[Benchlife] R9 480 (Polaris 10 >100w), R9

[Silverforce11]

It's not going to happen because they have 4 chips, Polaris 11/10 and Vega 11/10. So there's already 4 tiers of performance and power usage.

Btw, 1x 6pin (75W) in a GPU can hit 150W due to 75W from the MB. A single 1x 8pin (150W) is good to 225W for the same reason.

Polaris 10 is a Pitcairn replacement, it's a mainstream & small chip. 7850/7870 through and through. If it can get anywhere near Fury X, that's a huge achievement. It's target is the 390X.

Similar to Pascal GP106, it's target is the 980. It's not normal for a jump to be two tiers, ie. Polaris 10 = Fury X, or GP106 = 980Ti.

Edit: Not saying it cannot happen, but it would mean a miracle architecture, the second coming of Jesus in silicon.

 

[Adored]

The 7870 was dual 6-pin, the 7850 was single. Any 7850 I ever saw pulled under 100W though and I doubt the 7870 went above 150W even though it had 2 6-pins.

 

[Silverforce11]

http://arstechnica.com/gaming/2016/...aming-processors-coming-soon-is-one-for-xbox/

^ From AMD's conference call, Lisa Su emphasized the semi-custom wins in the gaming business.

All three new consoles. Some are ramping in Q2, ready for a Q3 release. I think this is the NX first, then Neo as the time-line from Sony's documents suggests October.

The current consoles will still be sold and not phased out immediately. Xbox360 only stopped being produced recently. This means they still have 28nm APU production for these consoles and so the income stated above is for the new consoles on top.

 

[AtenRa]

Just a thought,

Due to double patterning at the M1 and M2 metals on the 14nm process, we may not see the largest possible density from the first 14nm products (Polaris). AMD may not use more than one or two M1 layers in order to lower cost and that will decrease Density. So Polaris 11 and 10 may not have the highest density possible.

 

[jpiniero]

Keep in mind that on the MBP, Apple only uses discrete GPUs on the upgraded 15" models. The 27" iMac does come with a dGPU by default but I have to think that might change if they use Skylake 4+4e on it and only make it available on the upgraded model.

The current consoles will still be sold and not phased out immediately.

I think the expectation should be that the current PS4 won't be phased out at all. Sony's probably going to sell the current PS4 for cheaper, maybe $299 and they will have two models going forward.

 

[airfathaaaaa]

and how do we know they have 4 chips? for all we know they can pull another 7xxx and offer one desing with many disabled features and then "leak" bios to enable them 1/1.5 year after their release giving them a free boost...

 

[Mopetar]

I think the top end Polaris 10 of this year will wind up between a 390X and the FuryX.

This is probably the safest bet. When the showed off Polaris 10 they had it running Hitman at a frame locked 60 fps at WQHD (2560 x 1440) which suggests it's got better performance than the 390X.

Fury X has 65 FPS in one benchmark (although another had it at 72.3 FPS) so depending on how much extra power AMD is hiding behind that cap Polaris 10 could be competitive with Fury X (at least at 1080 and 1440), but I'll bet it falls a bit short of that in most games and is most comparable to just the Fury.

That makes me think that Polaris 10 gets released as the 480 and 470. What I don't know is whether or not the 480 will be a full chip, and I suspect not just for the sake of yields. Another reason it gets the 480 brand is that until they release Vega they'll be up against NV's 1070 and unless AMD really hit it out of the park while NV simultaneously craps the bed, NV will have the performance edge with the 1070 or whatever their supposed ~300 mm^2 chip ends up getting called. Vega 11 will likely be the 490 that's designed to go up against the 1070 and Vega 10 will be the Fury replacement that needs to compete against the 1080.

I'm also thinking we see a 480X about half a year from the Polaris 10 launch where it uses GDDR5X and potentially is a full Polaris 10 chip or just the best binned Polaris 10 chips. I think that this is the chip where we'll see better than Fury X performance from Polaris 10. It also gives AMD a good way to hit a price point between the 480 and 490 so that they've got a something in just about every bracket by early-mid 2017.

