[Ars] AMD confirms high-end Polaris GPU will be released in 2016

[3DVagabond]

The way AMD is talking about it the full Polaris 10 is going to be in that price range, not Polaris 11. I'll say $599 if it can get close to Fury X performance.

When did they say this?

That would be the point IF... AMD themselves said that they have to bring that kind of performance lower in the price/performance bracket.

This is what I read as well.

 

[jpiniero]

When did they say this?

The press conference that AMD had. They made it sound like (full?) Polaris 10 was going to be in the 950's price range which is like $150-$175. Given the costs of 14FF and AMD's desire to not be "the budget brand", I would not be surprised with $199 if it's closer to 380X performance rather than say 960.

 

[3DVagabond]

exactly. Nvidia will have a GP106 card faster than GTX 980 at 100w tdp. I think GP104 with 4096 cc and GDDR5X will be >2x faster than gtx 980 at 180w.

What are you basing nVidia's performance on? Besides the shrink the only improvements I've read about are HBM2/GDDR5X and NVLink.

Even though AMD is keeping the GCN naming almost everything in the chip is new. I would expect a bigger performance gain from AMD if I were speculating.

 

[Kenmitch]

What are you basing nVidia's performance on? Besides the shrink the only improvements I've read about are HBM2/GDDR5X and NVLink.

Even though AMD is keeping the GCN naming almost everything in the chip is new. I would expect a bigger performance gain from AMD if I were speculating.

Didn't Raja state it's the greatest leap so far?

 

[3DVagabond]

Comparing Fiji with GM204 is stupid. Underclocking and undervolting a 50% larger GPU to match the power efficiency of a smaller GPU is nothing great. Fiji must be compared to GM200 at its released clocks. If AMD want to compete they need to match Nvidia in terms of perf/sp vs perf/cc, perf/watt, perf/transistor and perf/sq mm. Thats when they can really compete otherwise Nvidia will be always calling the shots.

Perf/mm and perf/W yes. Perf/CC isn't critical. It's different uarch.

 

[3DVagabond]

This confirms what I posted earlier.

It will cost more to produce a GPU of any given performance on 28nm than on 14nm.

This should stop any further suggestions that AMD will reuse present designs. They will have to sell at a discount to the new gen because of power consumed and will also cost more to produce. No one is that incompetent.

Folks, the 2 GPU belief needs to be revisited.

They could have chips left over to use. In which case I believe they would do a rebrand.

 

[3DVagabond]

Didn't Raja state it's the greatest leap so far?

Without looking it up, I thought he was referring to efficiency. Which again bodes to more than simply a shrink.

 

[Head1985]

232mm2 is not enough.They need something to compete with GP104 witch will be around 300-320mm2 and also around 30-40% faster than TITANX.

 

[Kenmitch]

Without looking it up, I thought he was referring to efficiency. Which again bodes to more than simply a shrink.

Noooooooooooooooo! Perf/watt is cool unless it's gimped by watts.

Koduri: Yes. We have two versions of these FinFET GPUs. Both are extremely power efficient. This is Polaris 10 and that’s Polaris 11. In terms of what we’ve done at the high level, it’s our most revolutionary jump in performance so far. We’ve redesigned many blocks in our cores. We’ve redesigned the main processor, a new geometry processor, a completely new fourth-generation Graphics Core Next with a very high increase in performance. We have new multimedia cores, a new display engine.

From here http://venturebeat.com/2016/01/15/a...-to-full-graphics-immersion-with-16k-screens/

I'm leaning towards performance currently. Based on no performance per watt....Guess it might have been purposely omitted.

 

[raghu78]

Perf/mm and perf/W yes. Perf/CC isn't critical. It's different uarch.

I think AMD realized that shader efficiency and utilization was a major problem with Fury X. Thats why its a key focus area for Polaris. In the end Polaris and Pascal chips with 4096 cores will face off this year. So if AMD wants to compete they need to get on par in terms of perf/sp vs perf/cc.

My guess for specs
Polaris 11 - 4096 sp, 2048 bit HBM2, 128 ROPs, 8 GB, 512 GB/s
GP104 - 4096 cc, 256 bit GDDR5X, 128 ROPs, 8 GB, 448 GB/s.

My guess is both chips will end up close to 350 sq mm. So it boils down to who has the better architectural efficiency.

 

[Glo.]

I think AMD realized that shader efficiency and utilization was a major problem with Fury X. Thats why its a key focus area for Polaris. In the end Polaris and Pascal chips with 4096 cores will face off this year. So if AMD wants to compete they need to get on par in terms of perf/sp vs perf/cc.

My guess for specs
Polaris 11 - 4096 sp, 2048 bit HBM2, 128 ROPs, 8 GB, 512 GB/s
GP104 - 4096 cc, 256 bit GDDR5X, 128 ROPs, 8 GB, 448 GB/s.

My guess is both chips will end up close to 350 sq mm. So it boils down to who has the better architectural efficiency.

Charlie said that the Polaris GPUs are GDDR5 only.

