NVIDIA Pascal Thread

[Kris194]

Could be fake but they're saying NVIDIA is about to phase out the Geforce GTX 980 Ti in favour of the new Pascal cards. Will start shipping the GPUs to manufacturers soon, still enough time to demo the new cards at Computex. Launch might happen 1-2 months later in July/August.

www.zolkorn.com/news/nvidia-starting-stop-delivery-out-gm200-for-gtx-980ti

Why are you surprised? It doesn't make sense to manufacture GTX 980 Ti while new cards will be cheaper and more powerful. It was obvious.

 

[airfathaaaaa]

Perhaps they're planning a paper launch at Computex, with limited availability initially.

micron has SAMPLED gddr5x its not ready for anyone to use at any level of production yet

 

[RussianSensation]

Why are you surprised? It doesn't make sense to manufacture GTX 980 Ti while new cards will be cheaper and more powerful. It was obvious.

Exactly. There was a poster with an HD7970 RMA in another thread that couldn't get an R9 390/390X RMA step-up since Diamond is no longer receiving these chips from AMD -- also EOL products. In managing inventory of outdated generations, it makes sense that NV/AMD will stop manufacturing of more expensive/outdated products.

Unless NV starts off the Pascal GP104 generation sand-bagging their clock speeds and/or their 2X perf/watt marketing is lies, 970's successor should beat 980Ti.

970 = 58%
980Ti = 83%

970 -> 1070
Bare minimum: 1.5X perf/watt for 1070 = 87%
1.75X perf/watt for 1070 = 102%
2X perf/watt for 1070 = 116%

980 -> 1080
Bare minimum: 1.5X perf/watt for 1080 = 101%
1.75X perf/watt for 1080 = 117%
2X perf/watt for 1080 = 134%

Let's not forget that 980Ti is not a full chip that 780Ti was. If NV's claims of 2X the perf/watt are true, 1080 could be up to 50-60% faster than 980Ti in some of the most GPU limited games at 1440p/4K. With the competition not having anything remotely close until end of 2016/early 2017, I could see NV bumping up the price of 970->1070 to $399-449 and 980->1080 to $599-649.

 

[SteveGrabowski]

Let's not forget that 980Ti is not a full chip that 780Ti was. If NV's claims of 2X the perf/watt are true, 1080 could be up to 50-60% faster than 980Ti in some of the most GPU limited games at 1440p/4K. With the competition not having anything remotely close until end of 2016/early 2017, I could see NV bumping up the price of 970->1070 to $399-449 and 980->1080 to $599-649.

You think they'd sell well at that price? What made the 970 such a success was the $330 price tag. If that was $400 I doubt they would have sold anywhere near as many, but at $330 they were able to shift the new normal up from the $250 they were getting when everyone bought GTX 760s. Putting 70 series at a price low enough to get mass adoption while making a really lame 60 series card made them a boatload of money this generation while putting a ton of pressure on AMD, why stop a winning formula?

 

[jpiniero]

You think they'd sell well at that price? What made the 970 such a success was the $330 price tag. If that was $400 I doubt they would have sold anywhere near as many, but at $330 they were able to shift the new normal up from the $250 they were getting when everyone bought GTX 760s.

Exactly, higher but not a huge increase. The new normal for x70 will be $399. The 1199 would be $499 and so on. Given how expensive this node is, increasing the transistor count is only just going to increase their costs so if they want to maintain margins this is how it'll have to be.

 

[Qwertilot]

Well, if they can get even the 1070 > than the 980ti (or even = but much lower TDP etc really), then it looks like there won't be any sane competition from AMD until Vega hits.

So yes, they'd stick a premium on them for a bit. Why ever not?

Once vega/the bigger Pascal hit proper availability in 6-12 months after launch they'll get moved down a price bracket or two.

The same way that AMD, if left uncontested for low end 14nm stuff, won't be precisely shy of charging a bit of a premium there

We'll find out.

 

[Actaeon]

You think they'd sell well at that price? What made the 970 such a success was the $330 price tag. If that was $400 I doubt they would have sold anywhere near as many

If Nvidia makes more money per GPU sale at $400 then they won't need to sell as many as they would have at $330 to generate the same amount of profit.

