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

[tential]

The answer is simple: They won't. They don't have the money for a full new line-up, so we're going to see more rebrands. The multi-die talk is pointless because it's not currently feasible even if we accept the slim chance that it's already possible. You're just trying to be optimistic, hoping that AMD hs some kind of secret sauce to make them more competitive. Trust me, i wish that what you were saying were possible so AMD would stand a chance. Sadly, there are plenty of facts against it and none for it. The logic that you're using behind that question also could have applied to the 200 and 300 series, which is why we saw this exact same speculation both times. Nothing has changed. If you want to make tis arguement, find some actual, specific evidence. Your leading question is only proof that you lack actual evidence.

Basically you're saying pascal is already the better chip for the high end for next gen?

 

[imaheadcase]

History has shown putting stock in one person to make a great piece of hardware has always failed.

 

[Arachnotronic]

History has shown putting stock in one person to make a great piece of hardware has always failed.

This Koduri guy is probably the only AMD exec that I actually have a real degree of confidence in. He appears to have identified what AMD screwed up in graphics and seems to be doing his best to try to fix things with whatever resources he's got.

Whether he will be successful, only time will tell, but after losing top notch guys like Eric Demers and Carrell Killebrew, AMD GPU division has been sorely in need of good leadership, and we'll know over the next 3-4 years if he is as good as some hope.

 

[Techhog]

Basically you're saying pascal is already the better chip for the high end for next gen?

No. I'm saying that there will me more Pascal chips than Polaris chips this year.

 

[Zanovar]

The 980ti is definitely worth it.... Fury x sucked but the 980ti was definitely worth it. Just like the 290 and even 390 is worth it, and worth it to go cf on both in my opinion due to how well cf scales.
The 970/sli is more debatable as we can see instances where it's hindered by its 3.5 in sli. In single card it's fine though, although in this price bracket I do think you should keep sli/cf open since it is great price to performance usually.

LOL the fury x didnt and doesnt suck.priced wrongly perhaps ?

 

[Headfoot]

I'm guessing we'll see 5 graphics cards, or maybe even 6. 3 SKUs per die. Probably 4 initially (2 skus per die, 1 full sku and 1 salvage sku) with the other 2 to be special new die salvaged parts that will be aimed to best counteract whatever nVidia's counter to the first wave of Polaris cards. Although I also put equal odds that Polaris 10 (small) comes out a few months prior to Polaris 11.

 

[maddie]

Basically you're saying pascal is already the better chip for the high end for next gen?

No. I'm saying that there will me more Pascal chips than Polaris chips this year.

Which strongly implies agreement with what tential wrote.

 

[Techhog]

Which strongly implies agreement with what tential wrote.

He specifically said "for the high-end." Again, I don't see how you're getting this from my statement. I'm saying that there won't be a new mid-range (pricing-wise) chip from AMD this year, since that's what's being said here. The only way there could be is for your multi-die dream to come true, and since i don't believe that's possible, I expect Polaris to only compete with GP107/6 and GP104 this year. I also expect that GP100 will not see consumer release at all until next year, which is also when we'll see a large 5th-gen GCN chip.

 

[tential]

He specifically said "for the high-end." Again, I don't see how you're getting this from my statement. I'm saying that there won't be a new mid-range (pricing-wise) chip from AMD this year, since that's what's being said here. The only way there could be is for your multi-die dream to come true, and since i don't believe that's possible, I expect Polaris to only compete with GP107/6 and GP104 this year. I also expect that GP100 will not see consumer release at all until next year, which is also when we'll see a large 5th-gen GCN chip.

I'm not trying to imply anything by the way I'm just trying to understand your comments. So you are saying then amd will have a high end chip and a low end chip and no mid range?

If so, what do you think amd will use as the midrange? The old chips currently out?

I'm not trying to be leading or condescending lol just genuinely trying to figure out what you're making of amds 2 chips that we know of and what that means for the next gen.

 

[MrTeal]

Out of curiosity, how many chips do you think nVidia will release as consumer GPUs this year? GP104 seems to be a given, but other than that there seem to be a lot of different opinions on what we'll see.
Last gen we got GK104 (294mm²) in March 2012, GK107 (118mm²) in April, and GK106 (221mm²) in September, with GK110 (561mm²) following in February (Titan) and March (780) the following year. I don't think we'll have to wait a full six months this time to get GP107, 106 and 104, but I would be surprised if nVidia did launch top to bottom in the span of a few months.

