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

[Armsdealer]

They need to come out and clarify the steam VR number. It seems like an outlier, and they're morons to have put it on a marketing slide.

 

[tviceman]

RX 470 already scores as high as a R9 290x Quiet mode in Firestrike Standard, is all looking good.

Incorrect.

http://www.guru3d.com/articles-pages/powercolor-radeon-r9-390-pcs-8gb-review,23.html

Looks slower than 290x by ~13%.

 

[PeckingOrder]

I'm not going to bother looking up fire strike scores, but this leak shows the RX 470 only being 33% faster than a 270x, and a 290x/390 is way more than 33% faster than a 270x.

Sooooo......

So the Chinese benchmarks that half of this forum believed were real have turned out to be false as expected (or benched OCed with nitrogen cooling)

 

[DiogoDX]

I'm not going to bother looking up fire strike scores, but this leak shows the RX 470 only being 33% faster than a 270x, and a 290x/390 is way more than 33% faster than a 270x.

Sooooo......

What?

9090/5787 = 1.57 = 57% faster

Average on the games was 79% faster.

 

[Timorous]

What?

9090/5787 = 1.57 = 57% faster

Average on the games was 79% faster.

Indeed, and based on the 7850 and 7870 it's not unreasonable to expect the 480 to be about 20% faster than the 470 which puts it pretty much neck and neck with the 390X and the 980.

 

[tviceman]

The last few months of hype and the inevitable disappointment:

1. AMD will beat Nvidia to release because it's a new node. (false)
2. Samsung 14nm FF is superior to TSMC FF and achieves 2.5x density, so 230mm2 Polaris should easily match Fury X and will be a disappointment if not. (false)
3. Polaris 10 early benchmarks aren't disappointing because it's a cut down chip. (false)



Now we're looking at RX 480 potentially being slower than GTX 980 and R9 390x, and definitely being slower than an OC'd GTX 980. Man this is turning into a huge disappointment.

 

[tviceman]

What?

9090/5787 = 1.57 = 57% faster

Average on the games was 79% faster.

Sorry that was eyeball math when I looked away. I edited my post with a link to show scores of various other cards in fire strike, and I was still correct in my assumption that RX 470 doesn't equal 290x/390 speeds in firestrike.

 

[PhonakV30]

28.1 fps /180w = 0.1561 , 46 fps / 110w = 0.4181
==> 0.4181 / 0.1561 = 2.678

27.6 fps/ 180w = 0.153 , 60fps/110w = 0.5454
==> 0.5454 / 0.153 = 3.557

76fps/180w = 0.422 , 121/110w = 1.1
==> 1.1 / 0.422 = 2.605

But compare to Maxwell.It's much lower than this.AMD couldn't make it 90w, why?

 

[tviceman]

76fps/180w = 0.422 , 121/110w = 1.1
==> 1.1 / 0.422 = 2.605

But compare to Maxwell.It's much lower than this.AMD couldn't make it 90w, why?

270x doesn't use 180 watts. It uses much less. http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1070/22.html

 

[Olecki]

They need to come out and clarify the steam VR number. It seems like an outlier, and they're morons to have put it on a marketing slide.

In April Rod Taylor from AMD (http://arstechnica.co.uk/gadgets/2016/04/amd-focusing-on-vr-mid-range-polaris/) stated that Polaris is aiming in 970/290 performance level. If you compare VR score it nearly exactly match it. There was so mych hype about Polaris 10 that now many people couldn't believe this.

 

[Borisi]

I don't know much about the Steam VR test, but it seems that the CPU matters a lot... And RX 480 makes it barely using a i7 5960x so it seems that it won't run VR in like 90% of rigs... including everything with an AMD CPU.

Wow, if this is real they're absolute morons trying to hype the card for VR.

