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

[R0H1T]

Figuring in the difference in HBM vs. GDDR5 could make the overall card a bit less than 2.5 even if the GPU does it.

There is a very good chance that if the Polaris' are released with standard GDDR5 then we'll get to see better clocked &/or higher binned versions being (re)released with GDDR5x next year, kinda like 7970Ghz edition. Ideally the 2.5x efficiency gain number should have included some reference to the type of memory it'd get, since that makes a huge difference to the overall figures, therefore worst case its 2x perf/W for GDDR5 laden cards & up to 3x with HBM2 but that's just my humble guesstimate

 

[Silverforce11]

Hold on, so now we are supposed to care about 1080p performance? Come on Silver.

That's where the majority who buys these GPUs play at. AMD have said they are pitching Polaris for the perf per dollar crowd, it will be cheaper than you think.

 

[AtenRa]

@ RussianSensation

Pro Duo is made for 4K and VR, period

Polaris 10 is made for 1080/1440p and VR, period. (this is the vast majority of todays Gamers)
At those resolution it will be close in some/faster in others than Fiji. period

Polaris 11 is Mobile first and entry Desktop. It will bring console performance to Laptops at affordable price and low TDPs.

ps. Polaris 10 is not a Fury X replacement like GP104 over GM200 that some people suggest.

 

[Silverforce11]

We have no idea the level of updates in these GPU architectures but if Polaris is in the new NX console, that's huge... I mean, we've seen the console effect before and how that has benefited Hawaii, having that for a NEW product like Polaris in which the Xbox One and PS4 also refresh with potentially Polaris chips as well, that would be massive.

http://www.gamasutra.com/view/news/254327/Nintendo_joins_graphics_standards_body_Khronos_Group.php

^ Strong indication combined with other dev leaks that Nintendo NX is Polaris based and also using Vulkan as it's programming API.

Apparently word on the street is that Nintendo will make a small loss per NX, because they are paying a high price for the 14nm FF AMD SOC. Early dev reports that current PS4 games can be direct ported to the NX and it runs it barely loaded, basically the hardware is miles ahead.

If true, and I hope it is, it means Nintendo is serious about it's console platform not just being gimmick based but serious in performance.

Now it makes me concern about Polaris for PC, with a push for notebooks and the NX competing for 14nm wafers and all..

 

[ShintaiDK]

Polaris 10 to be named 480, around Fury X performance, 130 watt

Benchlife says no such thing. No wonder you didn't link a source.

https://benchlife.info/computex-2016-pre-show-with-amd-arm-intel-nvidia-04142016/

 

[JDG1980]

Interesting that the new cards are said to be R9 470 and R9 480. No "X" suffix. That indicates to me that we may be seeing only cut-down SKUs at first, much like we did with Tonga. The rumors so far have Polaris 10 with 2560 shaders on-die, but only 2304 enabled on the card (four CUs disabled). Polaris 11 is often speculated to have 1024 shaders, but it wouldn't surprise me at all if it actually had 1152 or 1280 on-die, and the first retail cards disabled two to four CUs. The best (fully-enabled) bins could be reserved for higher margin mobile products, like this year's 15" Retina MacBook Pro.

Later this year, we could see the R9 470X and R9 480X, with fully enabled GPUs once yields improve. Maybe GDDR5X as well. Presumably, the Tahiti-sized Vega GPU will be called the R9 490(X) when that comes to market near the end of 2016 or beginning of 2017, and the big (Hawaii-sized?) Vega will get a special name, either some variation of Fury or something new.

As for pricing, both the names and chip sizes may offer some clue. The current R7 370 card debuted at $149, and the R9 380 at $199. Back in 2012, the 7750 (cut-down 123mm^2 Cape Verde) started at $109, while the 7850 (cut-down 212mm^2 Pitcairn) began at $249. Given these precedents, I wouldn't be surprised to see the R9 370 at $149 and the R9 380 at $249.

 

[JDG1980]

Now it makes me concern about Polaris for PC, with a push for notebooks and the NX competing for 14nm wafers and all..

I wouldn't worry too much. That's what binning is for; the top bins will go to premium notebook SKUs while the lower bins go to desktops.

Besides, while Nvidia has to compete with Apple for TSMC wafers, GloFo will have AMD as their primary customer. That stupid WSA might actually help AMD a bit for once this time.

 

[Silverforce11]

Interesting that the new cards are said to be R9 470 and R9 480. No "X" suffix. That indicates to me that we may be seeing only cut-down SKUs at first, much like we did with Tonga. The rumors so far have Polaris 10 with 2560 shaders on-die, but only 2304 enabled on the card (four CUs disabled). Polaris 11 is often speculated to have 1024 shaders, but it wouldn't surprise me at all if it actually had 1152 or 1280 on-die, and the first retail cards disabled two to four CUs. The best (fully-enabled) bins could be reserved for higher margin mobile products, like this year's 15" Retina MacBook Pro.

That makes the most sense. Apple's refresh will likely take all the full dies coming out of GF. Both Polaris 11 and 10. Then AMD actually has won some notebook designs for these chips as well.