Nvidia will charge what they think they can get for it so it'll depend heavily on what AMD puts out at the same price range. It is easier to move a price down to react to competition and the market than it is to increase prices if you decide you want more money.

 

[Adored]

Surely the recent benches suggesting 256-bit buses means that both are likely to have similar performance? Unless Nvidia actually designed a midrange around GDDR5X availability while AMD didn't, which I find really unlikely.

 

[Actaeon]

I am speculating here, but could we see the 1070 = GDDR5 & 1080 = GDDR5X?

I thought I read they GDDR5X doesn't need any specific hardware on the die over GDDR5, so it should be pretty easy to use the same GP104, just cut down and different memory types to create the two products.

With splitting the distribution of GDDR5X into the expensive 'niche' card instead of the more mainstream card, that helps with the production/quantity issues discussed. It also creates some meaningful performance separation between the two products and a good reason to spend extra on the 1080.

 

[tweakboy]

Can it do 4k res ? DX12 can,

 

[Qwertilot]

Surely the recent benches suggesting 256-bit buses means that both are likely to have similar performance? Unless Nvidia actually designed a midrange around GDDR5X availability while AMD didn't, which I find really unlikely.

Well..... The rumors seem to have AMD starting quite small (up to ~x60 equiv perhaps), then filling the range in with bigger chips with HBM2 in ~6(?!) months. No huge need for GDDR5X really. Maybe for a 2017/18 low end refresh if they can't sensibly go all HBM.

If NV are (as vaguely indicated) trying to release a fairly large/chunky 1070/80, then filling in both directions over time - huge compute chip with HBM2 to get into consumer stuff and smaller things later(?) - then they might well have decided they needed it rather more.

Or, of course, all of this could be nonsense

For instance, given the massive power discrepancy the die shrink is going to give, its hard to see NV leaving getting out the 950/60 equivalent quite as long as they did with Maxwell.

 

[jpiniero]

I am speculating here, but could we see the 1070 = GDDR5 & 1080 = GDDR5X?

I thought I read they GDDR5X doesn't need any specific hardware on the die over GDDR5, so it should be pretty easy to use the same GP104, just cut down and different memory types to create the two products.

I think it's more likely that both will be GDDR5 but will be refreshed to use GDDR5X sometime next year.

 

[RussianSensation]

You think they'd sell well at that price? What made the 970 such a success was the $330 price tag. If that was $400 I doubt they would have sold anywhere near as many, but at $330 they were able to shift the new normal up from the $250 they were getting when everyone bought GTX 760s. Putting 70 series at a price low enough to get mass adoption while making a really lame 60 series card made them a boatload of money this generation while putting a ton of pressure on AMD, why stop a winning formula?

NV realized there were still a lot of HD4000-7000 series hold outs that didn't upgrade to 285/280X/290/290X and capitalized on the opportunity to get AMD users to switch. Now that these gamers are in the NV eco-system with a 970, unless Polaris can compete on performance (doubtful given the rumors of it bringing 970/290X performance to lower tiers -- i.e. That's not an upgrade path for 970 users), NV can easily raise the price to $399-429 and get the same customers to upgrade for 50%+ boost over the 970. If I have a 970 now and 1070 is 50% faster for $399-429, I am not waiting until December 2016/January 2017 for Vega. AMD released their road-map so NV knows this.

Another possibility of course is the return of 3-tier mid-range cards:

680 -> 1080
670 -> 1070
660Ti -> 1060Ti

If Pascal is 2X perf/watt, 1060Ti alone could ~ 390X/Fury. Then this could be a $299-329 card.

It's also possible that the performance gap between 1070 and 1080 could grow to 20-25%. We have seen a larger separation between 970 and 980 as opposed to 670 and 680.

Either way, unless Polaris can beat Fury X by 20-25%, I don't see it competing with GP104 and that means way less pricing pressure on NV in 2016. On the performance/watt, Nano is still a good card. Since AMD has way less $$$ than NV, I wouldn't doubt it if the Nano/Fury/Fury X got big price cuts (aka 290/290X did when Maxwell launched) to stay relevant until Vega. I think instead AMD will do something crazy like Polaris "390 performance" for $199-229 and "390X performance" for $249-269. This allows AMD to regain a ton of market share on the low end while using Fury/Nano/Fury X as $349-399 mid-range. 1080 then should have the market all to itself for most of 2016. Unfortunately that is how I see it given everything coming out from AMD's rumormill is small die/low power/mobile-laptop focus. I am of course speculating but why didn't AMD show us Polaris 10 against GTX980?

