NVIDIA Pascal Thread

[Grooveriding]

2017 Q1 for just the Tesla of GP100. Looks like no GP100 Geforce for at least a year Crappy.

 

[SteveGrabowski]

Damn, Q1 2017 for the Tesla card.

 

[Kenmitch]

"GP100 is in volume production today"

Shipping....Pause....Soon.

 

[xthetenth]

Gonna be a long time before this is on the desktop.

Agreed. Looking good for Christmas 2017 though.

 

[poofyhairguy]

2017 Q1 for just the Tesla of GP100. Looks like no GP100 Geforce for at least a year Crappy.

Didn't we already know that? I thought the whole plan was mid range chips this year to double dip next year, ala Maxwell and Kepler.

 

[xpea]

P100 is by far the most amazing and most ambitious chip ever created

 

[jpiniero]

600mm2 same size as GM200 but on 16nm

Going 1/2 DP must be taking up a ton of space. It's only 75% faster than GM200 in SP. So you figure that's the upside of any possible dGPU from nVidia for maybe 2-3 years.

 

[MrTeal]

Didn't we already know that? I thought the whole plan was mid range chips this year to double dip next year, ala Maxwell and Kepler.

I think most people were, but there was still some hope out there. Is Q1'17 for system builders with P100 modules in their systems, or PCIe based Tesla cards?

I'm still hoping that we might see a Titan in summer '17, but we'll see.

 

[Grooveriding]

Didn't we already know that? I thought the whole plan was mid range chips this year to double dip next year, ala Maxwell and Kepler.

Yeah it was expected, but this confirms it.

The X00 cards are the impressive cards. 980 was a huge disappointment for someone on a 780ti, if we're getting mid range geforce cards as flagships again, I'd like to see a better improvement than that from the new geforce flagship. Upgrading for 10%, and 20% after driver shens down the road, is underwhelming.

 

[Glo.]

http://www.bitsandchips.it/english/52-english-news/6620-rumor-nvidia-pascal-timeline

So Fottemberg was ultimately right.

P.S. http://www.bitsandchips.it/english/...scal-in-trouble-with-asyncronous-compute-code

P.S.2 https://devblogs.nvidia.com/parallelforall/inside-pascal/

Anyone understands anything from that? . 300W TDP for Tesla P100.

 

[MrTeal]

http://www.bitsandchips.it/english/52-english-news/6620-rumor-nvidia-pascal-timeline

So Fottemberg was ultimately right.

Not really. He was saying GP100 at GTC, but this isn't a launch of a card, it's an announcement of Pascal. Titan in Q4 2016 / Q1 2017 is right out now, if they're talking about shipping the first systems with P100 in Q1.

 

[Maverick177]

Specs: 3840 CUDA cores, Base clock 1328 MHz, Boost: 1480 MHz, DP: 5304, 16GB HBM2 Samsung, 300W TDP, 15.3B sistors, 610mm^2.

https://devblogs.nvidia.com/parallelforall/inside-pascal/

Edit: Meh.

 

[Kenmitch]

I bailed out on the feed....Nothing of interest too me so far. If any Pascal for the mortals news drops I'm sure it'll be posted in here anyways.

Just helped my son build a new gaming rig with a 6700k. I advised him to hold off on a GPU purchase for the time being. His 7870 does his rig no justice currently. Eagerly awaiting news from both camps to decide on what GPU will work best for him.

 

[SteveGrabowski]

Damn I was hoping to hear about GP104

 

[IllogicalGlory]

That clock is through the roof. I wonder what kind of overclocking we're going to see. Same as Maxwell or are they being more aggressive at stock?

Wow. 300W too. Beast. Only 3584 Cuda cores though.

 

[Glo.]

Specs: 3840 CUDA cores, Base clock 1328 MHz, Boost: 1480 MHz, DP: 5304, 16GB HBM2 Samsung, 300W TDP, 15.3B sistors, 610mm^2.

https://devblogs.nvidia.com/parallelforall/inside-pascal/

Edit: Meh.

3584 CUDA cores.

 

[xpea]

inside Pascal:
https://devblogs.nvidia.com/parallelforall/inside-pascal/

again, who said Pascal is Maxwell on 16FF ?

 

[n0x1ous]

That clock is through the roof. I wonder what kind of overclocking we're going to see. Same as Maxwell or are they being more aggressive at stock?

Wow. 300W too. Beast. Only 3584 Cuda cores though.

thats fp32 cores. + 1792 fp 64 cores means it is actually 5376

Also confirmed 610mm2 - even bigger than GM200. New node and right to the reticle limit right away. incredible

 

[Pinstripe]

I'm not sure whether Tesla P100 and GP100 are even the same chip.

Edit: Yes it is.

 

[nvgpu]

Those people that said Pascal is just Maxwell shrink don't know what they're talking about as usual.

 

[poofyhairguy]

While a GP100 SM has half the total number of CUDA Cores of a Maxwell SM, it maintains the same register file size and supports similar occupancy of warps and thread blocks.

Sooooo....no async?

 

[R0H1T]

That clock is through the roof. I wonder what kind of overclocking we're going to see. Same as Maxwell or are they being more aggressive at stock?

Wow. 300W too. Beast. Only 3584 Cuda cores though.

I think that's the limit with Pascal, they've done this in order to counter Intel on the HPC front.
I don't expect more than 5~15% OC potential with the fully enabled GP100 chip, that too with copious amounts of cooling.

 

[antihelten]

Going 1/2 DP must be taking up a ton of space. It's only 75% faster than GM200 in SP. So you figure that's the upside of any possible dGPU from nVidia for maybe 2-3 years.

Yes and no. P100 is 15B xtors versus 8B for GM200, an increase of 87.5%. P100 is 5376 CUDA cores (FP32 and FP64) versus 3168 on GM200, an increase of 69.7%.

So an average CUDA cores is roughly 10% more xtors on P100, not really a huge amount, especially when considering that P100 is 33% FP64 cores, versus only 3% on GM200.

The more interesting thing is arguably that FF16+ isn't scaling as well as some predicted, with only an 88.4% increase in transistor density, whereas people where predicting a 120% increase relative to 28nm. At least it appears to clock incredibly well.

Those people that said Pascal is just Maxwell shrink don't know what they're talking about as usual.

You're right, Pascal is Maxwell shrink plus the FP64 CUDA cores from Kepler (GK110) added back in

 

[MrTeal]

Sweet, two Xeons and 8 P100's for only $130k.

 

[Creig]

Also confirmed 610mm2 - even bigger than GM200. New node and right to the reticle limit right away. incredible

Does that 610 mm^2 include everything on the substrate or just the GPU? My initial guess is that it's the whole shebang.