Not looking so hot for a PC lineup. Hopefully they have enough production so there's full dies in 480X.

 

[antihelten]

Benchlife says no such thing. No wonder you didn't link a source.

https://benchlife.info/computex-2016-pre-show-with-amd-arm-intel-nvidia-04142016/

Just to expand on this, this is what Benchlife said (google translate):

In addition to the Computex AMD will announce Bristol Ridge, is the 7th generation APU, but also a continuation of the earlier topic, let Polaris for the first time show in front of everyone.

It is now known that information, Computex published Polaris will Baffin and Ellesmere, but no accident, then, Baffin is the Radeon R9 470, and Ellesmere is the Radeon R9 480 series. Reliable sources mentioned, Baffin consumes less than 50W, and Ellesmere is higher than 100W, about 110 ~ 135W, so the two series graphics cards do not belong to high-end products.

Other than that there is no info on polaris in the article, there is however a bit of info on Pascal, if anyones interested:

NVIDIA part is you can look forward to a new Pascal GPU architecture appearance. While the just-concluded GTC 2016, NVIDIA has announced the use of Pascal GPU architecture of Tesla P100, but that belongs to Deep Learning computing field, the average person can not touch the world. Comparison of the average consumer is looking forward to the arrival of GeForce graphics card, and this desire may debut in Computex 2016.

If not unexpected, Computex 2016 will be at least a GeForce GTX 1080 with two GeForce GTX 1070 graphics card, which will be used to determine the GeForce GTX 1080 GDDR5X memory, as the GeForce GTX 1070 is biased selection GDDR5.

 

[airfathaaaaa]

That's where the majority who buys these GPUs play at. AMD have said they are pitching Polaris for the perf per dollar crowd, it will be cheaper than you think.

the majority who buys a top end card will most certenly use dsr/vsr if they dont have a big monitor

 

[Silverforce11]

the majority who buys a top end card will most certenly use dsr/vsr if they dont have a big monitor

If it's $299 that isn't top-end. That's basically low mid-range.

Raja and Roy have said several times, they are aiming at the 970/290 price segment for Polaris 10.

 

[airfathaaaaa]

If it's $299 that isn't top-end. That's basically low mid-range.

Raja and Roy have said several times, they are aiming at the 970/290 price segment for Polaris 10.

tbh if their perfomance is above that or even on par with a 390 its still a high end card...given how those cards are perfoming now

 

[maddie]

If it's $299 that isn't top-end. That's basically low mid-range.

Raja and Roy have said several times, they are aiming at the 970/290 price segment for Polaris 10.

Is that really the quote?

I thought they said they were bringing, with Polaris, the minimum VR specs to lower price points.

Translates into Polaris 10 being cheaper than 970/290 prices.

 

[maddie]

tbh if their perfomance is above that or even on par with a 390 its still a high end card...given how those cards are perfoming now

It won't be in relation to the entire 14nm product stack.

I really like how Koduri appears to be handling things. A very long term vision with good ideas on how to get there.

 

[airfathaaaaa]

It won't be in relation to the entire 14nm product stack.

I really like how Koduri appears to be handling things. A very long term vision with good ideas on how to get there.

till we see the full potential of the 14nm process we will need at least 2 gen's more not to mention that since we wont see the full line up till mid 17(?) people will still measure them with this gen cards

 

[antihelten]

Is that really the quote?

I thought they said they were bringing, with Polaris, the minimum VR specs to lower price points.

Translates into Polaris 10 being cheaper than 970/290 prices.

This is the actual quote (from here):

we can produce GPUs that will run the minimum spec of VR at a lower cost, in larger volume, consuming less power and running faster

In the same presentation AMD also said that the minimum spec was 290X/970 and that those cards retailed for $350.

Bringing the price down to $300 would be too small of a jump imho (especially since you can already get a 970 for less than $300). As such the maximum price point for such a product would be $250 imho.

So $250 and at least 290X/970 level performance (the AMD quote mentions "running faster", hence the "at least" part).

Honestly that is pretty much what I would expect from AMD (and Nvidia) anyway. Right now the closest thing to the $250 price point would be the 380X (MSRP of $230), so 290X/970 level performance would represent an improvement of 45-50%, which is pretty much what I would expect from a node jump.

 

[maddie]

till we see the full potential of the 14nm process we will need at least 2 gen's more not to mention that since we wont see the full line up till mid 17(?) people will still measure them with this gen cards

What I meant was that all of the GPU designs for this generation ( 1st attempt at 14nm) should be completed. It is within this range of performance that Polaris 10 is seen as mid-range and will be priced as such. The 2 Vega designs will probably be the high end and premium GPUs.

The Navi info (scalability) plus Koduri's statement that the small die strategy is the future for AMD plus interposers allow very large composite products. As some of us have been saying and Nvidia has partially shown with GP100 (most said Fiji interposer was the limit), there is no technical limit to the size of a bare interposer up to the wafer level. The limit is the connective pathways which as we have seen can be done in sections. You can go past 600 mm^2 GPUs in the future.

