Overclock.net › Forums › Industry News › Rumors and Unconfirmed Articles › [TT] NVIDIA should launch its next-gen Pascal GPUs with HBM2 in 2H 2016
New Posts  All Forums:Forum Nav:

[TT] NVIDIA should launch its next-gen Pascal GPUs with HBM2 in 2H 2016 - Page 16

post #151 of 724
Quote:
Originally Posted by Avant Garde View Post

If GP100 will be 50% faster than GP200 than all those 980Ti owners will be super ashamed, dishonored and highly insulted by back-stabbing company and their new MUCH FASTER product released almost not even a year later. Now, that would be unprecedented scandal smile.gif

BUT on the other hand, if they don't release FAST GP100 and don't release it SOON they will be facing almost certain loss of GPU market pie as AMD launch Polaris with HBM2 so.... Tough decisions for management.

If GP100 is only 50% faster than GM200 I'll lachen.gif and then just skip Pascal entirely. Nothing short of 2x improvement over GM200 will impress me at this point.
post #152 of 724
Your expectations are sky high biggrin.gif
    
CPUMotherboardGraphicsRAM
Intel Core i5 3570K Asus Maximus V Gene EVGA GTX1080 FTW Kingston HyperX Predator 8GB 2400MHz 
Hard DriveHard DriveHard DriveHard Drive
Samsung SSD 850 EVO 1TB Samsung SSD 830 128GB Western Digital Caviar Blue 1TB Wester Digital Caviar Black 1TB 
Optical DriveCoolingOSMonitor
LG Corsair A50 Windows 7 Ultimate 64bit Asus ROG Swift PG348Q  
MonitorKeyboardPowerCase
Dell P2414H Lenovo Seasonic G-Series 650W Thermaltake 
MouseMouse PadAudioAudio
Logitech G900  SS 4HD Logitech Z4 FiiO E10K amp/dac 
  hide details  
Reply
    
CPUMotherboardGraphicsRAM
Intel Core i5 3570K Asus Maximus V Gene EVGA GTX1080 FTW Kingston HyperX Predator 8GB 2400MHz 
Hard DriveHard DriveHard DriveHard Drive
Samsung SSD 850 EVO 1TB Samsung SSD 830 128GB Western Digital Caviar Blue 1TB Wester Digital Caviar Black 1TB 
Optical DriveCoolingOSMonitor
LG Corsair A50 Windows 7 Ultimate 64bit Asus ROG Swift PG348Q  
MonitorKeyboardPowerCase
Dell P2414H Lenovo Seasonic G-Series 650W Thermaltake 
MouseMouse PadAudioAudio
Logitech G900  SS 4HD Logitech Z4 FiiO E10K amp/dac 
  hide details  
Reply
post #153 of 724
Quote:
Originally Posted by magnek View Post

If GP100 is only 30% faster than GM200 then it's beyond an epic fail, and nVidia should be ashamed to even release a turd like that.

Exactly.
Super P's rig
(20 items)
 
  
CPUMotherboardGraphicsRAM
5960x ASUS X99-A II Asus GTX 1080 Ti Corsair Vengeance DDR4 3000 
Hard DriveHard DriveHard DriveOptical Drive
MyDigitalSSD BPX NVMe Samsung 850 EVO Seagate Momentus XT 500 GB External DVDRW 
CoolingCoolingOSMonitor
EK-XLC Predator 240 Swiftech 240mm Radiator Windows 10 Samsung 40" 4K - UN40KU6290 
KeyboardPowerCaseMouse
G710+ EVGA SuperNOVA 850G2 Fractal Design Define S G700s 
Mouse PadAudioAudioAudio
Vipamz Extended XXXL Asus U7 M-Audio AV40 Sennheiser HD 439 
  hide details  
Reply
Super P's rig
(20 items)
 
  
CPUMotherboardGraphicsRAM
5960x ASUS X99-A II Asus GTX 1080 Ti Corsair Vengeance DDR4 3000 
Hard DriveHard DriveHard DriveOptical Drive
MyDigitalSSD BPX NVMe Samsung 850 EVO Seagate Momentus XT 500 GB External DVDRW 
CoolingCoolingOSMonitor
EK-XLC Predator 240 Swiftech 240mm Radiator Windows 10 Samsung 40" 4K - UN40KU6290 
KeyboardPowerCaseMouse
G710+ EVGA SuperNOVA 850G2 Fractal Design Define S G700s 
Mouse PadAudioAudioAudio
Vipamz Extended XXXL Asus U7 M-Audio AV40 Sennheiser HD 439 
  hide details  
Reply
post #154 of 724
Quote:
Originally Posted by escksu View Post

No, I meant the big Pascal 30% faster. Thats what I read on the net. The new titan will be 30% faster.