 

[Mopetar]

The last few months of hype and the inevitable disappointment:

1. AMD will beat Nvidia to release because it's a new node. (false)
2. Samsung 14nm FF is superior to TSMC FF and achieves 2.5x density, so 230mm2 Polaris should easily match Fury X and will be a disappointment if not. (false)
3. Polaris 10 early benchmarks aren't disappointing because it's a cut down chip. (false)

I believe that the first two are due to problems with the 14 nm process at Global Foundries. The process tech is Samsung's but it's not Samsung that's making the chips. Given GF's history it's not hard to believe that they botched it considering AMD had shown off working silicon as early as January and there were plenty of other leaks that pointed to them being able to launch first.

The Samsung process does allow for better density than the TSMC process, but Apple dual sourced their SoC from both and it doesn't appear as though either offers much practical difference in terms of performance, or at least we can't tell based on how Apple binned the chips.

I'm also still under the impression that the 480 is a cut chip. Another poster speculated that since Apple is using AMD as their graphics partner for their refreshed computers that they've bought up the entire supply of uncut chips. Another take is that due to poor GF yields, AMD doesn't have enough left over after fulfilling their contract with Apple so we won't see any until later. I don't quite think we can put this one to bed yet.

 

[Armsdealer]

In April Rod Taylor from AMD (http://arstechnica.co.uk/gadgets/2016/04/amd-focusing-on-vr-mid-range-polaris/) stated that Polaris is aiming in 970/290 performance level. If you compare VR score it nearly exactly match it. There was so mych hype about Polaris 10 that now many people couldn't believe this.

It's coming in below 970 and below 390 in that particular bench. All other results we've seen thusfar are fine and not a disappointment. It's absurd for a 36 cu part (rx480) to underperform the 40 cu part (390) from the previous node. Something is off. If that particular bench is just an outlier that places the part on a bad light, they're foolish to have released it in marketing materials.

 

[PhonakV30]

270x doesn't use 180 watts. It uses much less. http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1070/22.html

with Error margin , still my math is correct, If peak gaming is 130w , this can be dropped to 80w or even lower for RX 470.
at worse scenario , Ratio is above 1.7x.
at best scenario , Ratio is around 2.6x.

 

[DDH]

Still have to wait for the independent reviews. But things are not looking good for me personally in terms of upgrade path.

Its times like this I wish there was a third player in the performance dgpu market. Nvidia I refuse to buy, AMD I would have bought but offers me no incentive to change my 290x. Perhaps I need to take a look at the second hand market for a bargain

 

[geoxile]

I believe that the first two are due to problems with the 14 nm process at Global Foundries. The process tech is Samsung's but it's not Samsung that's making the chips. Given GF's history it's not hard to believe that they botched it considering AMD had shown off working silicon as early as January and there were plenty of other leaks that pointed to them being able to launch first.

The Samsung process does allow for better density than the TSMC process, but Apple dual sourced their SoC from both and it doesn't appear as though either offers much practical difference in terms of performance, or at least we can't tell based on how Apple binned the chips.

I'm also still under the impression that the 480 is a cut chip. Another poster speculated that since Apple is using AMD as their graphics partner for their refreshed computers that they've bought up the entire supply of uncut chips. Another take is that due to poor GF yields, AMD doesn't have enough left over after fulfilling their contract with Apple so we won't see any until later. I don't quite think we can put this one to bed yet.

The 480 is using Polaris 10 XT, which suggests it's the full chip. And Samsung's 14nm LPP could be denser than LPE, which is what Apple used for the dual sourced A9.

It's probably likely that the Polaris 10 just has a bunch of new blocks (or proportionally bigger blocks) on it that are taking up space.

 

[Olecki]

It's coming in below 970 and below 390 in that particular bench. All other results we've seen thusfar are fine and not a disappointment. It's absurd for a 36 cu part (rx480) to underperform the 40 cu part (390) from the previous node. Something is off. If that particular bench is just an outlier that places the part on a bad light, they're foolish to have released it in marketing materials.