 

[antihelten]

As some of us have been saying and Nvidia has partially shown with GP100 (most said Fiji interposer was the limit), there is no technical limit to the size of a bare interposer up to the wafer level.

This isn't really true. With the current approaches used, the limit is the 1.5 times the reticle size. The reticle size for TSMC at 65nm is 26mm by 32mm or 832mm2. So 1.5 times the reticle size is 1248mm2.

TSMC have talked about moving past the 1.5 times limit, but that won't happen for some time (estimates are that it will coincide with the arrival of the 7nm node).

 

[maddie]

This isn't really true. With the current approaches used, the limit is the 1.5 times the reticle size. The reticle size for TSMC at 65nm is 26mm by 32mm or 832mm2. So 1.5 times the reticle size is 1248mm2.

TSMC have talked about moving past the 1.5 times limit, but that won't happen for some time (estimates are that it will coincide with the arrival of the 7nm node).

The reticle limit only applies to the projected mask area.

Let me ask this.

Can you project several masks onto the wafer in adjacent spots?
Can this total masked area exceed the individual masked areas?
Can you have interposer silicon left bare (unmasked)?
Can you slice the wafer anywhere you desire?

What conclusions result here?

It's not like the entire interposer is one big single masked area. It isn't.

 

[antihelten]

The reticle limit only applies to the projected mask area.

Let me ask this.

Can you project several masks onto the wafer in adjacent spots?
Can this total masked area exceed the individual masked areas?
Can you have interposer silicon left bare (unmasked)?
Can you slice the wafer anywhere you desire?

What conclusions result here?

It's not like the entire interposer is one big single masked area. It isn't.

Sure you could do that, but doing so whilst managing to have all the routing and metal layers line up, would be a pain in the ass, and as such is not really viable at the moment (it might be viable for product with extremely high selling prices where you're able to eat the cost, but that's not what we're talking about here).

And leaving interposer silicon bare is of limited use since your average GPU/CPU/HBM module generally only has a very small amount of their surface area that doesn't feature interconnections of some kind.

So yes if you don't care about actually using your interposer then you can make it as big as you want, but if you want an interposer that's actually functionally useful, then there's a limit to it's size.

 

[Slaughterem]

If it's $299 that isn't top-end. That's basically low mid-range.

Raja and Roy have said several times, they are aiming at the 970/290 price segment for Polaris 10.

Price points are of course driven by wafer yield. When you consider that A9 chip at Glo Flo is 96mm as compared to 104.5 at TSMC you can extrapolate that out to a 232 mm2 chip from Glo flo is comparable to a 252 mm2 TSMC chip. Then depending on each process yields from each foundry on a 300mm wafer this will lead to lower pricing on AMD products. Also consider this that a 610 mm2 GP100 would be a 558 mm2 Glo flo chip.
On a 300mm wafer 232 mm2 can yield 246 dies as opposed to 252 mm2 providing 225. I believe this is a significant advantage.

 

[dzoni2k2]

Price points are of course driven by wafer yield. When you consider that A9 chip at Glo Flo is 96mm as compared to 104.5 at TSMC you can extrapolate that out to a 232 mm2 chip from Glo flo is comparable to a 252 mm2 TSMC chip. Then depending on each process yields from each foundry on a 300mm wafer this will lead to lower pricing on AMD products. Also consider this that a 610 mm2 GP100 would be a 558 mm2 Glo flo chip.
On a 300mm wafer 232 mm2 can yield 246 dies as opposed to 252 mm2 providing 225. I believe this is a significant advantage.

As it stands right now, anything coming from GF is a huge plus for AMD because of the wafer supply agreement. You could almost say GF is a freebie.

In other words AMD would have to pay up even if they don't meet the agreed volume (which was always the case in the past).

 

[poofyhairguy]

Now it makes me concern about Polaris for PC, with a push for notebooks and the NX competing for 14nm wafers and all..

Honestly it makes sense, we all want new GPUs but laptops NEED new power efficient GPUs. Those old GCN cards are holding back Apple's expensive laptops.

I will take a cut Polaris 10 if it's cheap enough and it beats a 390x at 1080p (the resolution I game at because a PC is basically an Xbone on roids in my usage). The problem with the 285 wasn't that it wasn't a full Tonga, the problem was 2GB RAM.

 

[maddie]

As it stands right now, anything coming from GF is a huge plus for AMD because of the wafer supply agreement. You could almost say GF is a freebie.

In other words AMD would have to pay up even if they don't meet the agreed volume (which was always the case in the past).

That is actually a very good interpretation of the situation.

Any one knows the # wafers in the WSA?

 

[dzoni2k2]

That is actually a very good interpretation of the situation.

Any one knows the # wafers in the WSA?

Well 'freebie' up to the agreed volume of course. And like I said, if past is any indication, AMD usually agrees to more than they end up ordering. Quite a lot more actually. To the point it really hurt them financially and they were forced the amend the agreement many times.