Pascal Tesla availability clue

[Creig]

Something like a GPU is easy to make redundant blocks for tho to get higher yields.

Yes, there are redundant blocks. But not all defects are considerate enough to land in areas that can utilize them. Thus yields will always decrease as the size and complexity of the die increases.

 

[NTMBK]

Dont forget parametric yields, you may get the entire die intact but not all of them will run at your targeted frequency/voltages.

Remember the first GK110 dies in retail were the GTX 780, cut down and at low clocks.

Wasn't the Titan first? Though that was also a partially disabled chip.

 

[MrTeal]

Wasn't the Titan first? Though that was also a partially disabled chip.

Sept 2012 - GK110 to Oak Ridge
Nov 2012 - Tesla K20 & K20X - 2496/2688 CC
Feb 21, 2013 - Titan - 2688 CC
May 23, 2013 - GTX 780 - 2304 CC
Summer 2013 - Quadro K6000 - 2880 CC
Nov 7, 2013 - GTX 780 Ti - 2880 CC
Feb 18, 2014 - Titan Black - 2880 CC

It wasn't until mid 2013 that full GK110 appeared, which was a year and a half after the first 28nm GPUs launched.

 

[AtenRa]

Wasn't the Titan first? Though that was also a partially disabled chip.

Yes, TITAN has more shaders at lower clocks and GTX 780 has less shaders at higher clocks.

Both cut down versions of the GK110. Titan was released on February 2013, 11 months after GK104 and GTX 780 three months later on May 2013 or 14 months after GK104.

GTX 780Ti uses a full GK110 (if i remember correctly it had a metal respin) and was released in November 2013.

 

[NTMBK]

Sept 2012 - GK110 to Oak Ridge
Nov 2012 - Tesla K20 & K20X - 2496/2688 CC
Feb 21, 2013 - Titan - 2688 CC
May 23, 2013 - GTX 780 - 2304 CC
Summer 2013 - Quadro K6000 - 2880 CC
Nov 7, 2013 - GTX 780 Ti - 2880 CC
Feb 18, 2014 - Titan Black - 2880 CC

It wasn't until mid 2013 that full GK110 appeared, which was a year and a half after the first 28nm GPUs launched.

I was referring to retail, as AtenRa was discussing

 

[MrTeal]

I was referring to retail, as AtenRa was discussing

Well, Tesla was kind of retail, though obviously not a GPU. I wasn't disagreeing with you though, just providing the GK110 timeline. It's interesting from a historical perspective, since even with a couple very lucrative supercomputer contracts, the 561mm² didn't show up as a fully functioning die until 20 months after GK104 launched. That includes the >US$3000 Tesla cards and whatever nVidia charged Oak Ridge for the K20X's used in Titan. Making a huge die and yielding on a new process is tough, and 28nm isn't even considered a really contentious node like TSMC 40nm was.

 

[NTMBK]

Well, Tesla was kind of retail, though obviously not a GPU. I wasn't disagreeing with you though, just providing the GK110 timeline. It's interesting from a historical perspective, since even with a couple very lucrative supercomputer contracts, the 561mm² didn't show up as a fully functioning die until 20 months after GK104 launched. That includes the >US$3000 Tesla cards and whatever nVidia charged Oak Ridge for the K20X's used in Titan. Making a huge die and yielding on a new process is tough, and 28nm isn't even considered a really contentious node like TSMC 40nm was.

It helps that HPC algorithms are inherently designed to scale to massive numbers of processors, so multi-GPU cards aren't a problem. Hence the K10, easier to yield two medium sized chips than one large one.

 

[imaheadcase]

They mentioned three new processors: a new Xeon (Broadwell-EP), Xeon Phi mk2 (Knights Landing), and Pascal.

What processors do you think they were referring to if not the three above?

Pascal IN a cpu? *gasp*

 

[Glo.]

https://forum.beyond3d.com/threads/nvidia-pascal-speculation-thread.55552/page-39#post-1895652 Great post about the affect of FinFET on GPU dies.

 

[96Firebird]

I'm confused... Do people not know that GPUs are processors? Or is it sarcasm that I'm not picking up on?

 

[cytg111]

This thread is a rollercoaster... Is Pascal launch imminent or ~6 months out?

 

[Glo.]

16 GB HBM2 versions are possible right now, if Nvidia went with Samsung as HBM2 supplier. GDDR5X starts in summer. So no HBM2 32 GB GPUs and GDDR5X until late Q3.

What is worrying is the fact that Nvidia launches new GPU from Maxwell Lineup, and that there is no GPU dies on Zauba shipping.

 

[MrTeal]

16 GB HBM2 versions are possible right now, if Nvidia went with Samsung as HBM2 supplier. GDDR5X starts in summer. So no HBM2 32 GB GPUs and GDDR5X until late Q3.

What is worrying is the fact that Nvidia launches new GPU from Maxwell Lineup, and that there is no GPU dies on Zauba shipping.

No GPU dies on Zauba doesn't necessarily mean there aren't any, it just means they haven't imported any into India with a description that makes it easy to search.

 

[Atreidin]

This thread is a rollercoaster... Is Pascal launch imminent or ~6 months out?

The definitions used seem to be based on what one means by being released, like existing in some form to some customer somewhere, or in a form normal people could buy one from a retailer. Some people seem to be using the first case and pretending that is what everyone cares about, then flamboyantly trying to rub it in people's faces, when really the issue was the retail timeframe.

 

[jpiniero]

16 GB HBM2 versions are possible right now, if Nvidia went with Samsung as HBM2 supplier. GDDR5X starts in summer. So no HBM2 32 GB GPUs and GDDR5X until late Q3.

The Tesla has been rumored to be two die, so it's possible by 32 GB they mean 16 x 2.

The problem is that you can't just take the yields of smaller chips and apply it directly to larger chips. The larger the die you try to harvest from any given wafer, the greater the chance that any particular die will have a defect in it thus lowering yields. Here's a graphical representation of the situation:

Yeah I know, which is why I included cut down dies in the percentage. Get a defect in one of the compute units, just disable the unit.

This thread is a rollercoaster... Is Pascal launch imminent or ~6 months out?

Seems pretty likely that Tesla GP100 will launch in April. If there's a Titan included in this launch, it will be well over $1K. Everything else seems more like 4Q.