Hard to see how it all plays out with NV keeping their cards so close to their chest, whether it's because they won't be shipping as soon or because they're being especially coy, but they don't have as much as AMD to prove right now so I can't blame them for staying quiet, if for no other reason than to keep AMD in dark as much as possible.

 

[swilli89]

This is probably the safest bet. When the showed off Polaris 10 they had it running Hitman at a frame locked 60 fps at WQHD (2560 x 1440) which suggests it's got better performance than the 390X.

Fury X has 65 FPS in one benchmark (although another had it at 72.3 FPS) so depending on how much extra power AMD is hiding behind that cap Polaris 10 could be competitive with Fury X (at least at 1080 and 1440), but I'll bet it falls a bit short of that in most games and is most comparable to just the Fury.

That makes me think that Polaris 10 gets released as the 480 and 470. What I don't know is whether or not the 480 will be a full chip, and I suspect not just for the sake of yields. Another reason it gets the 480 brand is that until they release Vega they'll be up against NV's 1070 and unless AMD really hit it out of the park while NV simultaneously craps the bed, NV will have the performance edge with the 1070 or whatever their supposed ~300 mm^2 chip ends up getting called. Vega 11 will likely be the 490 that's designed to go up against the 1070 and Vega 10 will be the Fury replacement that needs to compete against the 1080.

I'm also thinking we see a 480X about half a year from the Polaris 10 launch where it uses GDDR5X and potentially is a full Polaris 10 chip or just the best binned Polaris 10 chips. I think that this is the chip where we'll see better than Fury X performance from Polaris 10. It also gives AMD a good way to hit a price point between the 480 and 490 so that they've got a something in just about every bracket by early-mid 2017.

Hard to see how it all plays out with NV keeping their cards so close to their chest, whether it's because they won't be shipping as soon or because they're being especially coy, but they don't have as much as AMD to prove right now so I can't blame them for staying quiet, if for no other reason than to keep AMD in dark as much as possible.

Well remember this, for a (assuming) larger chip, Tahiti on 28nm, AMD released a fully enabled chip right out of the gate. Now with a much more conservative design I expect AMD will release the full Polaris 10 to begin with as well. AMD may have played it too conservative with die size, but one of the pro's of this is that they may not have to cut any dies down for yield purposes.

 

[R0H1T]

Well remember this, for a (assuming) larger chip, Tahiti on 28nm, AMD released a fully enabled chip right out of the gate. Now with a much more conservative design I expect AMD will release the full Polaris 10 to begin with as well. AMD may have played it too conservative with die size, but one of the pro's of this is that they may not have to cut any dies down for yield purposes.

They don't have to disable anything, not unless the yields are horrific, they could just rinse & repeat the 7970 & its Ghz edition gig or the other one from half a decade back ~ http://anandtech.com/print/2556

With 14nm FF they should have enough legroom to release an updated variant with 30~40% higher clocks, if they go full on retard mode.

 

[swilli89]

They don't have to disable anything, not unless the yields are horrific, they could just rinse & repeat the 7970 & its Ghz edition gig or the other one from half a decade back ~ http://anandtech.com/print/2556

With 14nm FF they should have enough legroom to release an updated variant with 30~40% higher clocks, if they go full on retard mode.

I actually said something similar earlier in this very thread. I think we are in for a 4870 repeat, a true value game changer. AMD will be able to release a full die clocked higher than we are expecting and will also be enjoying the benefits of DX 12 + console synergies.

 

[Mopetar]

Well remember this, for a (assuming) larger chip, Tahiti on 28nm, AMD released a fully enabled chip right out of the gate. Now with a much more conservative design I expect AMD will release the full Polaris 10 to begin with as well. AMD may have played it too conservative with die size, but one of the pro's of this is that they may not have to cut any dies down for yield purposes.

That assumes the two process nodes behave the same which we already know isn't likely. Apple had to split production between TSMC and Samsung because the yields weren't good enough for their SoCs which are much smaller than Polaris 10.

If it was viable for NV or AMD to come out of the gate swinging with the big chips, they'd be doing it.

 

[swilli89]

That assumes the two process nodes behave the same which we already know isn't likely. Apple had to split production between TSMC and Samsung because the yields weren't good enough for their SoCs which are much smaller than Polaris 10.

