Nvidia's Pascal GP104 GPU may opt for GDDR5 over HBM *Updated GDDR5X JDEC Specs

[2is]

The above has nothing to do with the market dynamics he brought up.

I stopped reading here because, well... Yes it does.

 

[ShintaiDK]

If anyone here is seriously suggesting that nVidia and probably AMD won't release a middle sized chip as a flagship and charge a flagship price once 14+16nm launch, I've got a bridge to sell. Real cheap

14/16nm will have to last a real long time.

 

[Grooveriding]

I expect to the see the initial mid-range launched as high end raise the price bar from the past similar releases like 680/980. It will likely cost as much as a 980ti does now right now, $650-700.

I just want them to pull off some good stuff with the new node and chip design coming together, so we don't wind up hearing nothing but power consumption numbers to gloss over another crappy 25% performance increase. Probably won't even get a sniff of the 980ti equivalent until 2017

 

[RussianSensation]

I expect to the see the initial mid-range launched as high end raise the price bar from the past similar releases like 680/980. It will likely cost as much as a 980ti does now right now, $650-700.

I just want them to pull off some good stuff with the new node and chip design coming together, so we don't wind up hearing nothing but power consumption numbers to gloss over another crappy 25% performance increase. Probably won't even get a sniff of the 980ti equivalent until 2017

Even if NV increases perf/watt just 50%, instead of 2X, 980 will smoke the 980Ti.

147 x 1.5 = 221

Then don't forget, some newer games may use Asynchronous Compute, Voxel Global Illumination, etc. Add to that NV's driver focus will turn towards Pascal. If 980's 2016 successor isn't at least 25% faster @ 1440p/4K vs. the 980Ti and costs $650 US ($950 CAD), that's a fail. Of course this assumes 2016 games also become more advanced and aren't CPU limited, technologically dumbed-down console ports. Either way, there is absolutely no incentive for either AMD/NV to release a card 80-100% faster than 980Ti in 2016 and price it at $650-700 US when they can split the generation into 2 parts, or even 3 (680->780->780Ti). You know it!

Yep, that can happen if the same set of circumstances takes place. Those being, NVidia seeing that AMD's flagship card is slower then their mid-range cards. That's a caveat you are either unaware of or are intentionally leaving out.

I stopped reading here because, well... Yes it does.

It has nothing to do with AMD vs. NV. You can keep staying in denial that suddenly these companies will give you a card 75-100% faster in 2016. If AMD's card beats NV's by 5-10%, or NV's card beats AMD's by 5-10%, both companies will price next gen's mid-range chips at flagship prices. That's the point 3DVagabond was making and he is spot on. This is because in a marketplace of shrinking GPU volume unit sales, longer GPU upgrade cycles by gamers, more expensive wafers/lower nodes, lower yields associated with new cutting edge nodes (14nm/16nm FinFET), the focus is survival => profit & gross margin maximization.

Since AMD has no $ to compete on price/performance anymore, and their CEO desires to shift away from their budget image, it's more likely than not that AMD's next gen mid-range will cost $500-600+ out of the gate. NV will do the same. Both companies have seen how well this strategy has worked since 2012 so there is 0 reason to believe it will somehow change. If one of these companies released their card first, let's say Polaris at $550, NV will launch its card at $500-650, not $299-350 which is the historical price for mid-range next gen cards. If NV releases GP104 first for $500-650, AMD will also go with $500-650, not $299-350.

That's why it has nothing to do with AMD vs. NV but the overall market dynamics that consumers have created by accepting the new status quo. Frankly, it would be stupid for AMD/NV NOT to price next gen mid-range chips at $500-650. It's the consumers who vote with their wallet and if the consumers keep paying these prices, it's only logical to keep raising prices. Hence why Grooveriding suggests $650-700 as not out of the question.

