AMD Polaris Thread: Radeon RX 480, RX 470 & RX 460 launching June 29th

[Mopetar]

I may just cream my pants with excitement, if that price of $299 comes true for 3072 GCN cores for a 490 (about 33% more cores over 480).

Yep, very interesting development. I hope we could see this chip sooner rather than later.

I wouldn't call it a development. It's a single post from someone who by all counts should know that information and it doesn't mesh with any of the previous information that we have about Polaris 10 or make much sense in the context of what we already have.

A 3072 core Polaris 10 means that the die size rumors were completely off, AMD somehow managed to make their SP design about 50% more efficient in terms of transistor usage, or that AMD is packing their chip something like 50% more densely than Nvidia has managed, which is far, far beyond the density advantage that Apple saw with Samsung's process above TSMCs.

Never mind that all of this was done with out any leaks from more credible sources. This one falls under the too good to be true category, which probably explains why so many people are jumping on it because they want to believe.

 

[.vodka]

That sounds too good to be true. I'm taking it with a grain of salt for now. But if it is accurate... Nvidia could be in for some serious trouble.

AtlasRush seems to be saying that 470/480/490 would all be Polaris 10 SKUs. Further down the thread, he indicates that Polaris 11 would be the 460/450/440 (with "R" prefix instead of "RX"). If all this is true, it seems to indicate that GloFo 14LPP yields are fairly bad at this time. Not only would there be third-tier salvage parts for even the small Cape-Verde-sized Polaris 11 die, but the third-tier salvage for Polaris 10 (RX 470) would be cut down by a full 50% (1536 of 3072 shaders enabled). AtlasRush was also careful to say "NDA for 480 expires on 29th." Thus even if this is accurate, we may not see the full 490 until later. And if yields are really bad, it may take some time for availability to become widespread.

GF as usual could be screwing AMD up. There was a rumor a few weeks ago saying that Vega got pushed back to October and was the 490x in old naming. In that thread they say Vega isn't the 490, and it's still due for 2017.

What if the RX490 is actually full P10 that's coming in October? I mean, 22-33% more SPs over RX 480 certainly warrants the 490 branding. We also know there's something above $200-230 4 and 8GB RX 480. With all that time and poor yields there could be a decent amount of healthy dies to sell.


I think the pieces of the puzzle are starting to come together... Let's say 48CU dies are reserved for Apple like Tonga was. 44CU is Hawaii size, with the GCN4 (and by extension GCN3) improvements, that's one hell of a GPU for a Pitcairn replacement at 232mm^2 at the $300 bracket. Makes one wonder what Vega holds as a new uarch and HBM2 on top if this is what Polaris actually is. Damn.


Still, all we know for sure is RX480 is 2304SP, 36CU. The rest is speculation.

 

[JDG1980]

That seems very unlikely, but if it were true it would be deserving of the 490 moniker. At the same time, that makes the 480 a massively cut die, the likes of which we haven't seen since . . . well the 1070 actually, but that's pretty unexpected for a ~230 mm^2 chip to yield that poorly, even if I believe that Global Foundries has had issues with their new node.

I think the long stagnation on 28nm may have biased some expectations about what kind of average yields should be obtained. Nvidia doesn't seem to be doing particularly stellar on yields either, with even the ultra-specialized P100 being a cut-down chip. And the GP104 in the GTX 1070 is cut down by a full 25% of shaders, more than expected. While that latter decision could be the result of market segmentation, I think it has to do with the rush to production and the fact that TSMC is still a relative newcomer to 16FF+.

We'll probably be seeing mediocre yields for 6 months or so. Maybe up to a year. For reference, it was in October 2013 that AMD released the first wave of R9 200 series rebrands, which (temporarily) did away with cut chips for Pitcairn and Tahiti. That was about 22 months after the first 28nm GPU (7970) hit the review sites.

Also, it doesn't make sense for other reasons. All of the rumors have pointed towards Polaris 10 being around ~230 mm^2. If they managed to fit 3072 SPs in that area, it would be around 2.3 times as dense (I'm basing density off of SP/mm^2 as we don't have a transistor count for Polaris 10 yet) as their previous generation chips, when Nvidia has well under that with the 1070/1080.

