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

[railven]

Somewhere, there's someone holding onto this card and just waiting to sell it on eBay for $500.

No kidding. Had a family member looking to buy a cheap used video card. He linked me a Radeon 7500 and asked if it was worth the $75 (at the time the HD 7750 was going for about $100).

I was surprised things like this were permitted on Ebay. Then I remembered, Ebay.

EDIT:

Okay. That said, I don't think that a mid-range dual GPU is a good idea at all.

For a while, NV and partners thought the GTX 460 2WIN was a good idea.

Also, if it's a small enough die, it could take place of the old HD 4850X2/4870X2. Owning a HD 4870X2, they weren't bad designs/cards.

 

[stahlhart]

Get this thread back on topic.
-- stahlhart

 

[Head1985]

Everyone is going to be missing real high end for consumers until 2017 or later. nVidia did it twice (680 -> 780, 980 -> 980Ti) and AMD once, arguably twice (7970 -> 290, 290 -> Fiji maybe?).

Why does everyone think for some reason they're going to abandon the "delay high end chip due to long process lifespan product planning?" They wont. The initial release will be low and and mid range, marketed as mid range and high end.

Yeah GTX970 and GTX980 successors will be at 350-550USD with +- 300mm2 SKU and TDP 150-180w with 35-40% faster than TITANX.
AMD sure have something to counter it.They must have their own 300-350mm2 SKU to sell for 350-550USD just like NV.
Big pascal GPU with 500+mm2 SKU will be in 2017 and probably same on AMD side.

 

[MrTeal]

Yeah GTX970 and GTX980 successors will be at 350-550USD with +- 300mm2 SKU and TDP 150-180w with 35-40% faster than TITANX.
AMD sure have something to counter it.They must have their own 300-350mm2 SKU to sell for 350-550USD just like NV.
Big pascal GPU with 500+mm2 SKU will be in 2017 and probably same on AMD side.

Curious, a 300mm² die 35%-40% faster than Titan X, or than GM200? TX is in some ways a more crippled card than the 980Ti is, unless you unlock the TDP limit. Still, that would be a double the perf/watt of the 980 Ti and 2.7x the performance per area, with 1.4x better perf/$. I'd be ecstatic if it happened, but that's pretty optimistic. Even with the massive uarch improvements going from Fermi to Kepler, that transition had the 680 at 1.5x better perf/W, 2.2x the perf/area and 96% of the per/$ relative to the 580.

That's with nVidia having good scaling at the last transition. At 19x12, AMD got 1.47x better perf/W, 1.44x the perf/area, and 81% the perf/$ going from the 6970 to 7970. GCN1.0 ended up aging better than Kepler, but hopefully AMD can get that value proposition better this time.

 

[Head1985]

TSMC's 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed, around 2 times the density, or 70 percent less power than its 28HPM technology.

And with this they sure can manage shrink 398mm2 GTX980 to 300-320mm2 GPU with TDP 180w and with performance 40% above TITANX.
TITANX is only 40% faster than GTX980.So they need only 80% more performance.That is less than what they manage from GTX560TI to GTX680.

 

[MrTeal]

TSMC's 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed, around 2 times the density, or 70 percent less power than its 28HPM technology.

And with this they sure can manage shrink 398mm2 GTX980 to 300-320mm2 GPU with TDP 180w and with performance 40% above TITANX.
TITANX is only 40% faster than GTX980.So they need only 80% more performance.That is less than what they manage from GTX560TI to GTX680.

That's an or in there. Outside of any uarch improvements, if you shrank GM200 to 2 times the density you'd get it to 300mm², if you clocked it 65% higher you'd probably get 40% or more better performance (more, if you were entirely GPU limited). I don't think you get to do those two at the same time and enjoy a 70% power savings, as well.

Edit: At 19x12, the 680 was 70% faster at launch than the 560Ti, so 1.91x better perf/area, 1.52x better perf/W, and 71.4% of the perf/$
Same trend there, though worse than relative to the 580 because the 580 was more compute heavy. You get a lot more performance per area, better perf/W, and generally the same or worse perf/$. Generally the midrange are better perf/$ though, so comparing 680 to 560Ti exaggerates how bad perf/$ is going to the new gen.
Still a far cry from 2x the perf/watt of the 980 Ti, and 2.7x the perf/area, and 1.4x better perf/$. I could honestly maybe see the first one. The second might be a stretch. The third? Nope.

 

[Head1985]

But gm200 is bad example.It has TDP 250w and its only 30% faster than 165-180w GTX980
http://www.computerbase.de/2015-06/geforce-gtx-980-ti-test-nvidia-titan/4/
TPU review 29% faster.So that mine 40% was overreaction.
http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_Titan_X/29.html

So they need now only 70% more performance jump from GTX980 and it will be 40% faster than TITANX.
70% above GTX980 with new node should be easy with 180w TDP even with zero arch. changes.Only with node.