Besides, an after-market 980Ti has nearly 2X the perf/watt of a 290X at 1440p:
100% vs. 53%
http://www.techpowerup.com/mobile/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/24.html

That means even if Polaris has 2.5X the perf/watt of 290X, GP104 will slaugher it. It's my quick math based on what both firms have said about next gen perf/watt and where their 2013-2015 cards sit wet to each other. AMD needs to get 2.5X the perf/watt from Fury X to actually get to Pascal's 2X over Maxwell given how far behind they are right now.

^ This is why AMD is showing Vega improving perf/watt again over Polaris -- because 2.5X over GCN 1.0/1.1 isn't good enough to match Pascal's 2X over Maxwell. Notice how AMD kept focusing on bringing VR GPU spec to the masses? That literally smells like HD4850/4870 strategy of bringing Polaris price/performance and low power. Remember, Pitcairn, at least for the first 2 years of its life, was not a GTX670/680 competitor. That's why I just do not see Polaris 10/11 competing against 1070/1080, maybe against 1060/1060Ti.

Another way of looking at it is we have a leak of 2304 shader 256-bit Polaris card. Even if we assume the same 35% increase in IPC AMD could get on Stream Processor that NV managed with Kepler --> Maxwell, then bump GPU clocks 30%, that means 2304 Polaris under these conditions is only as good as a ~ "4044 shader" Fury X. Even with only 50% perf/watt over 970, 1070 alone would already beat such a mythical Polaris 10 card. 1300-1350mhz GPU clocks and 35% per Shader core IPC increase are some wildly optimistic assumptions I just made for Polaris. I think it won't be anywhere close to that!

This is why I am pessimistic on AMD in 2016 as far as the desktop $349+ market segments go.

 

[Maverick177]

Which all boils down to either Polaris is 2.5 Perf/W to Hawaii or Fiji.

At 1440P, Fiji is around 56% more efficient than Hawaii, that means if Polaris is 2.5 Perf/W to Hawaii, Polaris would be only be 1.6 Perf/w to Fiji, kinda underwhelming.

 

[SteveGrabowski]

I think instead AMD will do something crazy like Polaris "390 performance" for $199-229 and "390X performance" for $249-269.

Kind of sad that would be considered crazy since that's pretty much where we were just under a year and a half ago with the 290 and 290x.

 

[SteveGrabowski]

Thanks for the breakdown, RussianSensation. You seem to know the market a lot better than I do. I just wonder if Nvidia's going to be able to create huge buzz about a $400+ gpu.

 

[Vesku]

If this rumor is correct and it is also true that AMD has stopped shipping Hawaii chips then we may be in for some healthy price competition. Assuming no pricing collusion.

 

[raghu78]

Either way, unless Polaris can beat Fury X by 20-25%, I don't see it competing with GP104 and that means way less pricing pressure on NV in 2016. On the performance/watt, Nano is still a good card. Since AMD has way less $$$ than NV, I wouldn't doubt it if the Nano/Fury/Fury X got big price cuts (aka 290/290X did when Maxwell launched) to stay relevant until Vega. I think instead AMD will do something crazy like Polaris "390 performance" for $199-229 and "390X performance" for $249-269. This allows AMD to regain a ton of market share on the low end while using Fury/Nano/Fury X as $349-399 mid-range. 1080 then should have the market all to itself for most of 2016. Unfortunately that is how I see it given everything coming out from AMD's rumormill is small die/low power/mobile-laptop focus. I am of course speculating but why didn't AMD show us Polaris 10 against GTX980?

Besides, an after-market 980Ti has nearly 2X the perf/watt of a 290X at 1440p:
100% vs. 53%
http://www.techpowerup.com/mobile/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/24.html

That means even if Polaris has 2.5X the perf/watt of 290X, GP104 will slaugher it. It's my quick math based on what both firms have said about next gen perf/watt and where their 2013-2015 cards sit wet to each other. AMD needs to get 2.5X the perf/watt from Fury X to actually get to Pascal's 2X over Maxwell given how far behind they are right now.

