Overclock.net › Forums › Industry News › Hardware News › [digitimes] TSMC 20nm wafer production delayed temporarily, but will not affect shipments
New Posts  All Forums:Forum Nav:

[digitimes] TSMC 20nm wafer production delayed temporarily, but will not affect shipments - Page 4

post #31 of 51
Well at least we can still expect Nvidia to push out more GK110 variants 780, 780TI, 780TI 6GB:)
post #32 of 51
Quote:
Originally Posted by fateswarm View Post

You are unaware GM107 is half the size of the GK104. They made a tiny chip that can be easily controlled in stability. It would totally fail on bigger dies.

Again, having this scaling is the ENTIRE point of Maxwell. It is literally the single design focus, designing for efficiency and scaling. I mean, sure, it could "totally fail" but not just because you said so. You're going to have to offer significantly more evidence that this won't work for me to believe you.
post #33 of 51
Quote:
Originally Posted by coachmark2 View Post

GPU architectures generally scale more or less linearly with added stream processors/CUDA cores.

That's not the problem. Of course they are highly parallel with little overheads of parallelism (that CPUs often have).

The point is you can't just take a tiny tablet-grade chip, enlarge it, about 4 X the size or more and then overclock it on top (compared to the frequency you had) and then say TADA! Titan Maxwell!

It doesn't work like that. Yields will be insane(ly low). Stability will be low. In fact it may not even run for 2 seconds.

There is a reason large K is harder to stabilize than small K. So a tiny chip will be more easy to stabilize. I seriously doubt the relationship is linear either. Chances are enlarging the chip of the Titan won't give anything important. But shrinking the smallest chip that was available apparently proves marvelous.


The AMD example was an argument. I wasn't responding to him. It was basically "AMD has proven they are around +/- 5% difference in performance for similar die sizes with NVIDIA. If they couldn't make a large die that is spectacularly faster than NVIDIA's, on the same size, so can't NVIDIA".
PC
(9 items)
 
  
CPUMotherboardGraphicsRAM
4790K 4.6G 1.24v/1.74v, 5.1G 1.35v validation. GA-Z97X-Gaming 7 Tri-X R9 290 1100/1350 +0.012v G.Skill 2400 c10 
CoolingMonitorPowerCase
Noctua NH-D15 24EA53 IPS 76Hz OC EVGA 1000 G2 Phanteks Enthoo Pro 
Audio
SoundMagic E10 
  hide details  
Reply
PC
(9 items)
 
  
CPUMotherboardGraphicsRAM
4790K 4.6G 1.24v/1.74v, 5.1G 1.35v validation. GA-Z97X-Gaming 7 Tri-X R9 290 1100/1350 +0.012v G.Skill 2400 c10 
CoolingMonitorPowerCase
Noctua NH-D15 24EA53 IPS 76Hz OC EVGA 1000 G2 Phanteks Enthoo Pro 
Audio
SoundMagic E10 
  hide details  
Reply
post #34 of 51
Quote:
Originally Posted by Mand12 View Post

Again, having this scaling is the ENTIRE point of Maxwell.

Correction. That's the propaganda of it.
PC
(9 items)
 
  
CPUMotherboardGraphicsRAM
4790K 4.6G 1.24v/1.74v, 5.1G 1.35v validation. GA-Z97X-Gaming 7 Tri-X R9 290 1100/1350 +0.012v G.Skill 2400 c10 
CoolingMonitorPowerCase
Noctua NH-D15 24EA53 IPS 76Hz OC EVGA 1000 G2 Phanteks Enthoo Pro 
Audio
SoundMagic E10 
  hide details  
Reply
PC
(9 items)
 
  
CPUMotherboardGraphicsRAM
4790K 4.6G 1.24v/1.74v, 5.1G 1.35v validation. GA-Z97X-Gaming 7 Tri-X R9 290 1100/1350 +0.012v G.Skill 2400 c10 
CoolingMonitorPowerCase
Noctua NH-D15 24EA53 IPS 76Hz OC EVGA 1000 G2 Phanteks Enthoo Pro 
Audio
SoundMagic E10 
  hide details  
Reply
post #35 of 51
I don't think you guys are understanding what fateswarm is trying to say. Basically you can't just take a design that works like the 750Ti and then suddenly decide you are going to make the same chip but 200% bigger, it doesn't work like that or we would have seen a replacement for the 7970 long after release. You have to go back and re-design the chip a little to make it run even a little bit stable. Other than heat a power issues, it's why you see the biggest chips overclocking more poorly than the smaller chips. Not always the case of course as we have a lot of custom PCB design and such effecting the overclocking capabilities. But put on a level playing field I'm sure a 7870 would overclock further than a 7970 on the same volts.
post #36 of 51
Quote:
Originally Posted by ObscureParadox View Post