In simplest terms, AMD has no $ or desire to compete on price/performance anymore, while NV would love nothing more than raising prices higher and higher. This creates the perfect storm of rising GPU prices, or at the very least a repeat of the last gen which is $550 980->$550 GP104 (mid-range). As long as GP104/Polaris is even 10% faster than 980Ti at $550, it'll be recommended. Professional review sites get free samples to review and these free samples are sent to get consumers to buy products after reading reviews. The only people who can get AMD/NV to go back to the old prices are consumers. If anything, it makes WAY more sense to price every new faster card at $650-700, because why not? I guarantee it if Polaris/GP104 is only 15-25% faster than 980Ti and costs $650-700, it'll sell like hot cakes. No need to bring out the big guns or lower prices because consumers keep paying, so...

 

[2is]

ok :thumbsup:

 

[Silverforce11]

Assuming both companies pursue this strategy, with closely competing products. Where do you see AMD's market-share for 2016-2017?

If NV is late, AMD could pull off the 4800 & 5800 series come-back after the disaster 3800 series.

Fermi wasn't its true potential until the 580, with a redesign along with several metal layer respins.

I'm not an NV shareholder and as a gamer, I actually look forward to AMD making a come back to keep the competition real so we won't get accustomed to minor iterations with a price hike like Intel has been doing to us all.

 

[MrTeal]

Even if NV increases perf/watt just 50%, instead of 2X, 980 will smoke the 980Ti.

147 x 1.5 = 221

Depends on how big it is. GM204 is also 35% larger than GK104 was. AFAIK the largest die that's come out of TSMC's 16FF is A9X at 147mm², and while GPUs are easier to bin for defects than SOCs going for a 400mm² gaming chip might be pushing it. It wouldn't be shocking to see the first Pascal die come in at 300mm² and the same number of shaders as full GM200, assuming they don't radically change anything like they did from Fermi to Kepler. It would still be faster through uarch improvements, might clock higher and consume less power, but not blow GM200 out of the water. Good upgrade for someone like me, but not someone already on a 980 Ti unless nVidia and games really push the new features.

How long before we get to play with GP100, who knows. All the initial GK110s went to Oak Ridge, and they didn't launch K20 until 8 months after GK104. Even then that wasn't a GPU, and it took another 3 months to launch in a cut down form in the Titan and then another 3 months to launch an 80% enabled die in the 780 at a more reasonable price. It wasn't until 20 months after Kepler launched in the 680 that gamers finally got to buy a fully enabled GK110. If GP100 is the compute powerhouse to replace GK110 that most expect it to be, I don't think anyone will get a GPU out of it for quite awhile. There's plenty of existing contracts and Telsa cards that can be sold at $3000 to absorb poor yields first.

 

[2is]

If NV is late, AMD could pull off the 4800 & 5800 series come-back after the disaster 3800 series.

Fermi wasn't its true potential until the 580, with a redesign along with several metal layer respins.

I'm not an NV shareholder and as a gamer, I actually look forward to AMD making a come back to keep the competition real so we won't get accustomed to minor iterations with a price hike like Intel has been doing to us all.

The 3800 series was a gem after the disastrous 2900 series

 

[MrTeal]

The 3800 series was a gem after the disastrous 2900 series

Oh, glory days, back when the 3850 launched at $180 and destroyed the 18 month old $500 7900 GTX and the year old $450 1950 XTX.
Somehow I doubt little Polaris will have the same success relative to the 390X and 980.

 

[Silverforce11]

The 3800 series was a gem after the disastrous 2900 series

They were both pretty awful TBH. Wasn't good until 4800 series.

 

[2is]

The 3870 couldn't compete with the likes of an 8800GT but it was FAR better than the 2900 series which was underperforming, consumed way too much power and was loud as heck.

 

[RussianSensation]

They were both pretty awful TBH. Wasn't good until 4800 series.

Ya, 3870 was junk. It was just 6-10% faster than HD2900XT. To make matters worse, 3870 offered just 83-87% of 8800GT's performance. That's akin to Fury X offering just 83-87% of the GTX970's performance. If GP104 is NV's 8800GT and Polaris flagship is a repeat of 3870, AMD might as well close shop.