If its 60% then it should be decent enough for 4K60. If not I dont think it will suffice.

If true, pretty sad.

The Maxwell Titan was 35% faster than Kepler Titan Black at 28nm. The Titan Black was close to 100% faster than a GTX 580 with a new architecture and a move from 40nm to 28nm.

With the move to 16nm and Pascal, anything less than 66% would be huge disappointment.
Upstairs Rig
(11 items)
 
  
CPUMotherboardGraphicsRAM
4770k Asus Maximus VI Hero evga 1080 Ti with Hybrid mod Corsair Vengeance Pro 2133 mhz 
Hard DriveHard DriveCoolingOS
Samsung 850 EVO 500gb WD Caviar Black Corsair h100i GTX Windows 8.1 64bit 
MonitorPowerCase
xb280hk EVGA Supernova 1000 G2 Corsair Carbide Air 540 
  hide details  
Reply
Upstairs Rig
(11 items)
 
  
CPUMotherboardGraphicsRAM
4770k Asus Maximus VI Hero evga 1080 Ti with Hybrid mod Corsair Vengeance Pro 2133 mhz 
Hard DriveHard DriveCoolingOS
Samsung 850 EVO 500gb WD Caviar Black Corsair h100i GTX Windows 8.1 64bit 
MonitorPowerCase
xb280hk EVGA Supernova 1000 G2 Corsair Carbide Air 540 
  hide details  
Reply
post #155 of 724
Quote:
Originally Posted by Avant Garde View Post

Your expectations are sky high biggrin.gif

Not really, just going by past precedence that nVidia set for themselves.

580 (GF110) to 680 (GK104) was a 30% jump, 580 to 780 Ti/Titan Black (full GK110) was almost a 100% jump. Fermi to Kepler also involved a process shrink (40 to 28nm), architecture change, and memory update (much faster GDDR5 on Kepler).

Then 780 Ti to Titan X (full GM200) was another 50% jump with just an arch update and no process shrink or memory update.

So to sump it up:

GF110 to GK104: 30% jump
GF110 to GK110: 100% jump
GK110 to GM200: 50% jump

So if GM200 to GP100 is only a 30% jump, that's beyond pathetic.
post #156 of 724
Quote:
Originally Posted by mcg75 View Post

If true, pretty sad.

The Maxwell Titan was 35% faster than Kepler Titan Black at 28nm. The Titan Black was close to 100% faster than a GTX 580 with a new architecture and a move from 40nm to 28nm.

With the move to 16nm and Pascal, anything less than 66% would be huge disappointment.

It is possible.

It depends if big titan is just a bit over 400mm2.

Nvidia said just a bit over 4tf dp for the card. If they have a single precision to double precision ratio of 2 to 1, that would mean a 8.x tf card. This would equal about 35 percent increase with a bit of architectural improvements being minimal.

16nm is simply too expensive. At this early stage, the bigger the die, the bigger the performance and bigger the cost. If people want 2x performance from a gm200 expect to pay 1k minimum and 1200 for the titan x version. A die that's twice as large a titan x costs more than 2x as much due to the immaturity of the process and the cost of finfet chips.

400mm2 cards are probably probably similar in pricing for cost of production as gm200 cards. Add in the HBM and the cost is quite a bit more.

The cost ramifications of finfet are a game changer for the pricing of cards. Cost of sales which is the biggest cost for these companies consist primarily of water costs and this is the first time wafer price doubled rather than the typicaly 15% between nodes.
post #157 of 724
Quote:
Originally Posted by tajoh111 View Post

It is possible.

It depends if big titan is just a bit over 400mm2.