^ This is why AMD is showing Vega improving perf/watt again over Polaris -- because 2.5X over GCN 1.0/1.1 isn't good enough to match Pascal's 2X over Maxwell. Notice how AMD kept focusing on bringing VR GPU spec to the masses? That literally smells like HD4850/4870 strategy of bringing Polaris price/performance and low power. Remember, Pitcairn, at least for the first 2 years of its life, was not a GTX670/680 competitor. That's why I just do not see Polaris 10/11 competing against 1070/1080, maybe against 1060/1060Ti.

Another way of looking at it is we have a leak of 2304 shader 256-bit Polaris card. Even if we assume the same 35% increase in IPC AMD could get on Stream Processor that NV managed with Kepler --> Maxwell, then bump GPU clocks 30%, that means 2304 Polaris under these conditions is only as good as a ~ "4044 shader" Fury X. Even with only 50% perf/watt over 970, 1070 alone would already beat such a mythical Polaris 10 card. 1300-1350mhz GPU clocks and 35% per Shader core IPC increase are some wildly optimistic assumptions I just made for Polaris. I think it won't be anywhere close to that!

This is why I am pessimistic on AMD in 2016 as far as the desktop $349+ market segments go.

RS firstly you have no idea what is the state of Pascal GPU silicon ? Nvidia might make a formal demo of some Pascal chip at GTC 2016. But until they do that there is nothing but rumours. You seem to think GP104 will launch as soon as June while I could say the opposite and tell it might launch by year end. We are yet to hear any disclosure whatsoever from Nvidia on even rough launch timelines. AMD atleast said Polaris will launch in mid-2016 which puts it in June-July timeframe. btw AMD has made a wise decision to focus on entry-mid range next gen GPUs using GDDR5 as the more advanced memory technologies like HBM2 and GDDR5X won't ramp till late in 2016. These new memory technologies won't be available in very high volume and would not be suitable for lower price segments. the way I see it AMD is following a sensible strategy of launching bottom up like what Nvidia did with Maxwell. On a bleeding edge process this strategy helps with yield learning and allows for the process to mature before launching bigger GPUs. AMD will launch 2 Polaris GPUs and 2 Vega GPUs and compete against the GP107/GP106/GP104/GP100 lineup.

GP107- Polaris 11
GP106- Polaris 10
GP104- Vega 11
GP100 - Vega 10

The way I see it AMD is more likely to have any time to market advantage than Nvidia in each of the contesting products. Anyway Nvidia is in a position of dominance so even if its late by a few months it does not matter. What matters is does AMD Polaris bring major efficiency improvements to compete against Nvidia Pascal in terms of perf, perf/watt, perf/$ (for the consumer) and perf/sq mm (for AMD to have good margins).

 

[raghu78]

Which all boils down to either Polaris is 2.5 Perf/W to Hawaii or Fiji.

At 1440P, Fiji is around 56% more efficient than Hawaii, that means if Polaris is 2.5 Perf/W to Hawaii, Polaris would be only be 1.6 Perf/w to Fiji, kinda underwhelming.

The 14LPP process node alone provides 2x perf/watt gain over 28nm. For Polaris 10 we are likely talking about a 2560 sp chip with a 256 bit GDDR5 memory controller. AMD mentioned a 70/30 ratio for efficiency gains due to process node and architecture. AMD has mentioned that improving throughput and efficiency were the key goals of Polaris. The way I see it AMD has worked on resolving the bottlenecks of the current GCN architecture and try and extract as much perf as possible from those sp/CUs. The goal seems to have been to imitate Nvidia Maxwell and have more efficient and higher performing sp rather than lots of sp. By saying that AMD Polaris has perf closer to R9 390/R9 390X we begin to contradict that. So imo Polaris 10 will provide 2.5x better perf/watt than Fiji.

 

[Silverforce11]

AMD will launch 2 Polaris GPUs and 2 Vega GPUs and compete against the GP107/GP106/GP104/GP100 lineup.

GP107- Polaris 11
GP106- Polaris 10
GP104- Vega 11
GP100 - Vega 10

Agreed with this.

How fast either side comes to market is going to be the main factor as well as good the prices are.

Vega 11 and 10 are the real upgrade target for folks on current high-end, not Polaris.

 

[jpiniero]

RS firstly you have no idea what is the state of Pascal GPU silicon ? Nvidia might make a formal demo of some Pascal chip at GTC 2016.