 

[raghu78]

Even if NV increases perf/watt just 50%, instead of 2X, 980 will smoke the 980Ti.

147 x 1.5 = 221

GTX 980 - 2048 sp, 5.2 billion transistors, 398 sq mm.
GTX Titan - 3072 sp, 8 billion transistors, 601 sq mm.

RS there are many factors at play here and its difficult to predict how much faster Pascal GP104 would be over GM204 and GM200. Maxwell was an extremely focussed design which served very well for gaming. Pascal is an architecture which is going to be designed for DX12 gaming, compute and VR. The architectural design choices which Nvidia made for Maxwell and Pascal could be very different. Moreover Maxwell was manufactured on an extremely mature 28nm node and GM200 is the largest Nvidia GPU ever. I do not see Nvidia going too aggressive on die sizes with the first gen FINFET GPUs as yields are still very hard on 200+ sq mm high performance chips which draw closer to 200w. There is a world of difference in yielding mobile SoC like A9X at 147 sq mm, 5W and 300-350 sq mm, 180-200w high performance GPUs. I would expect GP104 to beat GTX Titan easily due to an improved architecture and a vastly improved FINFET process node. I expect GP104 will be a very efficient chip.

DX 1080 - 4096 Pascal CUDA cores, 256 bit GDDR5X or 256 bit HBM2, 400+ GB/s, 180-200w. Atleast 20-25% faster than GTX Titan. I think we will see a repeat with the salvage SKU GTX 1070 edging out GTX Titan on avg performance and in DX12 games the gap will widen. Pascal will be Nvidia's first architecture designed for DX12 gaming, compute and VR. I am looking forward to see what design choices they made.

 

[RussianSensation]

DX 1080 - 4096 Pascal CUDA cores, 256 bit GDDR5X or 256 bit HBM2, 400+ GB/s, 180-200w. Atleast 20-25% faster than GTX Titan. I think we will see a repeat with the salvage SKU GTX 1070 edging out GTX Titan on avg performance and in DX12 games the gap will widen. Pascal will be Nvidia's first architecture designed for DX12 gaming, compute and VR. I am looking forward to see what design choices they made.

That's exactly my point. Even with a conservative estimate of 1.5X (50%) improvement in perf/watt, or basically half of what NV is aiming for, GTX980 147% ends up at 221% on that chart I linked (or 20% faster than 980Ti). Using your estimate of 20-25% faster than the Titan X, that would mean:

Forecast 1: 20% faster than TX => 196% x 1.20 = 235% or 29% faster than 980Ti

Forecast 2: 25% faster than TX => 196% x 1.20 = 245% or 34% faster than 980Ti

This shouldn't be too surprising considering GTX580 -> GTX680. That's what I am saying: next gen mid-range chips, if done right, should have no problem smoking 980Ti/Fury X.

Even just 15% more performance over Fury non-X/Nano will produce a card faster than the Fury X/980Ti!

It seems there is too much conservatism from some going into 14nm/16nm generation but the last time we saw something similar with Fermi -> Kepler, even a 660Ti beat out 580 and 660 was 95% as fast!

That means the 680>670>660Ti all beat 580 (this gen's Titan X), and 660 was almost as fast, faster than a GTX570 (this gen's 980Ti).

Also, the estimates you and I made are very conservative.

If GP104 fully unlocked with similar TDP to a 980 ends up at 245%, that's just a 67% increase in perf/watt over today's 980 (147%).

If NV truly achieves closer to 90-100% increase in perf/watt, that would imply 980's successor could in theory end up at 147% x 1.9 = 279% or a mind-blowing 53% faster than 980Ti @ 1440p.

 

[Head1985]

It will be same like it was GTX580>GTX680
GTX1080 will beat titanx around 30-40% and 40-50% vs GTX980TI.
1070 will be around 15% slower so still beat TITANX 15-25% and 980TI 25-35%.

