AMD Talks 3D Stacking at Rice Presentation

AlphaAtlas

[H]ard|Gawd
Staff member
Joined
Mar 3, 2018
Messages
1,713
AMD has talked up their "chiplet" based approach used in their upcoming products, and according to some reports, Marvell is already selling products based on the chiplet concept. But the next logical step from that approach is to move from 2D to 3D, where different dies are "stacked" on top of each other. In a recent presentation at Rice University, AMD confirmed that they're working on 3D stacking techniques in their future designs, and that it's a necessary step to keep the improvements coming, but didn't elaborate much beyond that.

Check out the memory and stacking talk in the presentation here.
 
Would be interesting to see the new heat transfer mechanisms for a 3d stacked chip. not much different than the human brain with a round head around it for cooling.
 
Would be interesting to see the new heat transfer mechanisms for a 3d stacked chip. not much different than the human brain with a round head around it for cooling.
There is a lot of talk about exotic cooling methods involving pumping various fluids through 3d semicondcutor structures, but I think the most obvious method would probably be to simply make sure each die has some good thermal contact with the die above and simply adjust clocks and voltages until the performance is reasonable. You can reduce a lot of power output by adjusting clock speed and voltage down just a little.

There was some research paper from Intel a while ago about a 3d stacked pentium 4 they custom built. They managed heat by not having hot blocks above eachother and adjusting clocks accordingly. Despite having lower clocks and the chip using less power, the overall performance was actually higher than a regular pentium 4 because of reduced latencies.
 
If they can make a connector that is on the edge long wise for the chiplet, stand it up on it's side, layer them with an offset it would increase surface area (and volume) while also increasing density.
 
Zen4 might be a stacked arch I guess? 2022 and on 5nm maybe?

from what I've read 7nm is the floor in terms of fab node size. Any smaller and the transistors cease to function properly. They lose their state and become unstable. Which is why AMD is moving to chiplets.
 
How many MHz do you think our brains run at? o_O
Commonly accepted values based upon brain wave states suggests about 4hz to 45hz. Not megahertz but hertz. Like a processor though during a single hertz many different tasks seem to be able to be run at once. So perhaps that's a billion things happening 45 times a second when you count reading in from every sensory input device (including every skin cells temperature or heat or dryness, etc)
 
Commonly accepted values based upon brain wave states suggests about 4hz to 45hz. Not megahertz but hertz. Like a processor though during a single hertz many different tasks seem to be able to be run at once. So perhaps that's a billion things happening 45 times a second when you count reading in from every sensory input device (including every skin cells temperature or heat or dryness, etc)
How would one go about overclocking such a device? Asking for a friend...
 
How would one go about overclocking such a device? Asking for a friend...
If your serious about this topic, brainwave entrainment audio files for a specific hertz along with visual stimulation at the same rate.. Additionally one might use brain neurological feedback and interference as a means of removing blockage and improving the syncing of hemispheres. A 16 lead system with audio feedback usually runs 15k but you can buy sessions for 200 to 300 each.
 
from what I've read 7nm is the floor in terms of fab node size. Any smaller and the transistors cease to function properly. They lose their state and become unstable. Which is why AMD is moving to chiplets.


AMD just announced that SI is going to be around for about ~10 more years and will be roughly 3nm. That is the wall.
 
How would one go about overclocking such a device? Asking for a friend...
Unless you're planning on improving the heat dissipation capacities of your brain somehow, how do you expect to accomplish such a feat without cooking off your neurons?
 
Back
Top