AMD's Future Ryzen SoCs May Feature New Chip-Stacking Technology

[lexluthermiester]

This only took what, 19 years after conception? Congrats AMD for being first in the X86/X64 market, but also a big damned BOO to both AMD and Intel for being so damn late to the stacked IC party.

 

[Dr. Dro]

Wasn't that a simple POP for the simple purpose of having a smaller package? It didn't do much other than keep the size of the thing really really small

I believe that is kind of the point, small package > lower physical distances > massively improved latencies, that comes with its own drawbacks though. Interesting read, perhaps from a what if scenario if 20A hadn't been shelved

https://hc34.hotchips.org/assets/program/conference/day2/Mobile%20and%20Edge/Meteor_Lake_Hotchips%20_%20Wilfred%20_%20final_submit%20(1).pdf

 

[ymdhis]

Why would that change something significantly in terms of Infinity Fabric? They'll run IF over silicon an TSVs or something but it will still be IF. The latency penalty has quite a bit to do with distance and less about going over substrate. Power is a different thing though.

Going through TSVs can make IF use less power and have improved latency. The gains might be minimal for desktop chips but could be more substantial for Threadripper/EPYC.

 

[londiste]

Going through TSVs can make IF use less power and have improved latency. The gains might be minimal for desktop chips but could be more substantial for Threadripper/EPYC.

Less power - absolutely. But would going through TSVs improve latency by itself (other than the distance reduction)?

 

[Nhonho]

There is a simple and old technology that inverts the hot side and the cold side of a piece of metal. I am searching for it on Google here, but I can't find it. I remember that they made some coolers for AMD's old Athlon and Duron socket A with this technology.

Does anyone know what this technique is? Post the link here if you find it please.

AMD could use it to transfer heat from the bottom side of the dies to the coolers' heatsink.

 

[TumbleGeorge]

There is a simple and old technology that inverts the hot side and the cold side of a piece of metal. I am searching for it on Google here, but I can't find it. I remember that they made some coolers for AMD's old Athlon and Duron socket A with this technology.

Does anyone know what this technique is? Post the link here if you find it please.

AMD could use it to transfer heat from the bottom side of the dies to the coolers' heatsink.

Thermoelectric effect, Peltier(other scientists also has works on this target) element.

 

[mechtech]

Stacked Chips......................old tech

 

[ymdhis]

Less power - absolutely. But would going through TSVs improve latency by itself (other than the distance reduction)?

The distance reduction alone should improve latency considerably. It is one of the reasons why core to core latency in Ryzen is measurably higher between cores in different chiplets.

And even if it doesn't improve latency directly, it can do it indirectly: less distance means less power, less power means higher clocks, and higher clocks can also reduce latency.

 

[londiste]

The distance reduction alone should improve latency considerably. It is one of the reasons why core to core latency in Ryzen is measurably higher between cores in different chiplets.
And even if it doesn't improve latency directly, it can do it indirectly: less distance means less power, less power means higher clocks, and higher clocks can also reduce latency.

AM5 CPU PCB size is 40x40mm. Lets assume the signal needs to go across the entire distance of package twice and a bit more - let say 100mm. Should be not really realistic and beyond a worst case scenario. For a signal to travel 100mm takes roughly 0.35nm. I doubt that distance reduction improves latency considerably.

Would IF be able to sustain higher clocks from either moving to silicon or from power reduction? I don't truly know - I think the GPU/HPC implementations have gone primarily wider for more bandwidth.

 

[mrnagant]

It seems like the IO die would need to be huge for chips that have 8+ chiplets on it. IO die could use cache as filler to make it large enough, rather than having empty space.

Gonna be interesting to see what AMD does.

 

[Nhonho]

Thermoelectric effect, Peltier(other scientists also has works on this target) element.

AMD and Intel should develop a standard to put this cooling technique into practice, to be common and controlled by the motherboard.
The cooler could be connected to the motherboard via a USB cable, for example.

 

[TumbleGeorge]

AMD and Intel should develop a standard to put this cooling technique into practice, to be common and controlled by the motherboard.
The cooler could be connected to the motherboard via a USB cable, for example.

If it was viable, it would have already been implemented in such a way.

 

[Nhonho]

If it was viable, it would have already been implemented in such a way.

I think it would be viable for home CPUs that get very hot and for server CPUs that have many cores and therefore also get very hot.

 

[londiste]

AMD and Intel should develop a standard to put this cooling technique into practice, to be common and controlled by the motherboard.
The cooler could be connected to the motherboard via a USB cable, for example.

I think it would be viable for home CPUs that get very hot and for server CPUs that have many cores and therefore also get very hot.

You have a very naive notion of how a Peltier element works.

Enthusiasts have tried cooling CPUs with Peltiers since time immemorial, CPU cooler manufacturers have given it a go and from ones you mentioned at least Intel has dabbled into doing something with it. It has not broken into mainstream and there are good reasons for it. There are application to Peltiers and that includes chip cooling at times but these are very specific very niche use cases.
- First, it consumes considerable amount of power to function in the first place.
- Second, also related to the first one - while its cold side does get cold the hot side also gets much more hot. Not only hot because of the heat it moves from cold side but because of the wattage it pulls to function.
- Third, there is a limit to the amount of heat a specific unit is able to remove. IIRC a CPU sized piece (say, 50x50mm) usually can handle like 100W or so.

From practical aspects, you still need a way to cool off that hot side. And on the cold side condensation tends to be a thing which electronics dislikes.
Btw, TEC is usually the keyword for Peltier in CPU coolers if you want to search and read up on it.