But AMD stated about 290, not 390 and score is matching perfectly to it. Compared to 970 is a little bit (3% ?) slower than 970 - it's not far behind. But what people were expecting from new process it's different story.

 

[tviceman]

I believe that the first two are due to problems with the 14 nm process at Global Foundries. The process tech is Samsung's but it's not Samsung that's making the chips. Given GF's history it's not hard to believe that they botched it considering AMD had shown off working silicon as early as January and there were plenty of other leaks that pointed to them being able to launch first.

The Samsung process does allow for better density than the TSMC process, but Apple dual sourced their SoC from both and it doesn't appear as though either offers much practical difference in terms of performance, or at least we can't tell based on how Apple binned the chips.

I see the excuses are already flying off the shelf before the cards even arrive. If GF had problems, then AMD would be getting bad yields, not bad density or bad perf/transistor. 3 months ago the very people talking up Polaris 10 in this thread said a 230mm2 GF 14nm FF chip should equal Fury X in performance and that R9 390x performance, let alone slower like what we're seeing now, would be a disappointment.

I'm also still under the impression that the 480 is a cut chip. Another poster speculated that since Apple is using AMD as their graphics partner for their refreshed computers that they've bought up the entire supply of uncut chips. Another take is that due to poor GF yields, AMD doesn't have enough left over after fulfilling their contract with Apple so we won't see any until later. I don't quite think we can put this one to bed yet.

Official AMD slide specifically says Polaris 10 and Polaris 11, not RX 480, 470 and 460. What more proof do you want?

 

[Silverforce11]

3. Polaris 10 early benchmarks aren't disappointing because it's a cut down chip. (false)

I believe it's a cut down chip. Just like Tonga's 384 bit bus and full shaders which were cut. We later learnt that from detailed die shots.

The reasons why I think it's 40 CU, is because of two reasons:

1. Polaris 11 was leaked as a 20 CU 1280 SP chip (manifest and SiSoft database). In GPU design, it is cost effective and time effective when scaling up or down to double up for a chip that's more than twice as big. 20 -> 40 CU. This is especially true for the same generation of architecture.

2. PS4 Neo confirmed 36 CU Polaris 10. Nobody sane would build a mass volume APU with only 36 CU and try to get most chips usable with all 36. It would have to be a redundant harvested part that makes it into the final volume PS4 Neo. This means >36 CU.

Time will prove me right or wrong, but it's the most logical design. In fact, we'll know soon enough once Apple's new Macs arrive.

One more #, only one Polaris 11 SKU RX 460, where's the harvested chip going? Is AMD selling cut down Polaris 11 to Apple? lol

 

[JDG1980]

So the Chinese benchmarks that half of this forum believed were real have turned out to be false as expected (or benched OCed with nitrogen cooling)

Any reason they couldn't have been conventional overclocks?

 

[Silverforce11]

These guys are claiming 100% certainty because they have the card from retail on hand to test for these results.

It is faster than 390X in this old benchmark. So what has changed that some of you now suggest it's slower than a 390X?

 

[JDG1980]

These guys are claiming 100% certainty because they have the card from retail on hand to test for these results.

It is faster than 390X in this old benchmark. So what has changed that some of you now suggest it's slower than a 390X?

Different clock rates (NDA traps, or maybe just different AIB cards with different factory OC) could account for most if not all of these disparate benchmarks. We aren't talking about that huge of a variation in performance, just 10%-20%.

 

[crisium]

Also it's AMD themselves giving us a VR score that most Hawaii cards beat.

 

[Armsdealer]

It's one steamVR bench that's purportedly from AMD marketing slides that shows rx480 with a 6.3, which is a bit over 10% slower than r9 390 according to tomshardware benches.

AMD PR at its finest - always leaving you with that lingering doubt that perhaps they've completely screwed up p10.

 

[ZGR]

Why is the 390 -> 390X gap from 79% to 97%? That doesn't line up with the scores above.