 

[Ajay]

It will be same like it was GTX580>GTX680
GTX1080 will beat titanx around 30-40% and 40-50% vs GTX980TI.
1070 will be around 15% slower so still beat TITANX 15-25% and 980TI 25-35%.

Because 28nm to 16FF+ will be identical to the shift from 40nm to 28nm? Pray tell, what specific process tech improvements and architectural changes will cause this wonderful result?

 

[raghu78]

That's exactly my point. Even with a conservative estimate of 1.5X (50%) improvement in perf/watt, or basically half of what NV is aiming for, GTX980 147% ends up at 221% on that chart I linked (or 20% faster than 980Ti). Using your estimate of 20-25% faster than the Titan X, that would mean:

Forecast 1: 20% faster than TX => 196% x 1.20 = 235% or 29% faster than 980Ti

Forecast 2: 25% faster than TX => 196% x 1.20 = 245% or 34% faster than 980Ti

This shouldn't be too surprising considering GTX580 -> GTX680. That's what I am saying: next gen mid-range chips, if done right, should have no problem smoking 980Ti/Fury X.

Even just 15% more performance over Fury non-X/Nano will produce a card faster than the Fury X/980Ti!

It seems there is too much conservatism from some going into 14nm/16nm generation but the last time we saw something similar with Fermi -> Kepler, even a 660Ti beat out 580 and 660 was 95% as fast!

That means the 680>670>660Ti all beat 580 (this gen's Titan X), and 660 was almost as fast, faster than a GTX570 (this gen's 980Ti).


Also, the estimates you and I made are very conservative.

If GP104 fully unlocked with similar TDP to a 980 ends up at 245%, that's just a 67% increase in perf/watt over today's 980 (147%).

If NV truly achieves closer to 90-100% increase in perf/watt, that would imply 980's successor could in theory end up at 147% x 1.9 = 279% or a mind-blowing 53% faster than 980Ti @ 1440p.

GTX 980 was a 400 sq mm part. I doubt we are going to see anything more than 300 sq mm for first gen FINFET flagship. Anyway I think we have to wait and see but I think 20-25% over Titan X is easy to achieve.

 

[Kenmitch]

That's exactly my point. Even with a conservative estimate of 1.5X (50%) improvement in perf/watt, or basically half of what NV is aiming for, GTX980 147% ends up at 221% on that chart I linked (or 20% faster than 980Ti). Using your estimate of 20-25% faster than the Titan X, that would mean:

Forecast 1: 20% faster than TX => 196% x 1.20 = 235% or 29% faster than 980Ti

Forecast 2: 25% faster than TX => 196% x 1.20 = 245% or 34% faster than 980Ti

This shouldn't be too surprising considering GTX580 -> GTX680. That's what I am saying: next gen mid-range chips, if done right, should have no problem smoking 980Ti/Fury X.

Even just 15% more performance over Fury non-X/Nano will produce a card faster than the Fury X/980Ti!

It seems there is too much conservatism from some going into 14nm/16nm generation but the last time we saw something similar with Fermi -> Kepler, even a 660Ti beat out 580 and 660 was 95% as fast!

That means the 680>670>660Ti all beat 580 (this gen's Titan X), and 660 was almost as fast, faster than a GTX570 (this gen's 980Ti).

Also, the estimates you and I made are very conservative.

If GP104 fully unlocked with similar TDP to a 980 ends up at 245%, that's just a 67% increase in perf/watt over today's 980 (147%).

If NV truly achieves closer to 90-100% increase in perf/watt, that would imply 980's successor could in theory end up at 147% x 1.9 = 279% or a mind-blowing 53% faster than 980Ti @ 1440p.

Or the 980's replacement could be 10% faster and use 1/2 the watts. Green is still in?

Time will tell.

 

[3DVagabond]

Or the IF 980's replacement could be 10% faster and use 1/2 the watts. Green is still in?

Time will tell.

FTFU.

If it uses 2x the power but is 10% faster than AMD then performance is all that matters.