Nvidia said just a bit over 4tf dp for the card. If they have a single precision to double precision ratio of 2 to 1, that would mean a 8.x tf card. This would equal about 35 percent increase with a bit of architectural improvements being minimal.

16nm is simply too expensive. At this early stage, the bigger the die, the bigger the performance and bigger the cost. If people want 2x performance from a gm200 expect to pay 1k minimum and 1200 for the titan x version. A die that's twice as large a titan x costs more than 2x as much due to the immaturity of the process and the cost of finfet chips.

400mm2 cards are probably probably similar in pricing for cost of production as gm200 cards. Add in the HBM and the cost is quite a bit more.

The cost ramifications of finfet are a game changer for the pricing of cards. Cost of sales which is the biggest cost for these companies consist primarily of water costs and this is the first time wafer price doubled rather than the typicaly 15% between nodes.


All of which makes me believe that the rumors we've been hearing are true and we are only getting small to mid-size Pascal this year and no big Pascal until at least next year.
post #158 of 724
Quote:
Originally Posted by tajoh111 View Post

It is possible.

It depends if big titan is just a bit over 400mm2.

Nvidia said just a bit over 4tf dp for the card. If they have a single precision to double precision ratio of 2 to 1, that would mean a 8.x tf card. This would equal about 35 percent increase with a bit of architectural improvements being minimal.

16nm is simply too expensive. At this early stage, the bigger the die, the bigger the performance and bigger the cost. If people want 2x performance from a gm200 expect to pay 1k minimum and 1200 for the titan x version. A die that's twice as large a titan x costs more than 2x as much due to the immaturity of the process and the cost of finfet chips.

400mm2 cards are probably probably similar in pricing for cost of production as gm200 cards. Add in the HBM and the cost is quite a bit more.

The cost ramifications of finfet are a game changer for the pricing of cards. Cost of sales which is the biggest cost for these companies consist primarily of water costs and this is the first time wafer price doubled rather than the typicaly 15% between nodes.

Is that just speculation, or do you have some sources for this? Kepler to Maxwell was an arch update only and look how much extra performance they wrung out.
post #159 of 724
Considering HBM needs less space on the die, we have a new architecture and 14nm how good in theory would a 550mm² Pascal flagship card be? Ignoring the fact how likely it is that we will see it, but how big a jump from the 980 Ti / Titan X would it be?

I am not good at this, but it should be more than 100% better shouldn't it?
current rig
(4 items)
 
  
CPUGraphicsGraphicsKeyboard
2500k R9 290X Tri-X GTX 680 Filco Majestouch 2 Brown Ninja 
  hide details  
Reply
current rig
(4 items)
 
  
CPUGraphicsGraphicsKeyboard
2500k R9 290X Tri-X GTX 680 Filco Majestouch 2 Brown Ninja 
  hide details  
Reply
post #160 of 724
Quote:
Originally Posted by zealord View Post

...how good in theory would a 550mm² Pascal flagship card be??

Let's do a pie in the sky Fermi estimation here.

Let's take the Titan X, full GM200.

601mm^2 and 8 billion transistors on the 28nm process.

Now we know the '28nm' and '16nm' are not actually that size, but let's assume the processes are the same relative shrink.

So the full 28nm GPU is capable of 13.31 million transistors per mm^2. Square root of 13.31 = 3648 transistors in a linear mm.

A relative die shrink of 16/28 would increase that to 3648 * 28 / 16 = 6385 transistors per mm

6385^2 = 40.76 million transistors per mm^2 on 16nm ff+.

That estimation matches the rumor mill of 17B transistors for GP100. 17000 / 40.76 = 417 mm^2. A perfectly reasonable size die for a 'big pascal' chip on a new immature process.

So 550mm^2 would get us a whopping 22.42 billion transistors on 16ff+. Now we have to assume per transistor performance losses compared to the Titan X due to FP32 being put back into the architecture. So comparing to the GK110 instead that is 316% of a Titan Black.

So a safe 3x Titan Black performance. Possibly 3.2x. A very comfortable double the performance of a Titan X, and likely more like 215% that of a Titan X.
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Rumors and Unconfirmed Articles
Overclock.net › Forums › Industry News › Rumors and Unconfirmed Articles › [TT] NVIDIA should launch its next-gen Pascal GPUs with HBM2 in 2H 2016