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

[LTC8K6]

Why not use GDDR5X and save some power?

 

[Stuka87]

Has anybody measured the die from RX470 picture? It's P10 or P11?

RX470 is most likely cut down P10, RX460 is most likely P11.

 

[casiofx]

These leaks looked more believable

http://videocardz.com/61005/new-amd-radeon-rx-480-3dmark-benchmarks

 

[AtenRa]

heh nice one

Exclusive: How AMD's 'The Uprising' Ad Campaign Will Incite A Radeon Rebellion

 

[LTC8K6]

I still can't believe that VR is going to be that big of a market.

I can't believe a huge number of people are going to wear those things on their head for very long.

 

[Silverforce11]

These leaks looked more believable

http://videocardz.com/61005/new-amd-radeon-rx-480-3dmark-benchmarks

That's definitely more accurate, as the 980 tend to beat the 390X in 3dMark but they equalize in games.

Also, slightly faster than 390X at stock is pretty good.

It's going to come down to how well it overclocks. If we're talking 20-25%, you have yourself a Fury X & 980Ti killer for very cheap.

If it's meh like Pascal, 10%, then you have a card that's on the heels of those former flagships for cheap.

Either way, it's good.

Note that I do not think 3dMark will reflect overall Polaris gaming performance, it's major draw card (Discard Accelerator) come into play in complex FPS and open world games.

The way I see it, the worse case scenario is 390 performance. It will be decent if its 390X perf, and if it's faster, it's a great product.

 

[AtenRa]

I still can't believe that VR is going to be that big of a market.

I can't believe a huge number of people are going to wear those things on their head for very long.

I give it two years from now, VR will be huge

 

[crisium]

Inteluser is right to be concerned. Based on current leaked die sizes, it looks like AMD is falling further behind this generation than last. Yes, as bizarre as it sounds, it appears that Hawaii vs Maxwell was more favorable to AMD than Polaris vs Pascal is on performance per mm. If P10 is 232mm, then compare it to 390X. Same ratio as GP104 to GM200, yet the 1080 stands tall over the Titan X and based on leaks 480 barely edges out the 390X. AMD is losing ground if the leaks are true!

Sure they can price it to be competitive, and isn’t that all that matters? Well no, not if you are the high end user hoping for competition. Die sizes have limit, die sizes determine profit. It means AMD will continue to fall further behind in revenue and can only compete by lowering margins on more expensive chips. Bad news for the long run. And unless Nvidia allows AMD to make much bigger chips (the true folly of AMD from 2007-2011... insanity) then they cannot compete for the crown. Unless Vega really benefits from HBM2 (it is a good advantage in die size, wattage, and bandwidth... this is the only hope), then it looks like Nvidia will have the high ground again, and possibly on a smaller die. It also looks like GP106 will easily compete with Polaris, and on a smaller die too.

 

[littleg]

I give it two years from now, VR will be huge

I actually have a feeling AR is going to be the big one. There's a huge proportion of the population for whom shutting themselves off from the rest of the world entirely isn't really an option.

 

[Maverick177]

High end GPUs tends to have more perf/W than lower tier GPUs, but you are right, it's a big concern for AMD: 232mm with 390X perf compared to 320mm with 65% more perf.

 

[crisium]

It's going to come down to how well it overclocks.

If the leaks of stock performance are true, this is the best hope for Polaris. I don't know why AMD would leave performance on the table, but they did it 4 years ago and may again. It would also leave room for a GDDR5X version with higher core clocks as well for a refresh next year, especially if GTX 1060 is clocked just enough to beat 480.

 

[coercitiv]

Based on current leaked die sizes, it looks like AMD is falling further behind this generation than last. Yes, as bizarre as it sounds, it appears that Hawaii vs Maxwell was more favorable to AMD than Polaris vs Pascal is on performance per mm. If P10 is 232mm, then compare it to 390X. Same ratio has GP104 to GM200, yet the 1080 stands tall over the Titan X and based on leaks 480 barely edges out the 390X. AMD is losing ground if the leaks are true!

How do you evaluate evolution of perf/mm when one vendor offers a 28% reduction in TDP (250W -> 180W) while the other a minimum of 55% reduction in TDP (275W -> less than 150W)?

Unless Polaris is inherently incapable of clocking higher than what these early leaks suggest, at this moment we have little idea of the maximum potential of the chip, hence any perf/mm analysis is superficial.

 

[Glo.]

I still can't believe that VR is going to be that big of a market.

I can't believe a huge number of people are going to wear those things on their head for very long.

VR does not have to be done through headset. It is only for display. VR is also AI, physics, ray tracing, thousands of things and features that combined add up to the experience.

VR is not only gaming. It is EVERY THING. Education, entertainment, films, books, games, cars, houses, absolutely everything.