Samsung's FinFET process is known to be denser than TSMC's. The Apple A9 smartphone chip was dual-sourced on TSMC and Samsung, and the Samsung version was about 9% smaller. And that was with the Samsung 14LPE process. 14LPP (the version GloFo adapted) is supposed to provide even smaller size, as well as better performance. Also, don't forget that even on the same process node, there is a trade-off between density and speed. Nvidia could have crammed more shaders in GP104 in the same die size if they wanted, but then they wouldn't be able to hit the 2000+ MHz clock speeds that has everyone so impressed. Since at least Kepler, Nvidia has preferred to go for less dense and higher clocking designs, while AMD's GCN products are generally oriented around denser designs with lower clock rate caps.

In addition, we've all been assuming that Polaris 10 is a 232mm^2 chip based on some information written in an old LinkedIn profile. This belief was bolstered by AMD's positioning of P10 as a "mainstream" part, the modest TDPs, and the low pricing announced at Computex. But it's at least possible that we could be wrong, that the 232mm^2 GPU was something other than Polaris 10 (a future product, an engineering sample that never made it into production, whatever else) and that Polaris 10 is actually closer to 300mm^2. Only time will tell. Fortunately, we don't have all that much longer to wait.

 

[Glo.]

Hawaii XT was 438mm2 die that had 2816 GCN core chip with 2816 GCN cores, AND 512 bit memory controller.

Here you might have 232mm2 GPU with 256 Bit memory bus. Impossible to pack 10% more CU's on a node that brings 60% higher density?

It amazes me, that everybody compared Polaris 10 to Hawaii Pro forgetting that that was cut down part.

 

[rainy]

Don't get your hopes up. 2304SP is all that's confirmed, and AMD always barely cuts their chips at all unlike Nvidia. 2304 could still be full, and if its cut then 2560 is the most likely.

Why not 2816 (44 CU) like in Hawaii?
300 dollars for such a card would make sense.

 

[Mopetar]

There was a rumor a few weeks ago saying that Vega got pushed back to October

If Vega were released in October, it would be moved up by at least 2 months based on information that AMD has publicly released. Their road map from March put Vega clearly in the late 2016 / early 2017 window.

I think that moderate availability for the holidays is the best-case scenario at this point.

I think you might have misread that particular rumor.

 

[Vesku]

Should know more by Monday 1PM, I'm expecting AMD E3 PC Gaming event is where they will drop some decent spec info.

 

[Mopetar]

Samsung's FinFET process is known to be denser than TSMC's. The Apple A9 smartphone chip was dual-sourced on TSMC and Samsung, and the Samsung version was about 9% smaller. And that was with the Samsung 14LPE process. 14LPP (the version GloFo adapted) is supposed to provide even smaller size, as well as better performance. Also, don't forget that even on the same process node, there is a trade-off between density and speed. Nvidia could have crammed more shaders in GP104 in the same die size if they wanted, but then they wouldn't be able to hit the 2000+ MHz clock speeds that has everyone so impressed.

Like I said though, to support 3072 SPs at the rumored size is too far out of bounds to be acceptable. It's the kind of thing that if could actually be pulled off would be talked about for years.

So either the chip is much larger than rumored, or we're more likely seeing a 2560 SP part for the full die. A 44 CU 2816 part is at least within the realms of plausible going from everything else we know at this point.

 

[Mopetar]

Should know more by Monday 1PM, I'm expecting AMD E3 PC Gaming event is where they will drop some decent spec info.

There was also some suggestion that they might have a part in Apple's WWDC event considering that they're going to be supplying GPUs for Apple's new notebook and desktop offerings.

I doubt we'd see any real technical specs there as Apple tends to avoid those details, but they might discuss some performance numbers or provide some information that gives a better indication of performance.

 

[oussama-tn]

I hope we get some more info in AMD event during E3 i am really looking for an upgrade

 

[Mopetar]

I hope we get some more info in AMD event during E3 i am really looking for an upgrade

What are you using now?

A 480 is going to be more of a sidegrade for anyone using a 290+ right now. At $200 though it's a great deal for anyone under that level though. What value Polaris 10 offers at $300 remains to be seen, but I'd be surprised if it's much more than a 25% improvement over the 480.

Anyone who wants a serious upgrade from AMD is likely going to have to wait another 4 months at the very least.

 

[garagisti]

Don't get your hopes up. 2304SP is all that's confirmed, and AMD always barely cuts their chips at all unlike Nvidia. 2304 could still be full, and if its cut then 2560 is the most likely.