Well I wouldn't expect any consumer parts to launch on Tuesday, only GP100 Tesla. Maybe you will get a demo of it running games but I think the focus is going to be on compute.

You seem to think GP104 will launch as soon as June while I could say the opposite and tell it might launch by year end.

July seems reasonable. It may only be mobile focused at first though.

AMD atleast said Polaris will launch in mid-2016 which puts it in June-July timeframe.

Would it really be surprising that Polaris is basically done today but they have to wait until GloFo gets going? IOW if they had done it at TSMC it would be out right now.

btw AMD has made a wise decision to focus on entry-mid range next gen GPUs using GDDR5 as the more advanced memory technologies like HBM2 and GDDR5X won't ramp till late in 2016.

Polaris isn't entry range parts. That's what the rebrands are for. Polaris is mid range and high end. Or at least it better be.

On a bleeding edge process

This isn't a bleeding edge process.

GP107- Polaris 11
GP106- Polaris 10
GP104- Vega 11
GP100 - Vega 10

Vega 11 going up against GP104 is not going to work for AMD, it'll be far too late and HBM2 is going to be too expensive for it to be in GP104's price range.

 

[3DVagabond]

Where is anyone getting their launch date info from? Has nVidia even given us a hint?

 

[Aristotelian]

Where is anyone getting their launch date info from? Has nVidia even given us a hint?

I think there have been hints that there will be a "Pascal Unveiling" at GTC on 4-7 April 2016, since it's a GPU Technology Conference. But, as jpiniero has said, there are some doubts that we will see any 970/980/980ti replacements unveiled there. But the question is - what would Nvidia unveil there then?

If everyone here is right about HBM2 and GDDR5X production being nowhere near enough to feature in current cards, is it possible that the 32 GB Pascal HBM2 pro card that there have been rumours about will feature at GTC, costing thousands of dollars per card to the professional/prosumer alike?

I get that AMD/Nvidia will first show off how efficient its new process is, and while this may be great for gaming laptops - I made a mistake buying a Razer Blade 2014 and it'll never happen again - where is the excitement for desktop gamers?

Last year at GTC the Titan X was announced for a launch in June, so maybe that "precedent" is being used by posters to infer that another high end video card will be announced by Nvidia this year, for purchase in June.

 

[raghu78]

Polaris isn't entry range parts. That's what the rebrands are for. Polaris is mid range and high end. Or at least it better be.

I said entry level and mid-range wrt next gen. I did not mean wrt last generation. Polaris 11 and Polaris 10 will be entry level and mid-range while Vega 11 and Vega 10 will be the high end and ultra high end.

Vega 11 going up against GP104 is not going to work for AMD, it'll be far too late and HBM2 is going to be too expensive for it to be in GP104's price range.

Right now nothing is known about Nvidia/AMD product launch timelines. But AMD Vega seems to be a very early 2017 launch from what anandtech inferred from their roadmap slide. This could even be pulled into late 2016 if the process is ready to yield big dies and if Nvidia puts the pressure on AMD with a GP104 launch in Q3 2016. As for pricing of GP104 and Vega 11 nothing is known so speculation is ill advised.

 

[Timmah!]

Well I wouldn't expect any consumer parts to launch on Tuesday, only GP100 Tesla. Maybe you will get a demo of it running games but I think the focus is going to be on compute.

July seems reasonable. It may only be mobile focused at first though.

Would it really be surprising that Polaris is basically done today but they have to wait until GloFo gets going? IOW if they had done it at TSMC it would be out right now.

Polaris isn't entry range parts. That's what the rebrands are for. Polaris is mid range and high end. Or at least it better be.

This isn't a bleeding edge process.

Vega 11 going up against GP104 is not going to work for AMD, it'll be far too late and HBM2 is going to be too expensive for it to be in GP104's price range.

So with all the talk how HBM2 parts wont be ready before the end of year, you expect exactly the one product to be launched, which is the most likely to feature them?

That said, i expect GP100 to be introduced too, i mean the details on the actual chip, its features, numbers of CC, metrics, performance numbers....but the actual boards with it not to be available until HBM2 are ready. Whenever that may be.

IMO, nothing will be launched on Tuesday. We will only have better understanding of what and when to expect.