Scientists Are Stacking Processor Cores on Top of Each Other

[Cazalan]

Ever major FAB is working on TSV technology. What's new about their approach?

 

[annymmo]

This would be great for memory.
Less space needed.
And they can only transfer as fast as the bus thus very predictable heat production.

 

[offerings12]

Intel probably already has these and is waiting for AMD to say they are about to release this NEW technology

 

[panders4]

They have already started doing this with several FPGAs, stacking the transistors. The idea being that the power required for communication between parts of the fabric can be drastically cut. It might actually represent a massive drop in latency between core communication and even a drop in power usage.

 

[hannibal]

[citation][nom]Dangi[/nom]this is the natural upgrade path to improve processors. [/citation]

Indeed! The Ivy Bridge did bring us 3d transistors, so 3d cores seems to next logical step! The heat problem will be there, but if you don't overclock it should not be too big problem... I don't think that Ivys overclokking difficulties are related to this, at least not very much, but it may be one reason behind the different behavior compared to Sandy. Ofcource there are other reasons, but it is not relevant in this case.
I would not be surprised id Hasvell will be 3d core chip, at least partially... If not then the next one most propably will be, if there will not be any major problems with this method.

 

[razor512]

Stacking is nothing new, even current gen CPU's do it for certain components. Doing it for the cores is new and has not really been considered for modern CPU's because they generate too much heat and you can't really create micro heat pipe for each transistor .

If someone can create a CPU they performs like current ones but is so efficient that it doesn't need any additional cooling, then the stacked design can work and may improve performance and lower production cost.

 

[shin0bi272]

[citation][nom]spp85[/nom]Stacking Processor Cores is a good idea, but how they cool the lowest cores ? Heat dissipation would be the biggest problem...[/citation]
you remember the model of the cpu in terminator movies? Cubes connected by little posts... that's what youd do here. A central post between the two cores for communications and your heat sink would have a recess hole cut in it so it would surround the entire thing. You could even cut the recess deep enough to allow the air from the fan to blow between the two cores.

Then of course expanding on that you could do 6 cores in a cube separated by a little space and connected in the middle by that same central post idea so they can all communicate and all but the bottom one have contact with the heatsink.... and I just revealed my idea from 2002... ahh well Its not like i was going to be able to patent it anyway...

 

[wiyosaya]

[citation][nom]elbert[/nom]The 3D design need more of a visual architectural design. Something like a processor sized vent with ribs. Copper heat pipes could be added to stiffen the design to move heat to the primary heat sink customers place on top. Air could also be blown though this design for cooling. Imagine if the processor were like old style heat sinks with one flat layer and several vertical. ...snip... A six layer stacked CPU could not possible vent enough heat to compete in the consumer space.[/citation]
At this time, a six layer stacked CPU could not vent enough heat to compete in the consumer space, however, copper is not the most efficient way to move heat at this scale. There is considerable research being done on heat dissipation in integrated circuits, and one of the most effective conductors of heat is graphene. Graphene's thermal conductivity is about a factor of about 10 times the thermal conductivity of copper. Here is a link to an article for anyone interested.

I suspect that the issue of heat dissipation is one of the reasons the researcher decided not to seek to immediately release this for commercialization. The heat dissipation issue is one of applied technology. The world already has materials that are capable, it is a matter of finding a way to apply the materials to ICs in a way that effectively dissipates heat. From that viewpoint, it is not an unsolvable problem.

 

[danwat1234]

[citation][nom]americanbrian[/nom]My only concern is that the heating effects will be increasingly difficult to manage. Clocking down the chips may be the only way to get them to not overheat, which will still be ok I guess but you lose single threaded speed for presumably higher core counts. We already see that this approach has been a mixed success on our desktops. It is kind of exciting though. It is like the Terminator 2 brain chip.[/citation]

On the 'Sarah Connor Chronicles' I believe I saw John shove a T2 brain randomly into a PCI slot and everything was honkey donkey until the brain tried to use his smart-phone.