 

[sontin]

I think it just depends on the supply side. With 28nm nVidia was supply constraint. GK110 was ready for production but TSMC was only able to bring up their second 28nm facility in november 2012.

 

[MrTeal]

I think it just depens on the supply side. With 28nm nVidia was supply constraint. GK110 was ready for production but TSMC was only able to bring up their second 28nm facility in november 2012.

GK110 was in production earlier, obviously. They were shipping in September 2012, just not into the consumer space. It would have been interesting to see how nVidia staggered their launch if they had more capacity.

They might get relatively lucky on 16FF+. If the 6s were still selling like hotcakes capacity might be a major concern, but with the slowdown in the smartphone market there might be more supply than they had planned on.

 

[GodisanAtheist]

If so, what do you think amd will use as the midrange? The old chips currently out?

The answer is simple: They won't. They don't have the money for a full new line-up, so we're going to see more rebrands.

I'm guessing we'll see 5 graphics cards, or maybe even 6. 3 SKUs per die. Probably 4 initially (2 skus per die, 1 full sku and 1 salvage sku) with the other 2 to be special new die salvaged parts that will be aimed to best counteract whatever nVidia's counter to the first wave of Polaris cards.

-These are basically the options we have on the table:

1) AMD fills the middle of their range with Fiji/Tonga, which would be nuts due to power/feature/performance disparities. However desperate times call for desperate measures. The 390/x series has apparently done rather well for itself despite being the 290/x with some more ram. Fury by another name may smell much sweeter...

2) Harvest the hell out of their new dies: 3 or even 4 dies from Polaris 11 depending on how bad yields are, 2 or 3 dies from Polaris 10. New line-up would look something like this:

490x - Polaris 11 full
490 - Polaris 11 - 10%
480 D: - Polaris 11 -30% (Non-X as it isn't a full chip)

470X - Polaris 10
470 - Polaris 10 - 10%
460 - Polaris 10 - 30% (Non-X as it isn't a full chip)

Maybe they bump Polaris 10 down to the 460X and the 475/470 or whatever are the real garbage chips of the Polaris 11 line. At the same time, if Polaris 11 is going to use HBM, I wonder how low AMD is willing to go with harvested chips if the HBM remains a fixed cost.

Anyhow 2016 is going to be fun!

 

[3DVagabond]

What AMD showed @ CES was 850MHz .82v (IIRC). Who knows how fast that chip can clock? Or if it was a full die? Nobody actually saw it. The size estimates given were based strictly on performance shown.

 

[PPB]

-These are basically the options we have on the table:

1) AMD fills the middle of their range with Fiji/Tonga, which would be nuts due to power/feature/performance disparities. However desperate times call for desperate measures. The 390/x series has apparently done rather well for itself despite being the 290/x with some more ram. Fury by another name may smell much sweeter...

2) Harvest the hell out of their new dies: 3 or even 4 dies from Polaris 11 depending on how bad yields are, 2 or 3 dies from Polaris 10. New line-up would look something like this:

490x - Polaris 11 full
490 - Polaris 11 - 10%
480 D: - Polaris 11 -30% (Non-X as it isn't a full chip)

470X - Polaris 10
470 - Polaris 10 - 10%
460 - Polaris 10 - 30% (Non-X as it isn't a full chip)

Maybe they bump Polaris 10 down to the 460X and the 475/470 or whatever are the real garbage chips of the Polaris 11 line. At the same time, if Polaris 11 is going to use HBM, I wonder how low AMD is willing to go with harvested chips if the HBM remains a fixed cost.

Anyhow 2016 is going to be fun!

If nV can get away with +3/4 mkt share with just 3 dies, AMD might be fine with just 2 until it gets some market share back. This time tho the die salvaging needs to get serious. So I agree with the salvaging up to 30% less shaders. This goes in line with a process that will start with low yielding and this wont rule out a third die sometime down the road for Polaris (obviously for a node expected to run for 4 years, we will see a fuckton of dies for said node, but I'm specifically talking about this very "generation" of skus).

 

[Vesku]

Given the likely performance gap they may just keep selling Hawaii inbetween Polaris 10 and Polaris 11. May be unnecessary if Polaris 10 can clock high enough, though.

 

[MrTeal]

What AMD showed @ CES was 850MHz .82v (IIRC). Who knows how fast that chip can clock? Or if it was a full die? Nobody actually saw it. The size estimates given were based strictly on performance shown.

