Nvidia's Pascal in 2016... any guesses on pricing?

[Head1985]

I'm guessing the successor to GM204 launches first in Q2 2016 with 8gb 12ghz GDDR5X vram, 35-40% faster than Titan X at 4K for $599-649. I'm also guessing there will be two cut additional SKU's based off the same chip for $499-529 and $399-449.

This but i think midrange pascal will cost max 500USD
GTX970 successor will be 30-35% faster than TITANX for 400USD
GTX980 successor 50% faster than TITANX for 500USD

 

[iiiankiii]

This but i think midrange pascal will cost max 500USD
GTX970 successor will be 30-35% faster than TITANX for 400USD
GTX980 successor 50% faster than TITANX for 500USD

I don't think it's going to be like that. It's going to be minor upgrades because they're going to milk the heck out of 14nm/16nm. They're going to slow roll this.

2016
GTX1070 = GTX 980TI @ $399
GTX1080 = 35%-40% faster than GTX 980TI @$549

2017
GTX1170 (rebrand) = GTX1080 @$399
GTX1180 (big pascal) = 35%-40% faster than GTX1170 @$549
GTX1180TI (big pascal)= 15%-20% faster than GTX1180 @$749
GTXTitanXX (big pascal)= 20% faster than GTX1180TI @$999

 

[b-mac]

This but i think midrange pascal will cost max 500USD
GTX970 successor will be 30-35% faster than TITANX for 400USD
GTX980 successor 50% faster than TITANX for 500USD

I would be estatic if this were the case but I have my doubts we will see performance increases at these pricepoints. Hope AMD and Nvidia prove me wrong!

 

[Soccerman06]

I would be estatic if this were the case but I have my doubts we will see performance increases at these pricepoints. Hope AMD and Nvidia prove me wrong!

Were looking at a 2 node jump, if we dont see a doubling in performance I would be extremely surprised. Die size vs die size at the very least.

 

[ShintaiDK]

Were looking at a 2 node jump, if we dont see a doubling in performance I would be extremely surprised. Die size vs die size at the very least.

They will cost likely reduce die sizes due to cost. The transistor cost is pretty much the same from 28nm (Up slightly). So add transistors=cost goes up.

I wouldn't be surprised to see GTX970/980 and 390/390X successor for that matter in the 250mm2 range.

 

[Ares202]

I wouldn't be surprised to see GTX970/980 and 390/390X successor for that matter in the 250mm2 range.

That's pretty much certain, anything bigger than 300mm2 would be a massive surprise for me. They will want 1-2 more generations out of 16nm FF so building a huge die from pascal just doesn't make business sense.

 

[Braxos]

13% more power 20-25% less watt. It worked till now why change it?

 

[tviceman]

They will cost likely reduce die sizes due to cost. The transistor cost is pretty much the same from 28nm (Up slightly). So add transistors=cost goes up.

I wouldn't be surprised to see GTX970/980 and 390/390X successor for that matter in the 250mm2 range.

GM204 increased transistors (over GK104) by 50% while increasing performance by ~75%. What is the touted transistor density increase for 16nm ff+ over 28nm?

 

[MrTeal]

That's pretty much certain, anything bigger than 300mm2 would be a massive surprise for me. They will want 1-2 more generations out of 16nm FF so building a huge die from pascal just doesn't make business sense.

nVidia got about 4x the transistors/mm^2 going from GT200 on 55nm to GK104 on 28nm, actually a little above 4x. If they can do 4x again going to 16nm, a 250mm^2 to 300mm^2 GP104 could slot in at 13M-16M transistors, which is still twice as many as GM200.

You wouldn't expect linear performance scaling with transistors if nVidia reintroduces all the FP64 features they've removed from Maxwell, but decently better than 980Ti performance in a <300mm^2 GTX1080 wouldn't be out of line at all. I can't see nVidia going much smaller than that. They're hardly a company that lacks ambition, and I don't think they're want to risk Arctic Islands outright winning the top performance crown before big Pascal comes out. That presupposed GP100 doesn't ship first, of course.

I'd expect the GP104 to launch at $599, with the x70 version launching at $399.

 

[tviceman]

65% more transistors with a 20% performance per transistor would be a 2X performance increase. I'd take that!