 

[Head1985]

I think AMD realized that shader efficiency and utilization was a major problem with Fury X. Thats why its a key focus area for Polaris. In the end Polaris and Pascal chips with 4096 cores will face off this year. So if AMD wants to compete they need to get on par in terms of perf/sp vs perf/cc.

My guess for specs
Polaris 11 - 4096 sp, 2048 bit HBM2, 128 ROPs, 8 GB, 512 GB/s
GP104 - 4096 cc, 256 bit GDDR5X, 128 ROPs, 8 GB, 448 GB/s.

My guess is both chips will end up close to 350 sq mm. So it boils down to who has the better architectural efficiency.

Gp 104 wont be 4096SP GPU.GTX980 is 400mm2 and if 16nm have 2x density that GPU will be 400mm2 and that will never happen.

 

[raghu78]

Gp 104 wont be 4096SP GPU.GTX980 is 400mm2 and if 16nm have 2x density that GPU will be 400mm2 and that will never happen.

GM204 aka GTX 980 had a 256 bit GDDR5 bus. GP104 can be around 350 sq mm as they will need to use only a 256 bit GDDR5X bus instead of needing a 512 bit GDDR5 memory bus to feed twice the cores. btw GP104 will also have 8 GB GDDR5X capacity. btw GTX 980 had 2 MB L2 cache. SRAM (L2 cache) shrinks the best, followed by logic and I/O (PCI-E and memory controllers) has the least shrinkage. So yeah I think Nvidia can ship a GP104 for 350 sq mm.

 

[Silverforce11]

232mm2 is not enough.They need something to compete with GP104 witch will be around 300-320mm2 and also around 30-40% faster than TITANX.

That sounds like Polaris 11.

Should be around 390X performance, maybe 100W if we extrapolate from Polaris 10 (~120mm2).

At this performance level, GDDR5 is more than sufficient. Keep it cheap, 8GB, makes for a great VR entry level GPU that is affordable. Makes sense with Roy's VR speech, saying ~$339 for VR-capable GPUs (R290, 970) is still too high and they are aiming to deliver better than that with Polaris.

Where is this VEGA rumor from, is it a code name exposed in drivers?

Because they need a big one if they want to compete at the top-end.

 

[Kenmitch]

Where is this VEGA rumor from, is it a code name exposed in drivers?

http://wccftech.com/amd-future-gpus-named-stars-galaxies-company-bids-farewell-island-code-names/

 

[boozzer]

I will again pop back the post on SemiWiki: https://www.semiwiki.com/forum/content/3884-who-will-lead-10nm.html

GloFo/Samsung Process is 2.2 times denser than 28nm TSMC.

Lets think about it for a second. 600 mm2 Die from TSMC like Fiji would be... 250mm2 on 14 nm. It is extremely close to that 232 mm2 die size.

Whats more, if reality turns out to be true with what AMD says about the architecture, it is cleaned up, revamped and redesigned same GCN architecture. It would also affect the density of the chip. Coincidence? Of course, apart from my educated guesses/guesses I do not have anything to prove it.

But think about it. What better way for AMD would be bringing VR performance to Pitcairn price/performance bracket? They can also cut it to 3072 GCN cores, and still have VR GPU.

I will not talk about performance/SIMD because we do not know anything apart from words from Mahigan on that side, but we may be really surprised...

Imagine a 3072 GCN4 core GPU with 199$ price point matching in performance Titan X...

if this can become true, 2016 will really be a good year.

 

[Silverforce11]

I don't want to get too hyped up, because we still don't know how high they can clock it on these new FF processes. So far, its only mobile SOCs which are different.

 

[Elixer]

I don't want to get too hyped up, because we still don't know how high they can clock it on these new FF processes. So far, its only mobile SOCs which are different.

It isn't just that, yield issues are going to be a thing as well, until they get the process down.

 

[JDG1980]

I don't want to get too hyped up, because we still don't know how high they can clock it on these new FF processes. So far, its only mobile SOCs which are different.

Yes, a lot of it does depend on the architecture. But one of the main reasons why Maxwell did so much better than 28nm GCN is that it was able to clock a lot higher. AMD has to know that its current inability to rase core clocks much over 1000 MHz is a bottleneck, and I'd expect to see that addressed in the move to FinFET. And the mobile SoCs prove that FinFET can be used to enable significantly (>30%) higher clock speeds compared to planar. Of course the chips have to be designed with that in mind; this is true for every process node.

 

[3DVagabond]

I will again pop back the post on SemiWiki: https://www.semiwiki.com/forum/content/3884-who-will-lead-10nm.html

GloFo/Samsung Process is 2.2 times denser than 28nm TSMC.

Lets think about it for a second. 600 mm2 Die from TSMC like Fiji would be... 250mm2 on 14 nm. It is extremely close to that 232 mm2 die size.

Whats more, if reality turns out to be true with what AMD says about the architecture, it is cleaned up, revamped and redesigned same GCN architecture. It would also affect the density of the chip. Coincidence? Of course, apart from my educated guesses/guesses I do not have anything to prove it.