Well, if it is possibly 2560 it may just retain same or higher clocks depending on power allowance, i.e how much of the 150W this 2560 shader equipped 480x/ 490 may use.

I may not be remembering correctly so, but is the "under 150W" for Polaris 10 range or is it for only 480?

 

[Armsdealer]

no hype bro, no hype. I am just reading this, and putting down hypes.

I want benchmarks before any super hype.

I don't think that's hype. A budget low wattage part on a new process should be about as good of a candidate for overclocking as anything. I expect rx480 to overclock to near 980ti stock in dx12.

The hype is 48cu full chip which sounds very unlikely given it implies more cores than gtx1080 in a significantly smaller die than gp104.

 

[garagisti]

If Vega were released in October, it would be moved up by at least 2 months based on information that AMD has publicly released. Their road map from March put Vega clearly in the late 2016 / early 2017 window.

I think that moderate availability for the holidays is the best-case scenario at this point.

I think you might have misread that particular rumor.

I think this information about Vega & what is but Polaris chippery has been consistent since about 9 months or more now. Nothing solid has come up about any delays in plans of launching Vega, which iirc was scheduled towards end on 2016/ Q4. Obviously with a big chip, sales numbers will be on lower side compared to Polaris, and comparatively AMD could start with a fairly lower stock.

 

[Glo.]

Like I said though, to support 3072 SPs at the rumored size is too far out of bounds to be acceptable. It's the kind of thing that if could actually be pulled off would be talked about for years.

So either the chip is much larger than rumored, or we're more likely seeing a 2560 SP part for the full die. A 44 CU 2816 part is at least within the realms of plausible going from everything else we know at this point.

It isn't that far fetched as you think. 2816 GCN core 1.1 GPU is Hawaii XT and it is 438 mm2 GPU with 512 bit memory bus that takes die space. Here you have 232mm2 die, that is not even 50% smaller. You do not need 512 bit memory bus, so you have die size saving. Secondly GF 14 nm process is 60% more dense than Hawaii XT. Is it really impossible to pack 256 more Shaders in that scenario?

 

[Armsdealer]

It isn't that far fetched as you think. 2816 GCN core 1.1 GPU is Hawaii XT and it is 438 mm2 GPU with 512 bit memory bus that takes die space. Here you have 232mm2 die, that is not even 50% smaller. You do not need 512 bit memory bus, so you have die size saving. Secondly GF 14 nm process is 60% more dense than Hawaii XT. Is it really impossible to pack 256 more Shaders in that scenario?

2816 * 1.6 (process) * 232 / 438 = 2400

Not seeing your logic here. Maybe my math is wrong somewhere?

 

[RussianSensation]

Pricing is from what we know:
RX 480 4 GB - 199$
RX 480 8 GB - 229$
? - <300$

From what I've read it seems most AIBs will launch 8GB cards first. Generally speaking AIBs add a $20-40 premium on top around launch. I would expect retail 480 8GB to cost $249-269. I would be pleasantly surprised if we see Asus Strix or Sapphire Nitro 480 8GB for $229-239. The Reddit thread which discusses a 3072 SP AMD card has info that intental AMD documents show Fury replacement for Q2 2017. Sounds like AMD will go completely backwards from HD7000/290 series -> HD7750/7770 P11 + HD7850/7870 P10 first, then Vega 10 is HD7950/7970 replacement and Vega 11 seems like the > 400mm2 HBM2 290/290X replacement (which naturally is a Fury/X replacement also).

 

[Mopetar]

I may not be remembering correctly so, but is the "under 150W" for Polaris 10 range or is it for only 480?

We could very well see a Polaris 10 part above 150W. In the past AMD has done something similar for this product category. The original Pitcairn had both 6-pin and 2x 6-pin configurations in the 7850 and 7870.

AMD might not release a reference design like that, especially if they want to maintain the performance/watt mantra with Polaris, but there might be some third party offerings that can handle the extra load with better cooling solutions.

The hype is 48cu full chip which sounds very unlikely given it implies more cores than gtx1080 in a significantly smaller die than gp104.

It's not that they can fit more cores in a smaller area. We already know that AMD has leaner SPs as Fiji packs 4096 in a ~600 mm^2 die whereas GM200 only fits 3072 in about the same area, and we see similar ratios across other chips as well.

The issue with a 3072 SP Polaris 10 is that it makes the chip much bigger than anticipated or means that AMD was insanely aggressive with density of Polaris 10. Fiji was around 12% more dense than GM200, but Polaris would be almost triple that amount even after accounting for Samsung's node advantages over TSMC.