 

[ShintaiDK]

GM204 increased transistors (over GK104) by 50% while increasing performance by ~75%. What is the touted transistor density increase for 16nm ff+ over 28nm?

Roughly twice the density.
https://www.semiwiki.com/forum/content/3884-who-will-lead-10nm.html

65% more transistors with a 20% performance per transistor would be a 2X performance increase. I'd take that!

Forget 20% per transistor. That's not happening.

 

[tviceman]

Maxwell GM204 got 20% more perf/transistor over Kepler GK104 on the same node. How is 20% out of the question with Pascal?

Edit: quick math shows GM107 got over 30% perf/transistor improvement over GK107. Looks to me like you are being pessimistic.

 

[ShintaiDK]

Maxwell was due to uarch improvements. You write 20% per transistor with a node in context as in the shrinked transistors adds performance(They would on low power devices)

But remember the joker. Transistor cost didn't go down since 28nm. So 65% more transistors roughly also means 65% higher cost for NVidia.

 

[jpiniero]

I definitely see clock speeds going up with Pascal, as that's the main way it will improve performance. Boost to 1.3 Ghz maybe?

 

[psolord]

I think that if we study what happened from the GTX 580 to the GTX 680, we will be around the ballpark of what to expect.

GF110 was 40nm, 520mm2 and had 3B transistors

GK110 was the first 28nm, 294mm2 and had 3.5B transistors.

This resulted to the following

So the 680 was 30% faster and 70% more efficient.

Going from 40nm to 28nm we had 2X more density, but going from 28nm to 16nm we have 3x more density.

Keeping in mind that we had 3 28nm iterations, we could see 4 iterations with 16nm, lol Especially if you take under consideration that shrinkage is getting ever more difficult.

The 980ti/TitanX have 8B transistors and are 600m2 at 28nm.

In order for Nvidia to give around 30% more performance, they would need around 10B transistors. I know that 10/8=1.25 but I guess the extra 5% will come from die savings from the simpler controller of HBM, IPC improvements and possibly higher frequencies. Heck it may even be more than 5%.

So for 10B transistors, we would need 600mm2+25% = 750mm2 / 3 for 16nm density = 250mm2.

If I am not completely wrong with my calculations, yeah that could be possible.

They could add 1.3 times transistors for the next iterations, so we could see

250mm2 X1.3 = 325 for the second

325mm2 x 1.3 = 423 for the third

423mm2 x 1.3 = 550 for the forth and call it a day


As for the prices the 680 came at 450 so here's hoping they can stay below 600 this time

 

[ShintaiDK]

I think that if we study what happened from the GTX 580 to the GTX 680, we will be around the ballpark of what to expect.

GF110 was 40nm, 520mm2 and had 3B transistors

GK110 was the first 28nm, 294mm2 and had 3.5B transistors.

This resulted to the following

So the 680 was 30% faster and 70% more efficient.

That's a good example of what to expect.

Going from 40nm to 28nm we had 2X more density, but going from 28nm to 16nm we have 3x more density.

Hold your horses there. TSMCs renamed 20nm to 16nm by added FF. By TSMCs own numbers 16nm is only roughly twice the density of 28nm. And it seems to be a point many people miss over and over.
https://www.semiwiki.com/forum/content/3884-who-will-lead-10nm.html

 

[psolord]

If it really is only 20nm, then we are looking at 1.96 density increase indeed.

So we will see a carbon copy of 28nm progress essentially, which is not too bad still, as long as they do a reality check regarding prices.

 

[psolord]

The 680 was 30% faster with 16% more transistors. Quite probably they can do it again. HBM2, turbo etc

 

[tviceman]

Maxwell was due to uarch improvements. You write 20% per transistor with a node in context as in the shrinked transistors adds performance(They would on low power devices)

But remember the joker. Transistor cost didn't go down since 28nm. So 65% more transistors roughly also means 65% higher cost for NVidia.

Pascal should also feature uarch improvements. If not, then it would just make sense to retool Maxwell around finfets and shrink it down as opposed to significantly overhauling architecture. From various rumored reports, Nvidia is using 16nm FF+ which is a step up from the normal 16nm FF. I think 15-20% improvement per transistor isn't out of the question, especially when the initial designs have to be very cost conscious.

 

[MrTeal]

That's a good example of what to expect.