 

[requiemsallure]

[citation][nom]joytech22[/nom]Well cooling would be a HUGE problem wouldn't it? Unless they use the outer edge of the cores as the thermal dissipation thing, and dissipate heat through a huge plate covering all four edges, which further extends to a larger heatsink of some sort?Or did I just solve the problem for them..?I don't know.. eh[/citation]

this might not be what they are doing, but when i read the article it reminded me of the PII days when they slotted the processor on a daughterboard and plugged that into the motherboard... on some variants... it was funny this was the first thing i thought of, instead of having it slotted into the mainboard, they integrate it into a daughter board and then slot a dual sided heatsink around it. as rediculous as it sounds it "could" (not saying it should) work that way, but it seems a bit farfetched to me...

 

[alidan]

[citation][nom]americanbrian[/nom]My only concern is that the heating effects will be increasingly difficult to manage. Clocking down the chips may be the only way to get them to not overheat, which will still be ok I guess but you lose single threaded speed for presumably higher core counts. We already see that this approach has been a mixed success on our desktops. It is kind of exciting though. It is like the Terminator 2 brain chip.[/citation]

single thread exclusive programs will most likely be dead by the time this comes, or we will have moved to something like graphene, which can handle high clock rates at low temps.

but with a full 3d core, you could cut the travel time for data by a crapton, meaning that even if the clock is slower, the data moves more efficiently, leading to an over all increase in speed, with a shrink of the chip size, or if the process to make the chips 3d is fast, you could see a 16 core chip with 4 cores with 4 stacks, and than adding hyper threading making it a 32 thread chip, as cheap as our current quad cores.

 

[teddymines]

It might be the case that it will come in a sealed package like requiemsallure mentioned, where all cooling support is internal, and you just fit it into a slot and give it power. Maybe put a couple such slots on a motherboard to support multiple packages.

 

[husker]

To everyone concerned about heat dispensation: Using the third dimension would help to lower total heat generated since circuitry would be shorter, resulting in less electrical resistance and heat. I'm not saying it would totally compensate, but it is a factor in its favor.

 

[eddieroolz]

I thought they meant literally stacking processor packages on each other...

 

[sceen311]

So one of the most powerful processors in existence, is 3D, one should only assumed artificial ones would follow suit.

 

[americanbrian]

The central chip is the one that will be most affected by the heat imbalance, we all agree.

Saying that the circuit paths are shorter is not necessarily true. The centre chip will need all its pin-out to go through the stack, which is a longer path then directly through the PCB on its own.

Also if we take all the cores pin outs out on 1 side we are talking maybe thousands of thru's through the pin outs. And this exacerbates the distance problem for the "far" core. This establishes another aspect of diminishing returns on core counts.

And you want to add cooling thru's too? Doubt it.

I like the idea of a fan or toroidal plane layout. However I don't think we can have it all. The distances of interconnects might have to increase, but may still be achievable.

I am imagining a classical "sun" shape with a central shared cache making the "circle" and each processor viewed edge on as the "rays". This would give some space to fit heat dissaption "technique of choice" to access both sides of all the chips and distances kept minimal. You can then take pin outs from the edge of the rays outward and place the whole thing in a cylindrical socket.

Somebody owes me something for that. I think it could actually work.

 

[Kurdain1]

They were talking about this an Intel 10 years ago as well as etching out all raw silicon out from the finished CPU so you are left with a lattice work of gates and interconnects allowing air to flow within the CPU. Extremely fragile but it'd be good for thermal issues.

 

[ProDigit10]

There's also a significant heat increase in these processors.
They'll literally have to build the processor inside the fan, or the fan around the processor, to maximize heat transfer.

 

[Kurdain1]

I don't think heat would be such an issue. You'd likely see an increase in via size, in some locations on the die at least, to better channel heat. There would likely be a redesign where the heat sink pulled heat from both sides of the die rather than just mainly the top like it is now.
Also I wouldn't be surprised if you'd see another drop in processor speed with a large increase in cores, just like the move from the single core speed race to todays multicore cpu's.

Another thing is that the signals would have far less distance to travel, which is a significant issue when you are trying to sync up GHz clock steps across a large die. Less circuitry to deal with this and less 'wire' for power to be lost on. On such a small scale every tiny mW of power adds up quickly.