 

[Silverforce11]

Don't forget the emphasis of Maxwell was on gaming performance by sacrificing DP compute and mix mode compute throughput. This was a decision to give them better perf/w being stuck at 28nm, as their CEO said in the launch stream.

NV need new Teslas, HPC powerhouses and for Pascal they would have to go back to a compute focused design.

Whereas AMD's GCN is already a compute focused design and Polaris is just enhanced GCN in every-way.

And no, I am not expecting big chips in 2016 for consumers. I fully expect NV to aim for a big chip for HPC market to stomp on Intel's advance in the HPC arena. Yields will be horrible, but at ~$6000 a piece for high end Teslas (which they already have back orders from supercomputer builders!!), it's very profitable.

AMD has to be ready for a big chip battle in 2017 when presumably yields improve, and NV has fulfilled their HPC orders, with excess big Pascal to be put into a GTX SKU.

 

[Kenmitch]

Yeah GTX970 and GTX980 successors will be at 350-550USD with +- 300mm2 SKU and TDP 150-180w with 35-40% faster than TITANX.
AMD sure have something to counter it.They must have their own 300-350mm2 SKU to sell for 350-550USD just like NV.
Big pascal GPU with 500+mm2 SKU will be in 2017 and probably same on AMD side.

Sarcasm?

Your performance to dollars algorithm needs more tweaking if not....Unless hell freezes over.

Dwindling dGPU sales = maximize profits....Translates to milk it for all she's worth.

 

[Silverforce11]

Dwindling dGPU sales = maximize profits....Translates to milk it for all she's worth.

It's not even dwindling dGPU sales, it's actually been a really good 2015.

But they will still maximize profits anyway.

We already saw mid-range chips going for $550. I wouldn't be shock to see ~300m2 chips in SKUs going for $650 or even $750.

 

[Head1985]

ehm nope..They need 300-400USD GPU it will be cutdown version of that GPU(GTX980 successor).You really think GTX970 successor will cost 550USD?
Zero chance.

 

[MrTeal]

But gm200 is bad example.It has TDP 250w and its only 30% faster than 165-180w GTX980
http://www.computerbase.de/2015-06/geforce-gtx-980-ti-test-nvidia-titan/4/
TPU review 29% faster.So that mine 40% was overreaction.
http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_Titan_X/29.html

So they need now only 70% more performance jump from GTX980 and it will be 40% faster than TITANX.
70% above GTX980 with new node should be easy with 180w TDP even with zero arch. changes.Only with node.

Of course. It's always easier to scale up lower performance GPU than a big one. GM200 is 50% bigger in everything or more than GM204. If you did straight up compute on it, it probably would be 50% faster. Other factors come into effect in games though.

With 0 uarch changes, how would you scale up the 980 exactly? You could give it 50% more SMX, a 50% wider bus, 50% more memory. With perfect scaling that's 50% faster. Clock it 13% faster, and that's 1.7x in raw performance right there. Of course, you just described a Titan X at 1400 boost.

 

[3DVagabond]

You think a Fury Nano would be very profitable at $300?

How about when faced against NV's GP104 harvested SKU, the 970 successor, with ~100W vs Nano's 175W? Using much cheaper GDDR5 I bet too.

Suicidal to pit a 28nm product against next-gen uarch on 16nm FF. Exactly like they've been doing on the CPU side, out-dated uarch, out-dated node vs Intel.. didn't go well.

Not talking permanently. Just until they can come out with a proper 14nm solution. Besides, when is nVidia going to have 16nm Pascal? It may be a while?

 

[3DVagabond]

I'd still argue that AMD hasn't truly done it yet. Tahiti doesn't seem like something conceived as a midrange chip. If you're comparing die sizes, don't forget about AMD's small die strategy.

This time, however, they don't have a choice. All we can hope for (well, those of us who aren't Nvidia\Intel fanboys anyway) is that they go for Nvidia's throat and aim for a $500 price point for the high-end

I think they could do Polaris 11 as highend @ $550, their established price point. Then for the true highend 500-600 mm^2, they could design a truly premium product and go for the premium price. It would actually have to be special though. Maybe full cover block, pump integrated with the rad, an expandable loop, larger available rads and CPU/chipset blocks, premium PCB components, etc. IF it backs it's position with the performance charge $1K for it.

 

[Silverforce11]

ehm nope..They need 300-400USD GPU it will be cutdown version of that GPU(GTX980 successor).You really think GTX970 successor will cost 550USD?
Zero chance.