If it was viable for NV or AMD to come out of the gate swinging with the big chips, they'd be doing it.

Well, nvidia did exactly that with a die size of about 316mm^2. Also important to note that Apple made that decision a full year before AMD's production schedule which makes a massive difference. At ~230mm and almost unlimited fab capacity, on a year+ mature node (14nm LPP is merely a refinement of LPE, many of the lessons learned will still apply), and they are most definitely shipping uncut dies on day 1.

 

[Mopetar]

Well, nvidia did exactly that with a die size of about 316mm^2. Also important to note that Apple made that decision a full year before AMD's production schedule which makes a massive difference. At ~230mm and almost unlimited fab capacity, on a year+ mature node (14nm LPP is merely a refinement of LPE, many of the lessons learned will still apply), and they are most definitely shipping uncut dies on day 1.

316 mm^2 isn't the big die though, and we know that Polaris 10 isn't AMD's big die either and Vega 10 is expected to be around 460 mm^2. The only chip NV has officially unveiled is the P100 which is nearly double the size at 600 mm^2 and I believe they said they were disabling a few SMs to improve yields.

We very well may get some full dies on day from both AMD and NV, but it wouldn't surprise me if we see some harvested versions of those chips as well with disabled SMs or CUs, if only because the demand is going to be strong.

The first products are mid-range, probably because the yields for anything bigger aren't as good. Otherwise we'd be getting the big chips first like we did in previous generations as you pointed out before, although both NV and AMD did release some smaller chips under the 40nm process for their previous generation before the big chips of the new generation, the big chips still came first, and usually had both full and cut versions available.

 

[hawtdawg]

Edit: Not saying it cannot happen, but it would mean a miracle architecture, the second coming of Jesus in silicon.

You're smoking some crack.

We're about to see the biggest relative node jump in GPU history. The previous node jump saw Nvidia and AMD release their mid-ranged GPU's first, just like they're about to do this time, with performance greater than the previous generation big GPU's.

Not only is it not even remotely crazy to think that a 135w 14nm GPU could be just as fast as a 275w 28nm GPU, it's actually what we should expect even if AMD didn't make so much as a single efficiency improvement from an architectural standpoint.

It's like you people forgot how GPU advancement used to work prior to getting stuck on 28nm.

 

[Silverforce11]

@hawtdawg

Polaris 10 isn't AMD's mid-range chip. This is where you made the mistake.

It's AMD's mainstream chip. It's like Pitcairn, not Tahiti.

AMD's mid-range chip is Vega 11. Their high-end will be Vega 10.

For Vega 11, I fully expect it will trash the current Fiji or GM200 SKU easily. For Polaris 10? That is a big ask. Maybe it can, under complex scenes, but it won't be the normal.

 

[hawtdawg]

@hawtdawg

Polaris 10 isn't AMD's mid-range chip. This is where you made the mistake.

It's AMD's mainstream chip. It's like Pitcairn, not Tahiti.

AMD's mid-range chip is Vega 11. Their high-end will be Vega 10.

For Vega 11, I fully expect it will trash the current Fiji or GM200 SKU easily. For Polaris 10? That is a big ask. Maybe it can, under complex scenes, but it won't be the normal.

Polaris 10 will be the Tahiti of 14nm. Vega 11 will probably be more like a Tonga.

Releasing a weak part would be suicide. They know good and well that Nvidia is going to be releasing something as good or better than a 980ti.

The claimed specs for Polaris 10 all fit the bill. a 2x perf/watt increase at 130 watts is right in the Fiji ballpark, and those perf/watt numbers should be attainable without even making an architecture change. Why on earth would it be so hard to believe that an optimized 2500ish shader part could hit Fiji numbers? Nvidia released a 1660 shader part that was as fast as the original Titan, and AMD hasn't done that type of redesign yet. Say what you will about their drivers, but AMD has always been just as capable at designing hardware.

 

[Silverforce11]

Polaris 10 isn't GP104 competitor (that's Vega 11's job), if the leaked prices are right, GP104 will be $649 and $499.

