Nvidia reveals key GP100 Pascal details at GTC 2016 (610mm2, 15.3B transistors, 300W TDP, 3584 cores, 1480MHz boost clock)

[don_svetlio]

21 minutes ago, i_build_nanosuits said:

02:27PM EDT - Chip on Wafer on Substrate, the largest such chip ever made blah blah blah deep learning

02:28PM EDT - And that's a wrap

 

02:31PM EDT -  ''Bolt me a freaking display port on the back of that thing i need to play arkham knight on that shit'' --- Jen-Hsun Huang

02:33PM EDT -  ''i still can't run that fucking smoke at 60FPS man!! you fucked up! '' --- Jen-Hsun Huang

02:35PM EDT -  ''also... 300W? really?! '' --- Jen-Hsun Huang

 

 

 

MY FOOD IS READY!

[Notional]

7 hours ago, patrickjp93 said:

It tells you nothing. The Tesla K80 uses 2 GK 210s which are cut down as well, and that's on the 28nm process with very high yields. If anything, this says Nvidia is confident AMD won't be able to compete.

Actually it tells us a lot. NVidia's strategy is to push out Pascal and HBM out first on the professional market, at very high margins (kinda what they did with the Titan's too). As these chips are huge and on a new node, the entire point is to pump out as many chips as possible (remember their AI server uses 8 p100 chips), so it's more important to have a lot of chips that are cut down a little, than have full fat chips.

The point is not that TSMC has worse yields than GloFo, just that NVidia wants this chip out on a high margin market afap (løl).

 

As for K80 it was not a good card. The Hawaii based Firepro cards were much faster after all.

[Majestic]

That much cache... these look expensive as hell..

[TheRandomness]

12 minutes ago, don_svetlio said:

MY FOOD IS READY!

Thermi resurrected?

[patrickjp93]

1 hour ago, Notional said:

Actually it tells us a lot. NVidia's strategy is to push out Pascal and HBM out first on the professional market, at very high margins (kinda what they did with the Titan's too). As these chips are huge and on a new node, the entire point is to pump out as many chips as possible (remember their AI server uses 8 p100 chips), so it's more important to have a lot of chips that are cut down a little, than have full fat chips.

The point is not that TSMC has worse yields than GloFo, just that NVidia wants this chip out on a high margin market afap (løl).

 

As for K80 it was not a good card. The Hawaii based Firepro cards were much faster after all.

I'm afraid the leading authorities on compute beg to differ. AMD's FirePro cards only boast a theoretical advantage. Their OpenCL drivers are still terrible. Until AMD starts producing far better numbers in LinPack, you cannot say their compute performance is better.

[226477_1454181668]

3 hours ago, TheRandomness said:

Thermi resurrected?

Looks like it is, another thing I noticed is that chip is very big. And it doesnt look like nvidia can make a bigger chip with that hbm 2 size constraint. I always though that 14nm would increase density, but it doesn't seem to have increased it. 

[Notional]

One thing I'm interested in, now that a lot of people have whined about Hawaii and it's TDP: Are they going to whine about Pascal p100 being 300w tdp, which is worse?

[Zah]

To make sure, these are not consumer grade chips, correct?  These are for professional and scientific applications?

[Shahnewaz]

29 minutes ago, Zah said:

To make sure, these are not consumer grade chips, correct?  These are for professional and scientific applications?

Yes. GP100 doesn't have any ROPs or Polymorph engines, meaning this part is aimed suqarely for the HPC market.

[shdowhunt60]

13 hours ago, Notional said:

One thing I'm interested in, now that a lot of people have whined about Hawaii and it's TDP: Are they going to whine about Pascal p100 being 300w tdp, which is worse?

Of course not. nVidia can do no wrong.