Even if Nvidia advertises their hardware as being VR capable it is not really capable in what people are currently thinking about VR. If you want to game on your hardware, than yes Nvidia may be good option. But for everything else? When you add Deep Learning, graphics, compute, huge data analysis at the same time?

I have said many times, that DirectX12 and Vulkan have came to life just for gaming purposes in VR. And we perfectly well know which hardware currently benefits the most from DX12.

One thing more. You know what is most important thing for VR? Compute performance of your GPUs. That is the factor that will determine the performance in your VR experience. If you want to judge a GPU don't look anywhere else, but for compute performance.

 

[linkgoron]

Inteluser is right to be concerned. Based on current leaked die sizes, it looks like AMD is falling further behind this generation than last. Yes, as bizarre as it sounds, it appears that Hawaii vs Maxwell was more favorable to AMD than Polaris vs Pascal is on performance per mm. If P10 is 232mm, then compare it to 390X. Same ratio as GP104 to GM200, yet the 1080 stands tall over the Titan X and based on leaks 480 barely edges out the 390X. AMD is losing ground if the leaks are true!

Correct me if I'm wrong, but if the 480 is a salvaged chip (i.e. not full Polaris), that 232mm isn't the "real" chip size.
It might be "effectively" just 200mm or 210mm?

 

[Headfoot]

Inteluser is right to be concerned. Based on current leaked die sizes, it looks like AMD is falling further behind this generation than last. Yes, as bizarre as it sounds, it appears that Hawaii vs Maxwell was more favorable to AMD than Polaris vs Pascal is on performance per mm. If P10 is 232mm, then compare it to 390X. Same ratio as GP104 to GM200, yet the 1080 stands tall over the Titan X and based on leaks 480 barely edges out the 390X. AMD is losing ground if the leaks are true!

Sure they can price it to be competitive, and isn’t that all that matters? Well no, not if you are the high end user hoping for competition. Die sizes have limit, die sizes determine profit. It means AMD will continue to fall further behind in revenue and can only compete by lowering margins on more expensive chips. Bad news for the long run. And unless Nvidia allows AMD to make much bigger chips (the true folly of AMD from 2007-2011... insanity) then they cannot compete for the crown. Unless Vega really benefits from HBM2 (it is a good advantage in die size, wattage, and bandwidth... this is the only hope), then it looks like Nvidia will have the high ground again, and possibly on a smaller die. It also looks like GP106 will easily compete with Polaris, and on a smaller die too.

Hawaii is going down as one of the top 5 best GPUs... It was the Fury line that was disappointing

 

[crisium]

Correct me if I'm wrong, but if the 480 is a salvaged chip (i.e. not full Polaris), that 232mm isn't the "real" chip size.
It might be "effectively" just 200mm or 210mm?

We don't know if it is cut. It may be, but there is currently no evidence of such. 36CU being "unusual" and "full Tonga went to Apple for a year" is not evidence, but does give hope.

But what if it is 2560SP full? That'd make 480 10% cut. Well, a 25% cut 1070 still essentially ties the Titan X. Unless 480 overclocks a lot more, which is possible, then it still looks like AMD made smaller gains jumping nodes than Nvidia. The other possibility is the rumours of 2816SP or even the dream of 3072SP, which would make me take back everything and AMD is in great shape.

No evidence though.

 

[swilli89]

Hawaii is going down as one of the top 5 best GPUs... It was the Fury line that was disappointing

Easily. The 290x/390x basically got a second wind due to game code catching up to its architecture as well as excellent support from the AMD driver team; all at a time when its original contemporaries (780 and 780ti) started to do worse and worse while Hawaii managed to keep up and then even start to surpass an entirely new generation from nVidia, Maxwell.

The real question that still lingers in my mind is how high Polaris 10 likes to clock. This is a heavily modified architecture on a much better clocking node. I would have to imagine that a ~232mm chip will want to clock up considering that a bigger Tahiti on early 28nm already could hit Poalris 10 base clocks when max OC'd.

 

[Paul98]

Easily. The 290x/390x basically got a second wind due to game code catching up to its architecture as well as excellent support from the AMD driver team; all at a time when its original contemporaries (780 and 780ti) started to do worse and worse while Hawaii managed to keep up and then even start to surpass an entirely new generation from nVidia, Maxwell.

The real question that still lingers in my mind is how high Polaris 10 likes to clock. This is a heavily modified architecture on a much better clocking node. I would have to imagine that a ~232mm chip will want to clock up considering that a bigger Tahiti on early 28nm already could hit Poalris 10 base clocks when max OC'd.