AMD have said Polaris 10 is aimed at below the 970/290 prices while giving higher performance.

 

[raghu78]

@hawtdawg

Polaris 10 isn't AMD's mid-range chip. This is where you made the mistake.

It's AMD's mainstream chip. It's like Pitcairn, not Tahiti.

AMD's mid-range chip is Vega 11. Their high-end will be Vega 10.

For Vega 11, I fully expect it will trash the current Fiji or GM200 SKU easily. For Polaris 10? That is a big ask. Maybe it can, under complex scenes, but it won't be the normal.

Silverforce it all boils down to the architecture. Just telling that Polaris 10 is a mid-range chip and thus cannot beat Fury X does not mean anything. Think about it this way if we are talking of 2.5x perf/watt gains for Polaris over 28nm GPUs then a Vega 10 at 250w will be atleast 2.25 times the perf of a Fury X at 275w. If thats the case I don't see why a Polaris 10 cannot beat a Fury X given that Vega 10 is most likely a doubling of Polaris 10 in terms of resources - 5120 sp, 80 CU, 8 shader engines, 128 ROPs . We know performance never perfectly doubles even when we double resources. So that means we are looking at Polaris 10 being 1.2-1.25 times Fury X perf if we take that Vega 10 is 1.8-1.9 times Polaris 10 performance.

Polaris 10 isn't GP104 competitor (that's Vega 11's job), if the leaked prices are right, GP104 will be $649 and $499.

AMD have said Polaris 10 is aimed at below the 970/290 prices while giving higher performance.

AMD said Polaris 10 is aimed at bringing VR to many more millions of gamers. That does not mean performance cannot be impressive. AMD seem to have 4 versions of Polaris 10 which allows a wide range of performance and pricing.

http://wccftech.com/amd-gcn-4-0-c99-flagship-polaris-rra-certification/

Anyway Raja Koduri stated that they are going to have GPUs across the entire performance stack

http://www.pcper.com/news/Graphics-...past-CrossFire-smaller-GPU-dies-HBM2-and-more

"There have been concerns that AMD was only going to go for the mainstream gaming market with Polaris but Raja promised me and our readers that we “would be really really pleased.” We expect to see Polaris-based GPUs across the entire performance stack."

 

[Silverforce11]

Silverforce it all boils down to the architecture. Just telling that Polaris 10 is a mid-range chip and thus cannot beat Fury X does not mean anything. Think about it this way if we are talking of 2.5x perf/watt gains for Polaris over 28nm GPUs then a Vega 10 at 250w will be atleast 2.25 times the perf of a Fury X at 275w. If thats the case I don't see why a Polaris 10 cannot beat a Fury X given that Vega 10 is most likely a doubling of Polaris 10 in terms of resources - 5120 sp, 80 CU, 8 shader engines, 128 ROPs . We know performance never perfectly doubles even when we double resources. So that means we are looking at Polaris 10 being 1.2-1.25 times Fury X perf if we take that Vega 10 is 1.8-1.9 times Polaris 10 performance.

I'm not saying it can't, nowhere did I say it cannot!

I've been saying it can, under some situations, like scenes with lots of overdraw due to the Polaris GCN's new features. I've been saying that since early this year with the reveal of Polaris. But to expect it to match Fury X in most games, is asking way too much for a very small chip.

Polaris 10 isn't mid-range. It's mainstream. AMD have said so already so people stop thinking it's mid-range.

Vega 11 is performance (mid-range), Vega 10 is enthusiast.

 

[Silverforce11]

https://youtu.be/J3IYhWVxCco?t=6m58s

^ AMD's Robert Hallock confirms Hitman was running 1440p on Polaris 10 with maxed settings, 60 fps locked, something Fury X cannot achieve at that area/view.

This is one of those scenarios where Polaris 10 can beat Fury X because it's not rendering the thousands of NPCs on that boat.

Some people will dismiss AMD's innovation here with the Primitive Discard Accelerator, labeling it as just another culling like Z-culling etc, they will be very wrong. All forms of software culling takes resources to perform since the objects have to enter the main rendering pipeline and get processed to determine whether they are visible or not first. DICE's approach uses compute shaders to calculate data for culling, it takes shader resources to do.

