Nvidia Pascal Speculation Thread

So far as I know 4 orders of magnitude is 10^(4) rather than 10*4.

 

If anyone is interested, by considering an order of magnitude is +10dB and thus half is +5dB, the factor of half an order of magnitude is ~3.16228, which is just sqrt(10).

The day I ran Crysis, I sure wished it had been 3.16x faster

 

1/2 is not 5dB. 3dB is double.

1/2 is 7dB.

10 dBm 10 mW
7 dBm 5.0 mW
6 dBm 4.0 mW
5 dBm 3.2 mW
4 dBm 2.5 mW
3 dBm 2.0 mW
2 dBm 1.6 mW
1 dBm 1.3 mW
0 dBm 1.0 mW

 

Just like 2.5D in the past it's a poor substitute for the real thing

 

If Crystalwell is any indication (not to mention various, less tangible PowerPoint slides) it's still pretty damn great.

 

The memory itself is stacked - it's just in a stack next to the GPU instead of on top of it.

 

So Volta will have the memory stacked on top of the GPU while Pascal will have the memory next to it? I've got a feeling tht Pascal and Volta won't change that much from Maxwell architecturally and the next 'big arch change' will happen presumably with Einstein (whatever the hell the next GPU arch after Volta).

 

That's over an order of magnitude difference in memory bus distance and probably capacitance right there.

It won't be easy to do the real thing (DRAM under logic might work, if not I guess only integrated heatpipes or microchannel water cooling can make it work) but there are massive gains to be had.

 

If a SoC is in the <10W range, can the exotic cooling be skipped?

 

We don't know anything about Volta other than that it's post-Pascal. But don't expect anyone to stack memory on top of the GPU. 2.5D stacking will be the preferred method for high performance GPUs.

 

The PS Vita has a stacked package, and one layer of Wide IO memory on the top of the processor.
The wiki on the Vita TV says the whole box has a max power of 2.8W.

The components of the package target low power, which is likely necessary not only for the Vita's mobile origins but to keep individual power contributions small enough for the limited heat transfer out of the stack.

 

NVIDIA has stated "In Volta, due after Maxwell, memory modules will be piled atop one another and placed on the same silicon substrate as the GPU core itself."

I think a lot of people think "substrate" is the "the GPU die", but it's not. As Ryan says, NVIDIA is using the 2.5D method where the GPU die and the 3D stacked RAM die are mounted to a common substrate. Here's a good article on the various 2D, 2.5D, and 3D strategies.

 

Just so no one is confused, that article is from 2013.

http://www.anandtech.com/show/7900/nvidia-updates-gpu-roadmap-unveils-pascal-architecture-for-2016

Volta was pushed back this year. Pascal is what we'll get in 2016.

 

As it stands now, Nvidia's roadmap should look like this:

Fermi --> Kepler --> Maxwell --> Pascal --> Volta --> Einstein (?)

 

That seems the case, the only doubt is Einstein which the only mention was someone from Nvidia while back that it would be after the release of Maxwell (or was it their 2020 arch and part of their Echelon target?).

 

Was there anything ever disclosed about Volta other than the stacked DRAM? If not, with stacked DRAM now part of Pascal (not to mention, the introduction of Pascal itself), it's clear that there were some changes in road map.

At that point, does it even matter what comes after Pascal and what name it has? If the only information of substance is a name, there's really not much to talk about, is there? We might as well call it Heisenberg.

 

Ultra flyfucking mode on..
"Two orders of magnitude" means 100x, and not 20x. That's why I think "half an order" is 3.16x, not 5x.

 

Exactly.

 

Can someone explain all those parts shown in the above picture.

If the four dice surrounding the Nvidia GPU are the memory modules then what are the eighteen large dice on the PCB?

Those eighteen large dice also have eighteen small dice next to them and what also looks like a capacitor/inductor. Are these power regulators?

 

They look like power related gizmos to me.