Yeah I am wondering if this is similar to what we saw at stock speeds for the HD 78xx and 79xx with 40% overclock headroom. But I am just waiting for real info on this now before I start speculating much at all. Or draw any conclusions

 

[Headfoot]

I would be very surprised if AMD used 100% of the headroom from 14nmFF on power reduction and none of it on frequency gain. The only way I could see that is if they dramatically beefed up per-shader average throughput and used the 14nmFF headroom to combat power creep / frequency drop from that. But still, I can't imagine that would consume all the headroom unless it was truly a massive change

 

[Stuka87]

Inteluser is right to be concerned. Based on current leaked die sizes, it looks like AMD is falling further behind this generation than last. Yes, as bizarre as it sounds, it appears that Hawaii vs Maxwell was more favorable to AMD than Polaris vs Pascal is on performance per mm. If P10 is 232mm, then compare it to 390X. Same ratio as GP104 to GM200, yet the 1080 stands tall over the Titan X and based on leaks 480 barely edges out the 390X. AMD is losing ground if the leaks are true!

Sure they can price it to be competitive, and isn’t that all that matters? Well no, not if you are the high end user hoping for competition. Die sizes have limit, die sizes determine profit. It means AMD will continue to fall further behind in revenue and can only compete by lowering margins on more expensive chips. Bad news for the long run. And unless Nvidia allows AMD to make much bigger chips (the true folly of AMD from 2007-2011... insanity) then they cannot compete for the crown. Unless Vega really benefits from HBM2 (it is a good advantage in die size, wattage, and bandwidth... this is the only hope), then it looks like Nvidia will have the high ground again, and possibly on a smaller die. It also looks like GP106 will easily compete with Polaris, and on a smaller die too.

First off, "Polaris vs Pascal" is a very poor comparison. Since Pascal is an architecture that has numerous chips in it. Pascal is the model for a volume model GPU. The correct comparison would be GCN4 vs Pascal.

How is AMD falling farther behind than last when nVidia has not even announced the competitor to Polaris yet? Not sure how you can say an as yet unannounced chip will "easily compete" with an as yet shipped chip.

And how is nVidia going to not allow AMD to make whatever they want? nVidia has ZERO control over what AMD does.

 

[Glo.]

Polaris 10 and 11 Device ID's are already in macOS Sierra Kext's.

 

[crisium]

First off, "Polaris vs Pascal" is a very poor comparison. Since Pascal is an architecture that has numerous chips in it. Pascal is the model for a volume model GPU. The correct comparison would be GCN4 vs Pascal.

How is AMD falling farther behind than last when nVidia has not even announced the competitor to Polaris yet? Not sure how you can say an as yet unannounced chip will "easily compete" with an as yet shipped chip.

And how is nVidia going to not allow AMD to make whatever they want? nVidia has ZERO control over what AMD does.

Did you do the die size and performance comparisons?

Keep in mind Polaris specs and performance are not set in stone. I'm making the assumption of 232mm for speculation purposes.

Nvidia was ahead in every metric last gen. Performance per mm, per watt, per TFLOP, and memory compression.

Inteluser is discussing performance per TLFOP. All of these are wishy washy based on speculation and rumours, but if that leaked 2304SP 480 is running at 1266Hz and is barely faster than the 2816SP 1050MHz 390X, then AMD has not improved at all there. Complete stagnation.

I'm also looking at performance per mm. Nvidia is already ahead here on 28nm, and Polaris 10 vs Hawaii and GP104 vs GM200 comparisons show a very similar ratio, again based on rumours only. Therefore if Polaris 10 does not have the same lead over Hawaii that GP104 has over GM200, based on current rumours, it is easy to speculate they are falling further behind here.

What other comparison can you make? Convince me they haven't lost ground, I have Freesync so I'd love to see see good high end competition. If you're not interested in discussing the rumours and want to believe something else is coming that is fair, but I'm curious as to what you are observing.

 

[Glo.]

I'm also looking at performance per mm. Nvidia is already ahead here on 28nm, and Polaris 10 vs Hawaii and GP104 vs GM200 comparisons show a very similar ratio, again based on rumours only. Therefore if Polaris 10 does not have the same lead over Hawaii that GP104 has over GM200, based on current rumours, it is easy to speculate they are falling further behind here.

Depends where you look. If you look at DX11 then yes - Nvidia wins. If you look at DX12 and compute power - AMD wins.

438mm2 die has roughly the same compute power as 601mm2 die.

Why compute? Because for DX12 and for VR this is most important metric.

 

[crisium]

True. Much more optimistic for AMD on the DX12 front.

 

[Glo.]

vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67c0 ("Ellesmere [Polaris10]")
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67c1
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67c2
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67c4
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67c7
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67c8
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67c9
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67ca
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67cc
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67cf
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67df ("Ellesmere [Radeon RX 480]")
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67e0 ("Baffin [Polaris11]")
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67e1 ("Baffin [Polaris11]")
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67e3
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67e7
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67e8 ("Baffin [Polaris11]")
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67e9 ("Baffin [Polaris11]")
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67eb ("Baffin [Polaris11]")
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67ef
vendor: 1002 ("Advanced Micro Devices, Inc. [AMD/ATI]"), device: 67ff ("Baffin [Polaris11]")

This is complete list of Device ID's for Polaris architecture.