What are you basing that on? Ryan's article seemed to say he had seen the die itself. It's hard to imagine he said this based on only seeing the card with cooler.

In any case, the GPU RTG showed off was a small GPU. And while Raja’s hand is hardly a scientifically accurate basis for size comparisons, if I had to guess I would wager it’s a bit smaller than RTG’s 28nm Cape Verde GPU or NVIDIA’s GK107 GPU, which is to say that it’s likely smaller than 120mm2.

 

[3DVagabond]

Given the likely performance gap they may just keep selling Hawaii inbetween Polaris 10 and Polaris 11.

They should be able to make Nano appealing at the right price. Depends on what it costs to manufacture though.

 

[3DVagabond]

What are you basing that on? Ryan's article seemed to say he had seen the die itself. It's hard to imagine he said this based on only seeing the card with cooler.

Well, I never saw where anyone said they actually saw the physical chip. I haven't scoured all of the articles though. We still don't know what the performance ceiling is on the chip though. Or, do we think 850MHz is topped out?

 

[maddie]

What are you basing that on? Ryan's article seemed to say he had seen the die itself. It's hard to imagine he said this based on only seeing the card with cooler.

He did see the die. From same article.

Similarly, while Raja Koduri held up an unsoldered version of the GPU used in the demonstration, again we were not allowed to take any pictures

To everyone, who will buy a 28nm GPU mid range card with the competition using a fraction of the power. This is the core of sales segment.

 

[MrTeal]

He did see the die. From same article.



To everyone, who will buy a 28nm GPU mid range card with the competition using a fraction of the power. This is the core of sales segment.

Who bought a 560Ti when it was slightly slower than a 7850 and used 65% more power for the six months before GK106 launched? nVidia priced it appropriately, and still sold GPUs in the $200 market. Not an ideal situation, but you make due with what you have.

They should be able to make Nano appealing at the right price. Depends on what it costs to manufacture though.

I would think the BOM cost of the Nano might be a little much to push it down in to the $200 range. Hawaii seems much more likely, and the performance relative to a 960-class Polaris 10 and whatever Polaris 11 comes in at would seem to make more sense.

 

[railven]

How likely is that one of the new modular things they mentioned (Memory Controller) can get retrofitted to Fiji and you get some kind of Frakenstein Fiji+GDDR5(X) variant?

Be interesting to see what AMD does to fill the holes.

 

[Silverforce11]

2 chips allow 4 SKUs with one harvest, 6 SKUs with a 2nd harvest or cut-down parts.

That's plenty to fill out low to high. With the option of a dual-Polaris 11 x2 SKU for the enthusiast segment.

And Robert Hallock said the Polaris design itself is capable of HBM or GDDR5 and they will use either depending on the segment due to price sensitivity.

Thus, it's very possible for a harvested Polaris 11 to be on GDDR5 to keep it affordable at the mid-range sweetspot pricing ~$300.

R490X2: Dual Polaris 11 w/ HBM2
R490X: Polaris 11 w/ HBM2
R490: 10% cut-down Polaris 11 w/ HBM2
R485: 20% cut-down Polaris 11 w/ GDDR5X

That's 4 SKU from Polaris 11 alone.

 

[MrTeal]

2 chips allow 4 SKUs with one harvest, 6 SKUs with a 2nd harvest or cut-down parts.

That's plenty to fill out low to high. With the option of a dual-Polaris 11 x2 SKU for the enthusiast segment.

And Robert Hallock said the Polaris design itself is capable of HBM or GDDR5 and they will use either depending on the segment due to price sensitivity.

Thus, it's very possible for a harvested Polaris 11 to be on GDDR5 to keep it affordable at the mid-range sweetspot pricing ~$300.

R490X2: Dual Polaris 11 w/ HBM2
R490X: Polaris 11 w/ HBM2
R490: 10% cut-down Polaris 11 w/ HBM2
R485: 20% cut-down Polaris 11 w/ GDDR5X

That's 4 SKU from Polaris 11 alone.

You couldn't just cut it down and put on GDDR5, it would have to be an entirely different die.

 

[Silverforce11]

You couldn't just cut it down and put on GDDR5, it would have to be an entirely different die.

Why, AMD have said Polaris was designed to support both GDDR5 and HBM and which they use will depend on the market segment due to HBM pricing, its in their videos.