I don't think you guys are understanding what fateswarm is trying to say. Basically you can't just take a design that works like the 750Ti and then suddenly decide you are going to make the same chip but 200% bigger, it doesn't work like that or we would have seen a replacement for the 7970 long after release. You have to go back and re-design the chip a little to make it run even a little bit stable. Other than heat a power issues, it's why you see the biggest chips overclocking more poorly than the smaller chips. Not always the case of course as we have a lot of custom PCB design and such effecting the overclocking capabilities. But put on a level playing field I'm sure a 7870 would overclock further than a 7970 on the same volts.

This is exactly how it works. Good post.
The girlfriend.
(15 items)
 
The Mistress
(13 items)
 
Media Server
(11 items)
 
CPUMotherboardGraphicsRAM
A8-6410 Lenovo Lancer 4B2 K16.3 R5 128 Shaders/M230 Hynix 8GB DDR3 1600 
Hard DriveHard DriveOSMonitor
Samsung 840 120 GB SSD Seagate Momentus 1TB 5400rmp Win 8.1 CMN1487 TN LED 14" 1366*768 
KeyboardPowerMouseMouse Pad
Lenovo AccuType 2900mAh/41Wh Elan Trackpad/Logitech M90 Super Flower 
Audio
AMD Avalon(Connexant) 
  hide details  
Reply
The girlfriend.
(15 items)
 
The Mistress
(13 items)
 
Media Server
(11 items)
 
CPUMotherboardGraphicsRAM
A8-6410 Lenovo Lancer 4B2 K16.3 R5 128 Shaders/M230 Hynix 8GB DDR3 1600 
Hard DriveHard DriveOSMonitor
Samsung 840 120 GB SSD Seagate Momentus 1TB 5400rmp Win 8.1 CMN1487 TN LED 14" 1366*768 
KeyboardPowerMouseMouse Pad
Lenovo AccuType 2900mAh/41Wh Elan Trackpad/Logitech M90 Super Flower 
Audio
AMD Avalon(Connexant) 
  hide details  
Reply
post #37 of 51
Lower end chips tend to overclock better than flagships because flagships are already pushed from the factory to a high level to remain competitive overall. An excellent example of this would be the GTX 470 which had a stock clock for 607 mhz next to the GTX 480's 700 mhz.

AMD and Nvidia only have so many ways they can create market differentiation in a given generation of GPUs. Generally speaking, Nvidia makes between 4 and 6 different chips per generation (To continue the Fermi example, there was GF100, GF104, GF106, and GF108). Within each chip, there are further opportunities for differentiation that can be attained by disabling certain parts of a "full featured" die. The GTX 470 had 32 of its CUDA cores disabled. This had the effect of dropping the performance to about 92% give or take of a 480. But then Nvidia lowered the core clock from 700mhz to 607mhz.

These two combined led to the chip consuming less power but also being slower than the GTX 480 by about 15-20%. This gave Nvidia two products to bring to market and to price appropriately.

On the other side of the coin, it would be sheer lunacy to think that silicon wafers could be scaled far beyond GTX 480 or GTX 780 or Tahiti sized and still retain any form of efficiency or usefulness in a home scenario.

If I could have a main point with regards to the entire discussion:

Scaling only works to a point. That I agree with you on. Nvidia did not have the practical ability to make a bigger or better GF100 or GF110, nor do they have any headroom left on GK110. However, when considering a design based upon the same 28nm process as the GeForce 6xx and clearly an entry/mid level GPU, all one must do is look to the past to see what and how Nvidia will use their new architecture.

Just compare the GTX 660 to the GTX 780ti. One is GK106 and the other is GK110. The difference in performance can be almost entirely explained as a combination of clockspeed and extra cores. Scaling up the design absolutely worked then, and there's no reason to expect it not to for Maxwell.
Heisenberg
(10 items)
 
Redemption
(18 items)
 
