Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Mar 19, 2015.
Here's some examples.
HBM does not need cooling, because it runs at lower voltage and considerably lower clock, therefore entire 4-layer brick eats less watts than gddr5 therefore generates less heat.
HBM by itself, yeah you can cool it using traditional methods. But they are moving towards 3D memory configurations. Right now both AMD's 390x and PASCAL (despite nvidia marketing saying 3D) is using a 2.5D configuration. Eventually the HBM will be directly ontop of the GPU die, that's when non-traditional cooling methods will need to be utilized.
3D is not for tomorrow .. it called TOP-PIM Throughput-Oriented Programmable
Processing in Memory
AMD research on this .. ( They are largely leading on this future option, in collaboration with SKHynix, Samsung etc )
http://www.dongpingzhang.com/wordpress/wp-content/uploads/2012/04/TOP-PIM_v6_public.pdf ( actuallly this presentation is more easy to understand and at the same time the more complete you could surely find about it )
Full 3D placement is far future and cooling solutions required are not meant for home PC. It will come 3-5years in advance to servers before we'll get it home, because it will cost a lot.
And main idea is not to place memory on top of GPU/CPU, but to put 2nd GPU/CPU with very same architecture like bottom one and that will save power considerably.
It is too bad that for now we can't cool more than 15-25W (not sure exactly as I only saw slide for moment) CPUs/GPUs per chip for 2 slices.
But if we look at AMD APUs, they are getting close to make this possible.
One Excavator Module (2-cores) eat up to 15 Watts at reasonable clocks. Placing 2 modules above each other may reduce power consumption to 10W per module.
But then there are so many other things and for now it will not fit thermal limits.
If it was possible to do with things AMD has now (power management), 3D stacking would likely get 8-core carrizo to 45~55W, but manufacturing price... I think there still needs to be some evolution.
For servers biggest issue is cooling solution. It must not degrade chips, it has to be serviceable and space is an issue too.
I can imagine it here (servers) in 2020 since by then AMD wants to accomplish their power efficiency target. And that will likely need 3D stacking of processing units.
Yeah, my Digital IC Design professor at RIT actually worked with IBM on/with 3D interposer tech. It was a pretty interesting class.
I don't know what about my post that made you guys think that I'm saying that this is coming tomorrow. I'm aware of how far 3D based integration is. What I'm saying is, most of the advanced cooling systems (between substrates) that are talked about are specifically for that application. Not the way HBM is being integrated currently.
I think you mean MB.....
Not to mention it was a 2003 video card...
Nice display there "vbetts" of pics and diagrams
so this means graphics cards will become smaller cause of the stacking memory