Nvidia PASCAL...next-gen GPU.

[cadaveca]

NV-LINK

3D Memory.


PIC:

 

[BiggieShady]

Lovely ... following live blog at http://www.anandtech.com/show/7894/nvidia-gtc-2014-keynote-live-blog

Seems Volta is Pascal now

 

[Hilux SSRG]

Seems Volta is Pascal now

I still think Volta will slot in for release in 2015/2016. Pascal is more a goal post with no timeline mentioned yet.

 

[Ferrum Master]

photoshopped

 

[True Christian]

I usually skip 2 generations of GPUs so I'll probably go with a GTX 1060 8GB in 3 1/2 years.

 

[TheoneandonlyMrK]

I still think Volta will slot in for release in 2015/2016. Pascal is more a goal post with no timeline mentioned yet.

Volta Is rumoured to have been pushed back after pascal. As I said tsv 3d stacked multi die chips are some ways off with only stacked cube memory (already heavily developed by others in the open ish) so volta was a big ask this soon .

I do think it's very very wrong that a ceo of a company can spout utter shit year after year and never get called on it , if i had invested in nvidia, I would not like the apparently vague and rapidly changing forcasts.

Looks to be an interesting prospect ,except no one has a clue yet exactly what the heck it is or how im going to be able to fit eight in my pc

No need to retort ill sumarise .
Tech is good but odd and strange
Nvidia yet again are divisive and foolish

 

[KainXS]

Denver vanished too on the mobile side, wonder whats up at nv

 

[15th Warlock]

Please excuse my ignorance, but... isn't that module lacking any sort of interface?

Dunno if that's the standard PCB format GPU developers like AMD or Nvidia use, but how do you connect it to a PC?

 

[Xzibit]

Please excuse my ignorance, but... isn't that module lacking any sort of interface?

Dunno if that's the standard PCB format GPU developers like AMD or Nvidia use, but how do you connect it to a PC?

Mezzanine Card

Connector/slot/plug should be on the other side, it should just fit on top of the mainboard and screw in to the support.

You can see the connector on the back in this video @ 1:19-1:26

 

[15th Warlock]

Mezzanine Card

Connector/slot/plug should be on the other side, it should just fit on top of the mainboard and screw in to the support.

You can see the connector on the back in this video @ 1:19-1:26

Thanks for your very informative reply

 

[newconroer]

I am super surprised this thread isn't thirty pages long, fanboyism or not. If Nvidia pulls this off, it really is the future.


Here's a good read (hope this isn't a repost)

http://wccftech.com/nvidia-pascal-a...-analysis-stacked-dram-nv-link/#ixzz2ycEpG1W7

I think they look awesome - understandably they may not be exactly like that, but I suspect they'll try to make them that way.


What tickles my fancy is seeing the redesign of the motherboards, I imagine the space they take width wise would cover almost three PCI E lanes and force PCI E for sound cards elsewhere.
Also, GPUs that lay flat! What type of heat /fan shroud would it require? Will they still be adjacent to the rear of the motherboard as to exhaust out the back?

 

[TheoneandonlyMrK]

Definitely Interesting , but it's the Nv link that's most interesting to me, I don't understand how they are going to sell it to motherboard maker's since it is soo proprietary and different in config and design from what we have now..
I'm sure the institutions organising the pciex standard have plans for the future too which could get interesting.
Then you have the rumours of intel allegedly moving to make discrete cards null by restricting the available pciex lanes more in future chips.
Could all get a bit odd or crazy good.

 

[MxPhenom 216]

Probably what I am going to wait for unless 2nd Generation Maxwell is as good as I hope.

 

[FordGT90Concept]

http://www.nvidia.com/object/nvlink.html

It's pretty obvious to me this is extremely proprietary stuff to the extent that the CPUs, motherboards, and GPUs have to be designed explicitly for it--typical Cray/IBM supercomputing engineering. This will never see use in consumer products--it's made to order for $100+ million machines. It would be like redesigning Xeon to expand it's FPUs by commandeering Xeon Phi cards.


From what I see, Pascel is just NVIDIA's interpretation of AMD Fury: die shrink, some tweaks, HBM2 controller.