I'd like to see AMD work towards designing a beefier core as some of the performance issues they've been having are due to underutilization of their hardware. DX12 should alleviate some of that, but in general fewer, stronger cores tend to win out over weaker, more numerous ones.

 

[garagisti]

I don't think that's hype. A budget low wattage part on a new process should be about as good of a candidate for overclocking as anything. I expect rx480 to overclock to near 980ti stock in dx12.

The hype is 48cu full chip which sounds very unlikely given it implies more cores than gtx1080 in a significantly smaller die than gp104.

You may very well be right about hype about the part with 48cu.

 

[Glo.]

2816 * 1.6 (process) * 232 / 438 = 2400

Not seeing your logic here. Maybe my math is wrong somewhere?

Math is good, just irrelevant for this.

Count the percentage of die size between 438/232, remember that 512 bit memory bus is created from 8 memory controllers compared to 256 bit - 4.

Hawaii XT ported to 14 nm GF process would be 175 mm2. So you have a lot of room for another few shaders in the core.

232 is exactly 0.53 die size of Hawaii XT. 47% smaller die.

 

[Glo.]

I don't think that's hype. A budget low wattage part on a new process should be about as good of a candidate for overclocking as anything. I expect rx480 to overclock to near 980ti stock in dx12.

The hype is 48cu full chip which sounds very unlikely given it implies more cores than gtx1080 in a significantly smaller die than gp104.

Fiji XT - 4096 GCN cores in 596 mm2 die size.
Titan X - 3072 in 601 mm2 die size.

GTX 980 - 398 mm2 - 2048 CUDA cores with 256 bit memory bus.
R9 390X - 2816 GCN cores - 438 mm2 with 512 bit memory bus.

R9 390X - 2816 GCN cores with 512 memory bit bus - 438 mm2.
GTX 980 Ti - 2816 CUDA cores with 384 bit memory bus - 601 mm2.

Core counts have absolutely different affect on die sizes for both vendors and directly comparing them is useless.

 

[geoxile]

http://www.anandtech.com/show/9330/exynos-7420-deep-dive

If the Exynos SoCs' GPU clusters are any indication, going from Samsung's 28nm to 14nm LPE saw a >50% reduction in area per unit. So >2x more space efficient. It might not be the most comparable comparison since Samsung 28nm and TSMC 28nm aren't equal, but they are similar at least.

 

[Armsdealer]

Fiji XT - 4096 GCN cores in 596 mm2 die size.
Titan X - 3072 in 601 mm2 die size.

GTX 980 - 398 mm2 - 2048 CUDA cores with 256 bit memory bus.
R9 390X - 2816 GCN cores - 438 mm2 with 512 bit memory bus.

R9 390X - 2816 GCN cores with 512 memory bit bus - 438 mm2.
GTX 980 Ti - 2816 CUDA cores with 384 bit memory bus - 601 mm2.

Core counts have absolutely different affect on die sizes for both vendors and directly comparing them is useless.

I totally get that a smaller memory bus is a factor in favor of more cu on p10, but 48cu isn't coming within the realm of reason mathematically:

3072 on 600mm2 => 2560 on 314mm2 is a 60% improvement (titanx to gtx 1080)

Let's say glofo 14nm is 10% better than tsmc 16nm (apple a9 results) so 76% better than tsmc 28nm.

Then:
2816 cores on 438mm2 becomes 2625 on 232mm2

Maybe 44cu is possible, but 48... I don't know maybe you're right

 

[geoxile]

I totally get that a smaller memory bus is a factor in favor of more cu on p10, but 48cu isn't coming within the realm of reason mathematically:

3072 on 600mm2 => 2560 on 314mm2 is a 50% improvement (titanx to gtx 1080)

Let's say glofo 14nm is 10% better than tsmc 16% (apple a9 results) so 60% better than tsmc 28nm.

Then, my math is exactly right:
2816 cores on 438mm2 becomes 2400 on 232mm2

If someone said 2400 is actually 2560, fine ok, but 3072 is hard to see.

That's LPE vs. FF+, not LPP.

 

[Armsdealer]

That's LPE vs. FF+, not LPP.

Yeah, maybe it's possible. I also made an error in math. Gtx 1080 is 60% better than tx. Wow I hope full p10 is 48 cu that would be bonkers