Hold your horses there. TSMCs renamed 20nm to 16nm by added FF. By TSMCs own numbers 16nm is only roughly twice the density of 28nm. And it seems to be a point many people miss over and over.
https://www.semiwiki.com/forum/content/3884-who-will-lead-10nm.html

The gate x metal number is for comparing density, it's not for indicating how many transistors you can pack in.

Using those GxM numbers, 65nm->28nm was 2.5x the density for TMSC. The 65nm version of GT200 (GTX280, etc) was 2.43 million transistors per mm^2 (1400 / 576mm^2) while GK104 was 3540M/294mm^2= 12.04 million transistors per mm^2. Tesla got an almost 5x increase in density going from 65nm to Kepler at 28nm.

 

[MrTeal]

Pascal should also feature uarch improvements. If not, then it would just make sense to retool Maxwell around finfets and shrink it down as opposed to significantly overhauling architecture. From various rumored reports, Nvidia is using 16nm FF+ which is a step up from the normal 16nm FF. I think 15-20% improvement per transistor isn't out of the question, especially when the initial designs have to be very cost conscious.

Maxwell was also a very gaming optimised design, with limited FP64 hardware. Unlike Fermi and Kepler where it was just disabled in the GeForce cards, it just wasn't there in Maxwell. Pascal is rumoured to serve as a replacement for Kepler in the HPC market as well as a gaming GPU, so they couldn't just shrink Maxwell. It's not unreasonable to assume that Pascal will be an HPC monster but that adding all that high precision compute hardware back in could offset any other uarch gains on a gaming performance per transistor basis, or even cause a decrease in it.

 

[mysticjbyrd]

I don't think it's going to be like that. It's going to be minor upgrades because they're going to milk the heck out of 14nm/16nm. They're going to slow roll this.

2016
GTX1070 = GTX 980TI @ $399
GTX1080 = 35%-40% faster than GTX 980TI @$549

2017
GTX1170 (rebrand) = GTX1080 @$399
GTX1180 (big pascal) = 35%-40% faster than GTX1170 @$549
GTX1180TI (big pascal)= 15%-20% faster than GTX1180 @$749
GTXTitanXX (big pascal)= 20% faster than GTX1180TI @$999

How much faster is the TItanx over the 980ti?

The 680 was 30% faster with 16% more transistors. Quite probably they can do it again. HBM2, turbo etc

Nvidia is using ddr5 in 1st gen pascal

 

[tviceman]

Maxwell was also a very gaming optimised design, with limited FP64 hardware. Unlike Fermi and Kepler where it was just disabled in the GeForce cards, it just wasn't there in Maxwell. Pascal is rumoured to serve as a replacement for Kepler in the HPC market as well as a gaming GPU, so they couldn't just shrink Maxwell. It's not unreasonable to assume that Pascal will be an HPC monster but that adding all that high precision compute hardware back in could offset any other uarch gains on a gaming performance per transistor basis, or even cause a decrease in it.

Wth respects to the flagships of Kepler and Maxwell, you are right. But GM204 had more compute capability and performance than GK104 so you are wrong in that aspect and GP104 is specifically what I am talking about in my posts (I have said GM204 replacement and/or GP104 several times).

 

[happy medium]

50% + performance increases in each tier at about the same price points.
Fastest gpu not exceeding 250 watts.

AMD will be 10% slower in each tier and be priced 50$ lower.
fastest gpu not exceeding 300watts

 

[MrTeal]

Wth respects to the flagships of Kepler and Maxwell, you are right. But GM204 had more compute capability and performance than GK104 so you are wrong in that aspect and GP104 is specifically what I am talking about in my posts (I have said GM204 replacement and/or GP104 several times).

Sorry about that, you are correct there. That being said, 75% faster than GK104 might be stretching it. When the GTX980 came out TPU had it as 50% faster than the GTX770 at 1080, and even in the 950 review from a couple months ago it was 60% faster. There's a few newer games where Kepler really tanks vs Maxwell, but debating why that is is the subject of another extremely long and bitter thread.

Considering the GTX980 is a hair under 50% larger (in transistors) and clocked over 10% faster that GTX770, it would be tough to expect a 20% increase in performance per transistor from uarch alone in the general case. That doesn't preclude specific cases if a game uses some particular new feature that Maxwell doesn't do well, of course.

I'd love it if I were wrong here though.