 
CPUGraphicsRAMHard Drive
Intel Core 2 Extreme @ 2.8GHz (upgrade) Intel GMA 950 2 x 2 DDR2 667MHz (upgrade) 120GB Mushkin ECO2 
Optical DriveOSMonitorPower
Replaced with 42 W/hr aux battery Mint 17.3 1280x800 Dell 65w Dell Adapter 
CaseOther
Latitude D630 Chassis Intel 7260-802.11ac 2x2 
CPUMotherboardGraphicsRAM
Not telling AsRock something Pixel Pusher 2 x 8 GB AMD Radeon RAM @ 1600mhz (All the VM's) 
Hard DriveHard DriveHard DriveOptical Drive
Crucial M4 128GB (8 second boot) Western Digital RE4 1TB Western Digital RE4 2TB Pioneer Blu-Ray Reader/Writer 
CoolingOSMonitorKeyboard
Cooler Master Hyper 212 Evo Windows 8.1 Pro x64 Dell U2713HM IBM Model M (CLICKY!!) 
PowerCaseMouseAudio
Antec HCG-900 NZXT Source 210 (it was cheap) Logitech G502 Schiit Modi 
AudioAudio
Sennheiser HD 598 M-Audio AV40's 
CPUMotherboardGraphicsRAM
Intel Core i5-4200U Microsoft BGA1168 Intel HD 4400 8 GB LPDDR3 Dual Channel @ 1600mhz 
Hard DriveOSMonitorKeyboard
SK Hynix 256GB SSD Windows 8.1 Professional 1920 x 1080p IPS Display Microsoft Touch Cover 2 
Power
48w PSU 
  hide details  
Reply
Heisenberg
(10 items)
 
Redemption
(18 items)
 
 
CPUGraphicsRAMHard Drive
Intel Core 2 Extreme @ 2.8GHz (upgrade) Intel GMA 950 2 x 2 DDR2 667MHz (upgrade) 120GB Mushkin ECO2 
Optical DriveOSMonitorPower
Replaced with 42 W/hr aux battery Mint 17.3 1280x800 Dell 65w Dell Adapter 
CaseOther
Latitude D630 Chassis Intel 7260-802.11ac 2x2 
CPUMotherboardGraphicsRAM
Not telling AsRock something Pixel Pusher 2 x 8 GB AMD Radeon RAM @ 1600mhz (All the VM's) 
Hard DriveHard DriveHard DriveOptical Drive
Crucial M4 128GB (8 second boot) Western Digital RE4 1TB Western Digital RE4 2TB Pioneer Blu-Ray Reader/Writer 
CoolingOSMonitorKeyboard
Cooler Master Hyper 212 Evo Windows 8.1 Pro x64 Dell U2713HM IBM Model M (CLICKY!!) 
PowerCaseMouseAudio
Antec HCG-900 NZXT Source 210 (it was cheap) Logitech G502 Schiit Modi 
AudioAudio
Sennheiser HD 598 M-Audio AV40's 
CPUMotherboardGraphicsRAM
Intel Core i5-4200U Microsoft BGA1168 Intel HD 4400 8 GB LPDDR3 Dual Channel @ 1600mhz 
Hard DriveOSMonitorKeyboard
SK Hynix 256GB SSD Windows 8.1 Professional 1920 x 1080p IPS Display Microsoft Touch Cover 2 
Power
48w PSU 
  hide details  
Reply
post #38 of 51
Quote:
Originally Posted by coachmark2 View Post

Lower end chips tend to overclock better than flagships because flagships are already pushed from the factory to a high level to remain competitive overall. An excellent example of this would be the GTX 470 which had a stock clock for 607 mhz next to the GTX 480's 700 mhz.

AMD and Nvidia only have so many ways they can create market differentiation in a given generation of GPUs. Generally speaking, Nvidia makes between 4 and 6 different chips per generation (To continue the Fermi example, there was GF100, GF104, GF106, and GF108). Within each chip, there are further opportunities for differentiation that can be attained by disabling certain parts of a "full featured" die. The GTX 470 had 32 of its CUDA cores disabled. This had the effect of dropping the performance to about 92% give or take of a 480. But then Nvidia lowered the core clock from 700mhz to 607mhz.

I don't mean a percentage overclock I mean a max frequency overclock. More cores = more chance for instability = more design work required.
post #39 of 51
Quote:
Originally Posted by fateswarm View Post

Correction. That's the propaganda of it.

Just because you say it's propaganda does not make it not true.
Quote:
Originally Posted by ObscureParadox View Post

Basically you can't just take a design that works like the 750Ti and then suddenly decide you are going to make the same chip but 200% bigger, it doesn't work like that

Why not? This is exactly what Nvidia has said their design approach to Maxwell would be. GM107 is 148 square mm, GK110 is 561 square mm. Just shy of four times the area. And in an architecture designed with scaling as a primary design goal, what makes you think it doesn't work like that? Seriously, I want some specifics here, this is a direct and honest question.