But think about it. What better way for AMD would be bringing VR performance to Pitcairn price/performance bracket? They can also cut it to 3072 GCN cores, and still have VR GPU.

I will not talk about performance/SIMD because we do not know anything apart from words from Mahigan on that side, but we may be really surprised...

Imagine a 3072 GCN4 core GPU with 199$ price point matching in performance Titan X...

if this can become true, 2016 will really be a good year.

Someone will quote this if it's not true, which it very likely won't be, as how AMD over hypes everything.

I don't want to get too hyped up, because we still don't know how high they can clock it on these new FF processes. So far, its only mobile SOCs which are different.

We know hardly anything. We know process, have an idea of perf/W and perf/clock of one Polaris chip @ 850MHz. A small ~120mm chip we/ve been told. We've discovered what is very likely evidence of a 232mm^2 chip. We don't know anything about it. We can speculate the CU's. We can speculate the perf at a similar clock to the smaller chip.

We have the statement of bringing VR level performance to a more affordable price point. We don't know whether that's an actual clue to Polaris perf/$, or simply a common sense strategy for VR to succeed.

I'm with you, I don't see Titan X perf for $199 inferred anywhere though.

 

[Silverforce11]

We know hardly anything. We know process, have an idea of perf/W and perf/clock of one Polaris chip @ 850MHz. A small ~120mm chip we/ve been told. We've discovered what is very likely evidence of a 232mm^2 chip. We don't know anything about it. We can speculate the CU's. We can speculate the perf at a similar clock to the smaller chip.

We have the statement of bringing VR level performance to a more affordable price point. We don't know whether that's an actual clue to Polaris perf/$, or simply a common sense strategy for VR to succeed.

I'm with you, I don't see Titan X perf for $199 inferred anywhere though.

I fully agree with that.

It was inferred 970/R290X or the minimum required for VR is in the $330 price bracket (too high according to AMD), and it's something AMD will address. I am certain (educated believe rather) Polaris 11 will be ~980/390X performance. Not sure about price.

Titan X perf is too much to ask for a 232mm2 chip.

Imagine if that holds, Polaris 11 ~390X performance for a little less, say $299... woooah.... ??

Not really. Custom R290X were pretty cheap last year. They run at 390X performance. We wouldn't be moving the perf/$ at all compared to what we had in 2014 and 2015!!

I still believe in the next-gen era, you still won't get better bang per buck compared to custom R290s.

 

[beginner99]

This confirms what I posted earlier.

It will cost more to produce a GPU of any given performance on 28nm than on 14nm.

This should stop any further suggestions that AMD will reuse present designs. They will have to sell at a discount to the new gen because of power consumed and will also cost more to produce. No one is that incompetent.

Folks, the 2 GPU belief needs to be revisited.

I don't think so. I expected the 2 polaris chips to replace 370x and 390x and then AMD has a full stack of GCN 1.2 or higher parts. Meaning they have the same features like video decode/encode, color compression and so forth.

But in the end all the 14nm hype will fall flat on it's face. Even if polaris 11 is bit faster than 390(x) and cost around $250 you still don't get a huge performance/$ increase compared to over a year ago when 290x inventory was cleared. You can get used 290(x) at $200 right now.

 

[flopper]

232mm2 is not enough.They need something to compete with GP104 witch will be around 300-320mm2 and also around 30-40% faster than TITANX.

Of course.
No reason to put everything out there early.
the 232mm seems to be a match for todays entusiast cards.
at a way lower tdp than before.
adding a 350mm or whatever size they can make and aim for seems plausible to replace furyx entusiast end.

 

[MrTeal]

I fully agree with that.

It was inferred 970/R290X or the minimum required for VR is in the $330 price bracket (too high according to AMD), and it's something AMD will address. I am certain (educated believe rather) Polaris 11 will be ~980/390X performance. Not sure about price.

Titan X perf is too much to ask for a 232mm2 chip.

Imagine if that holds, Polaris 11 ~390X performance for a little less, say $299... woooah.... ??

Not really. Custom R290X were pretty cheap last year. They run at 390X performance. We wouldn't be moving the perf/$ at all compared to what we had in 2014 and 2015!!

I still believe in the next-gen era, you still won't get better bang per buck compared to custom R290s.

Yeah, AMD isn't going to offer better perf/$ than the custom 290s that were available. It will likely increase decently relative to some of the poor options in AMD's lineup like Fury X, but nothing is going to touch the $220 R9 290s or $250 290X.

 

[JDG1980]

But in the end all the 14nm hype will fall flat on it's face. Even if polaris 11 is bit faster than 390(x) and cost around $250 you still don't get a huge performance/$ increase compared to over a year ago when 290x inventory was cleared. You can get used 290(x) at $200 right now.

Even so, that would still be a huge improvement if we got that kind of perf/$ across the board. It would mean GTX 980 Ti level performance at about $350. (The 980 Ti has about 40% better performance at 1080p than the 290X, and the 290X was being blown out at about $250 last year.)

And don't forget that all the new cards will have HDMI 2.0 output, hardware HEVC decoding support, and much lower power consumption. To many buyers, this makes a difference.