Nope, I'm expecting Pascal successor of the 970 is going to be ~$400 or more and not ~$330. When NV dropped the 970 at $330, we were all very surprised by the good bang for buck. You expect that treatment on a new node vs the 970 on a very mature 28nm?

 

[Arachnotronic]

I think they could do Polaris 11 as highend @ $550, their established price point. Then for the true highend 500-600 mm^2, they could design a truly premium product and go for the premium price. It would actually have to be special though. Maybe full cover block, pump integrated with the rad, an expandable loop, larger available rads and CPU/chipset blocks, premium PCB components, etc. IF it backs it's position with the performance charge $1K for it.

For the right performance level, $1k+ is fine by me. As long as I get what I pay for I'm cool.

 

[3DVagabond]

For the right performance level, $1k+ is fine by me. As long as I get what I pay for I'm cool.

Well nobody's been able to give us that so far and I really don't personally believe that it's possible to get "your money's worth" out of a $1000 gaming card. To each his own though.

 

[biostud]

Well nobody's been able to give us that so far and I really don't personally believe that it's possible to get "your money's worth" out of a $1000 gaming card. To each his own though.

The closet would be the 295x2 once the price was reduced to $999, but obviously not a single GPU card.

 

[Aristotelian]

I hope this time that the high end release (not the Titan series, but the -Ti series) date is similar for both AMD and Nvidia. I'm in a place to buy whichever comes first in order to complete my new build, but I'd like to actually see the competition not in terms of long term performance, but immediate high end releases on the new process competing with one another in price and performance. And I hope to buy two of the high end series from either company, put waterblocks on them, overclock to blazes, and catch up on some gaming...

 

[Krteq]

"Big Polaris" (or Vega) card spotted on zauba.

PCB with new code => 102-C993xx and unit cost is much higher then previous ones

Source: 3D Center (DE)

 

[Headfoot]

All AMD needs to do to stand out for enthusiasts is sell a bare card for watercooling folks, or even a binned bare card like eVGA did for their Kingpin 980ti (but naked, and less exorbitant). That and ensure an overclocking utility with overvolting is available day one, every time, no exceptions

 

[Kenmitch]

All AMD needs to do to stand out for enthusiasts is sell a bare card for watercooling folks, or even a binned bare card like eVGA did for their Kingpin 980ti (but naked). That and ensure an overclocking utility with overvolting is available day one, every time, no exceptions

The battle for the future depends more on past perceptions I'd think. Urban myth syndrome is the biggest hurdle I'd think.

 

[exar333]

Because ???

AMD still hasn't even executed on the Fiji x2 yet. Coming out of the gate with Polaris X2 seems highly unlikely...

 

[exar333]

Don't forget the emphasis of Maxwell was on gaming performance by sacrificing DP compute and mix mode compute throughput. This was a decision to give them better perf/w being stuck at 28nm, as their CEO said in the launch stream.

NV need new Teslas, HPC powerhouses and for Pascal they would have to go back to a compute focused design.

Whereas AMD's GCN is already a compute focused design and Polaris is just enhanced GCN in every-way.

And no, I am not expecting big chips in 2016 for consumers. I fully expect NV to aim for a big chip for HPC market to stomp on Intel's advance in the HPC arena. Yields will be horrible, but at ~$6000 a piece for high end Teslas (which they already have back orders from supercomputer builders!!), it's very profitable.

AMD has to be ready for a big chip battle in 2017 when presumably yields improve, and NV has fulfilled their HPC orders, with excess big Pascal to be put into a GTX SKU.

I don't disgaree with you on the focus of NV in the pro arean (Tesla, etc.). They desperately need a refreshed Tesla line. It likely will have been a few years since Kepler once Pascal gets out the door...

I do disagree with the lack of new 2016 consumer options. There probably hasn't been a time in the last ~10 years (since when the initial Crysis came out) when new hardware has been needed more. 4K displays will be front and center this year, along with UW 1440p displays and Oculus/VR setups. All of those (today) require multi-GPU to ideally function. 14/16nm is NEEDED NOW. The first company to have a compelling option will reap the benefits.

If there isn't plans to release high-perf parts in 2016, NV/AMD better be ready to slash their flagship prices and get folks to bite on multi-GPU setups or sales will dry-up by the end of the year. How many more 980GTIs or Furys can they sell at current prices? Can't be too much of a market left...

 

[Udgnim]

AMD still hasn't even executed on the Fiji x2 yet. Coming out of the gate with Polaris X2 seems highly unlikely...

I don't understand why AMD is even spending R&D money on Fiji x2 unless they see something in DX12 that will solve most multi GPU problems