You can't just decree "it's too hard, nop can't be done nuppers!" and just expect it to end with that. Literally the entire Maxwell endeavor was to improve two things: efficiency and scaling. To write off any actual improvement as propaganda just because they haven't done it before makes it impossible for you to accept that real innovation occasionally happens. The 750 data backs them up, whereas all you have is "it doesn't work that way."

Clearly it does.
Edited by Mand12 - 3/23/14 at 8:15pm
post #40 of 51
Quote:
Originally Posted by Mand12 View Post

Just because you say it's propaganda does not make it not true.
Why not? This is exactly what Nvidia has said their design approach to Maxwell would be. GM107 is 148 square mm, GK110 is 561 square mm. Just shy of four times the area. And in an architecture designed with scaling as a primary design goal, what makes you think it doesn't work like that? Seriously, I want some specifics here, this is a direct and honest question.

You can't just decree "it's too hard, nop can't be done nuppers!" and just expect it to end with that. Literally the entire Maxwell endeavor was to improve two things: efficiency and scaling. To write off any actual improvement as propaganda just because they haven't done it before makes it impossible for you to accept that real innovation occasionally happens. The 750 data backs them up, whereas all you have is "it doesn't work that way."

Clearly it does.

Kind of like how some people just write off free sync right?
Sorry AMD
(13 items)
 
   
CPUMotherboardGraphicsRAM
i5-8400 ROG STRIX Z370-I GAMING ZOTAC GTX 1080 Ti Mini 16gb Gskill Ripjaw V 3600mhz CL16 
Hard DriveCoolingOSMonitor
256gb m.2 Crucial noctua nh-l9i Windows 10 Pro AOC G2460PF 
KeyboardPowerCaseMouse
Mechanical Eagle Z-77 (Cheap) Corsair SF600 Fractal Design Node 202 Logitech G400 
Audio
Sennheiser PC 350 (hero mod) 
CPUMotherboardGraphicsRAM
AMD Ryzen™ 1700 Asrock Fatal1ty x370 Gaming K4 Zotac mini GTX 1080ti TridentZ 16GB (2x8GB) DDR4 
Hard DriveHard DriveOSMonitor
256GB M.2 SSD 1TB HDD Windows 10 Pro Acer XG270HU 
KeyboardPowerCaseMouse
Logitech G15 750W Thermaltake Toughpower Gold Rated Thermaltake View 27 w/ riser Logitech G40 
CPUMotherboardGraphicsRAM
fx-8350 @5ghz 990fxa-ud3 rev 1.1 PowerColor 290x @1150/1500mhz Gskill 8gig 1600mhz (2x4) @2033 mhz 
Hard DriveCoolingOSPower
1TB WD Caviar Black rs360 raystorm kit Windows 8 Pro 700 Watt Corsair 
CaseMouse
HAF 912  Logitech G400 
  hide details  
Reply
Sorry AMD
(13 items)
 
   
CPUMotherboardGraphicsRAM
i5-8400 ROG STRIX Z370-I GAMING ZOTAC GTX 1080 Ti Mini 16gb Gskill Ripjaw V 3600mhz CL16 
Hard DriveCoolingOSMonitor
256gb m.2 Crucial noctua nh-l9i Windows 10 Pro AOC G2460PF 
KeyboardPowerCaseMouse
Mechanical Eagle Z-77 (Cheap) Corsair SF600 Fractal Design Node 202 Logitech G400 
Audio
Sennheiser PC 350 (hero mod) 
CPUMotherboardGraphicsRAM
AMD Ryzen™ 1700 Asrock Fatal1ty x370 Gaming K4 Zotac mini GTX 1080ti TridentZ 16GB (2x8GB) DDR4 
Hard DriveHard DriveOSMonitor
256GB M.2 SSD 1TB HDD Windows 10 Pro Acer XG270HU 
KeyboardPowerCaseMouse
Logitech G15 750W Thermaltake Toughpower Gold Rated Thermaltake View 27 w/ riser Logitech G40 
CPUMotherboardGraphicsRAM
fx-8350 @5ghz 990fxa-ud3 rev 1.1 PowerColor 290x @1150/1500mhz Gskill 8gig 1600mhz (2x4) @2033 mhz 
Hard DriveCoolingOSPower
1TB WD Caviar Black rs360 raystorm kit Windows 8 Pro 700 Watt Corsair 
CaseMouse
HAF 912  Logitech G400 
  hide details  
Reply
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Hardware News
Overclock.net › Forums › Industry News › Hardware News › [digitimes] TSMC 20nm wafer production delayed temporarily, but will not affect shipments