 

[cadaveca]

http://www.nvidia.com/object/nvlink.html

It's pretty obvious to me this is extremely proprietary stuff to the extent that the CPUs, motherboards, and GPUs have to be designed explicitly for it--typical Cray/IBM supercomputing engineering. This will never see use in consumer products--it's made to order for $100+ million machines. It would be like redesigning Xeon to expand it's FPUs by commandeering Xeon Phi cards.


From what I see, Pascel is just NVIDIA's interpretation of AMD Fury: die shrink, some tweaks, HBM2 controller.

Yet why would it be anything BUT as you say? I think that understanding what truly holds GPUs back from greater performance makes design choice reasoning pretty obvious. PASCAL is a design meant to target specific needs, yet like any other product, to do so with the lowest cost possible.

 

[Steevo]

http://www.techpowerup.com/reviews/NVIDIA/GTX_980_PCI-Express_Scaling/21.html

Haha, we make dem buy new proprietary stuff so we can make more money, tell them they need it. hahahah.


Or NVLink is exactly as it sounds "EnvyLink" make sure you have the new NVlink, cause you need more, and more is always better, so you need more. More proprietary.

 

[Frag_Maniac]

...I don't understand how they are going to sell it to motherboard maker's since it is soo proprietary and different in config and design from what we have now..

If the Hydra chip could make it onto gaming MBs, I don't see why this can't. Hell, Hydra was just a way to use dissimilar GPUs on one board. This blows hydra out of the water practicality wise. How many people would really use an AMD card with an Nvidia one?

The CPU has to be compatible with it though, so CPU manufacturers would have to jump on board, though I'm sure many would be OK if just Intel did, since obviously AMD is in big competition with Nvidia.

I've seen an article talking of an NVLink version 1.0 configuration with multi GPU setup being ran with Pci Ex switches between the GPU and CPU though, indicating Nvidia have already planned for at least inter GPU NVLink connectivity on existing consumer MB architecture.

Currently only IBM has adopted this tech for their Power series chips, but I don't see why Intel couldn't do the same for consumer grade CPUs, with a lower version of it with enough bandwidth to at least equal that of today's CPUs and GPUs.

NVLink data rate varies from about 30GB/s all the way up to 200GB/s, depending on the version and application. That's why Nvidia says Pascal will be as much as 10x more powerful than Maxwell.

That said, I'd almost prefer Intel had come up with this tech, as it might stand a better chance of being universally accepted. Nvidia seem to get treated like an evil empire when they dream up stuff like this.

 

[FordGT90Concept]

There's a huge difference between video processing and high performance computing processing. Most of video processing is output to a display where high performance computing processing needs to be returned to the CPU so the CPU can save it or manipulate it and send it somewhere else. NVlink will no doubt be a big leg up on Xeon Phi which I'm sure has Intel concerned. Because of that, PCI-SIG easily could introduce a HPC dedicated slot just so Xeon Phi (and later Tesla) can use it. This could be good news for PC gamers but it will be a half a decade until it trickles down to us. I doubt games would even benefit much from it due to the very different nature of the work they do.

 

[Ferrum Master]

It does matter what it is called - Hypertransport, Quick Path, Mutiol lol... it needs to have specially designed north bridge/uncore part...

I cannot see it living really... supercomputer solutions doesn't count, they are so costly, they usually even implement their own interconnects and cluster arrays...

As I see HyperTransport allows already peripheral connection, so AMD is ready for that since Athlons came... raising parallelism and frequencies ain't a issue to get more bandwidth.

 

[bhaalkc]

I usually skip 2 generations of GPUs so I'll probably go with a GTX 1060 8GB in 3 1/2 years.

I plan to get 2x pascal gpu in sli next year.

 

[newconroer]

I plan to get 2x pascal gpu in sli next year.

That is my plan as well.

 

[fullinfusion]

Amd made this and all Nvidia did was copy this!

 

[RejZoR]

Well, technically, AMD could patent it and I'm sure parts of HBM are patented. They just decided to go with a non restrictive model. Otherwise they could force NVIDIA to stay with GDDR5...

 

[the54thvoid]

Amd made this and all Nvidia did was copy this!

Did AMD actually invent HBM? Or are they not just the first to utilise it?

 

[fullinfusion]

Did AMD actually invent HBM? Or are they not just the first to utilise it?

Simple answer... The 1st to implement HBM was AMD.

Nvidia will take this tec and Re-engineer the HBM shit and stomp AMD as usual for this go Around But----

But yes I say, as being a red Team fan.. umm that the Green team will prevail as the usual k