 

[Ajay]

GTX 980 was a 400 sq mm part. I doubt we are going to see anything more than 300 sq mm for first gen FINFET flagship. Anyway I think we have to wait and see but I think 20-25% over Titan X is easy to achieve.

It's going to depend allot on the yield curves vs frequency && area. Maxwell was already shoehorned into 28nm (was supposed to be 20nm). I'm sure, unless Pascal is a significant improvement performance wise over Maxwell, that we won't be seeing big increases in the first gen 16FF+ GM204 class GPUs (except for a perf/watt). The 6-9 month estimates for shipping retail GPUs is probably to get that yield curve up higher so that the next gaming enthusiast class CPU delivers what you predict (otherwise, how does NV get gamers to throw money at them).

Of course, NV could release earlier, maybe with a larger area and moderate frequency and charge early adopters through the nose, kind of like they did with the GK104 D:

 

[2is]

FTFU.

If it uses 2x the power but is 10% faster than AMD then performance is all that matters.

The flip side being, AMD is better by merely existing but not excelling at anything.

 

[GodisanAtheist]

Lisa Su recently mentioned that the 14/16nm generation is expected to last 5 years if not more, I imagine NV/AMD will attempt to get 3 major launches into the time frame to maximize yields and profits (which we have already seen in the 28nm gen):

200-300mm die as initial launch flagship ~20% faster than FuryX/980ti (7970/680 on 28nm)
400-500mm die 1.5/2 years later ~ 40% faster than FuryX/980Ti (290x/780ti on 28nm)
600mm monstrosity 1.5/2 years after that ~ 70% faster than FuryX/980ti (FuryX/980ti on 28nm)

Of course these numbers will inflate over time thanks to driver deprecation of the older cards and new game engines built around newer technologies (and dev relations programs), so it will ultimately feel like we're getting a lot more than we really are.

NV/AMD stumbled through this "strategy" on 28nm because they were not ready for it, but any illusion of timely node progression is gone at this point.

The traditional PC market is shrinking and I don't see it having some massive resurgence over the next 5 years, so NV/AMD have to stretch their dollar in this market as well. As gaming has moved into the mainstream, there are a lot of new buyers (larger slice of a smaller pie) who are calibrated to the new $500 for a 300mm chip scale who won't raise a stink about this (and honestly, its not like old-timers have anywhere else to turn).

 

[Silverforce11]

FTFU.

If it uses 2x the power but is 10% faster than AMD then performance is all that matters.

As a long time reader of this forum and other tech forum, I know that goal post shifting trend is true.

When the 480 used 2x the power the 5870 did, power usage was not important. When NV's dustbuster was ridiculously noisy, noise wasn't important.

When NV's GPU was slower in single card, 680 vs 7970Ghz, multi-GPU was important because the 680 SLI was smoother. When AMD CF was smoother, smoother was no longer important. When AMD's Fury X CF stomps on 980TI SLI, multi-GPU is no longer important.

The same for OC potential, when the 7950 pulled 50% OC, OC wasn't important. Now Maxwell OC greats, OC is THE metric.

If Pascal has functional Async Compute, then I bet AC will be important... otherwise, "Async Compute? Meh!"..

And if Pascal is a compute focused design that sacrifices perf/w and loses to Polaris in power efficiency, power usage will no longer be important again. We're going full circle.

Whatever happened to simple and logical metrics such as bang for buck? Are you people actually happy to pay ridiculous prices for your GPUs?

 

[Glo.]

FTFU.

If it uses 2x the power but is 10% faster than AMD then performance is all that matters.

So it is 4096 cores, clocked at around 950-1000 MHz. With TDP for each GPU at 100W.

Basically GTX980 with 100W of TDP.

 

[DooKey]

Goal post shifting isn't a NDF only past time. The ADF loves to do it just as much. Computer hardware company fans are all pitiful regardless of the color.

I'll be buying the "Flagship" GPU of whichever company releases first. If the opposing company delivers faster after that then I'll be buying that as a replacement. I need my shiny precious things.