A hardware accelerator that goes before the main rendering pipeline and auto discards non-visible objects is revolutionary.

 

[raghu78]

I'm not saying it can't, nowhere did I say it cannot!

I've been saying it can, under some situations, like scenes with lots of overdraw due to the Polaris GCN's new features. I've been saying that since early this year with the reveal of Polaris. But to expect it to match Fury X in most games, is asking way too much for a very small chip.

Polaris 10 isn't mid-range. It's mainstream. AMD have said so already so people stop thinking it's mid-range.

Vega 11 is performance (mid-range), Vega 10 is enthusiast.

When did die size become an indicator of performance when comparing GPUs across process nodes and architectures. btw you like others keep confusing performance and pricing. You could have excellent perf at mainstream prices. Polaris 10 seems to have 4 models from the recent leaks I mentioned. That gives them a wide range of performance and pricing. Anyway I am optimistic that Polaris will be able to impress and surprise everyone with its perf and pricing.

 

[kraatus77]

https://youtu.be/J3IYhWVxCco?t=6m58s

^ AMD's Robert Hallock confirms Hitman was running 1440p on Polaris 10 with maxed settings, 60 fps locked, something Fury X cannot achieve at that area/view.

This is one of those scenarios where Polaris 10 can beat Fury X because it's not rendering the thousands of NPCs on that boat.

Some people will dismiss AMD's innovation here with the Primitive Discard Accelerator, labeling it as just another culling like Z-culling etc, they will be very wrong. All forms of software culling takes resources to perform since the objects have to enter the main rendering pipeline and get processed to determine whether they are visible or not first. DICE's approach uses compute shaders to calculate data for culling, it takes shader resources to do.

A hardware accelerator that goes before the main rendering pipeline and auto discards non-visible objects is revolutionary.

Remember overclocker's dream ? yeah. nobody can trust that guy again. i seriously don't get amd fanboys, instead of waiting they always overhype every little piece of info. and when the product doesn't perform "wait for drivers hurr durr'' i mean. can't you guys just wait for release and stop hyping ?


This crap is doing more brand damage than nvidia fanboys doo with their no.1 motto " amd drivers suck"

 

[tential]

People harp on memory bandwidth and die size exclusively in these conversations and just seem to ignore architecture entirely. I truly don't understand these Pascal/Polaris discussions. Without understanding the architecture, random comments on diesize/memory bandwidth are meaningless.

 

[IntelUser2000]

It's like you people forgot how GPU advancement used to work prior to getting stuck on 28nm.

That's true for CPUs as well. What happened with Intel CPUs before 22nm? The biggest hyped process became the lowest gain in years.

Remember Fury's HBM? We've seen that it was a necessity just to be able to compete with Nvidia, rather than jump forward.

I think the companies are becoming better and better at hyping technologies. I don't think many of us expected both Nvidia and AMD were planning a staggered launch.

So that means we are looking at Polaris 10 being 1.2-1.25 times Fury X perf if we take that Vega 10 is 1.8-1.9 times Polaris 10 performance.

Isn't it clear by now that 2.5x perf/watt gain is over Tonga? And its 2x compared to current generation? We don't even know that will apply uniformly to all Polaris products. Also don't forget 14nm only applies to the computing part of the video card. It certainly doesn't apply to memory, and even I/Os on the chip gain far less. That's marketing for ya. I can't believe that top Polaris will be faster if at all than Fury X if they justified on releasing $1599 Radeon Pro Duo. A 20-25% gain would be too close to one that would cost 3-4x and get 80% gain situationally. Unless they make not-the-smartest decision.

$1599 Radeon Pro Duo
$400 Polaris 10

If Polaris 10 performs better than Fury X, what goes in between $1599 and $400?

 

[poofyhairguy]

People harp on memory bandwidth and die size exclusively in these conversations and just seem to ignore architecture entirely. I truly don't understand these Pascal/Polaris discussions. Without understanding the architecture, random comments on diesize/memory bandwidth are meaningless.

The problem is twofold:



1. There is a pent up desire for cards on a new node.



2. People want to ignore how long we will be on the next node.