm2

14nm FF has ~2.2 density = ~31,15 Transistor/mm2 (Assuming we keep the same design)

Polaris 11 = 232mm2

So Polaris 11 could have close to ~7200 Million transistors or one Billion transistors more than Hawaii/Grenada

For comparison, Fiji has 8900 Million Transistors

So even at the same clocks, Polaris 11 at 232mm2 could be faster than Hawaii/Grenada XT.

 

[railven]

I suppose this is possible, but can you imagine the marketing gymnastics to sell these separate lines with their hugely divergent power consumptions, especially with AMD strongly stated their power efficiency drive.

That's easy, cost. Someone here said it - efficiency is a new "marketing" term in itself. "Go Green." "Buy Carbon Dollars" or whatever the hell they are called.

Say you have two equal products:
GPU R7 470 (equal performance to current R9 390) @ $150-200
GPU R9 480 (R9 390 + 10%) @ 300-350

the 470 uses 220W, the 480 uses 110W. You will have the savvy buyers (and those cutting edge) opting for 480, but those on a budget would probably get the 470.

And now both users can run VR if the 290 is the base-point they are aiming for.

[All numbers above are made up, just used for my point]

As I said, the interpretation of the 2 GPU statement is the key. The popular one leads to marketing, pricing and performance complications. Witness the long discussions on what might be possible within that limitation.

The other interpretation of it meaning 2 classes of GPU allows a much more realistic and simpler scenario. A lot of problems disappear.

Occam's razor applied, in effect.

All my opinion of course.

At this point, who knows. My aggressive stance that they need to make more money (now) stems from the fact that NV is going to keep doing their little tricks and AMD is going to need resources to fight back. They've had amazing tech for years, often selling it for pennies compared to NV. If this is their chance to make some money, grow their brand (no longer the budget brand!), I say do what you can and go for it.

However, I also have zero confidence in their marketing teams so who knows if it will even matter if they have any advantages when they can't get joe public to acknowledge it.