At this week’s 2025 Intel Foundry event, there were a number of non-process-related capabilities that the company showed. One was the Intel Foundry Thermal capabilities. That included solder TIM and other more traditional approaches we have seen as well as options like liquid-cooling being done in package.
Intel Foundry Thermal Capabilities with TIM Options and In-Package Liquid Cooling Shown
Intel showed off its capabilities, characterizing hot spots on chips and the ability to design thermal solutions around that.
To those outside of the industry, it may not make sense at first, but looking at future designs where there are different types of tiles likely on different process nodes and performing different functions, it makes sense that some areas will get warmer than others, especially since each tile can interact with those around it.
Currently, Intel Foundry has a number of thermal interface materials. The “New TIM” sounded like it was liquid metal. Some TIM types are better for different applications so it makes sense that there are different options out there.
One of the more interesting showcases was adding liquid cooling into the processor packages directly. Here we can see this demoed with cross sections.
Intle had an exampel package here with openings for cooler and warmer fluid.
Here is another angle of that.
Intel also had other packages and their test fixtures on display.
One challenge is getting the cooler with its channels into an existing packaging footprint.
Here we can see another design for a larger chip. If you have used coldplates in today’s liquid-cooled server designs, you will notice how much smaller these are.
Intel says that it is more efficient to package the cooling block on the processor because there are fewer layers to go through versus a traditional chip where there can be TIM between the lid and the chips, then more TIM between the lid and the coldplate.
Final Words
What was shown at Intel Foundry 2025 was more of demonstration gear. We also heard a bit about what Intel is testing for future chips. Today’s designs are neat, but as we move to 2kW and 3kW accelerators even the internal structures will need to change. This is an area many will not think about when they hear about a chip foundry, but perhaps it makes sense since it is one less set of obstacles a chip design team needs to solve before having a working solution if the foundry can provide thermal solution design.
Of course, this was also focused more on top of package cooling solutions. We have heard some companies talk about a future where liquid may have to go into the actual chip package itself, not just on the top surface which would be another reason folks will need a team of thermal engineers designing in-package liquid-cooling blocks in the future.
We also had a weekend piece on a new cold plate product we were shown in our Substack here:
Looking at these photos reminds me of a video by CPU Galaxy that featured a processor that came out of a decommissioned Cray supercomputer. In that processor, the thermal hat was designed to allow refrigerant to flow directly on and over the exposed dies. There was no cold plate in that design as such, just an opening for the refrigerant inlet and another for the outlet and the fluid filled the internal space between the ceramic carrier and the thermal hat. The inlet opening was shaped in such a way that the refrigerant fluid sprayed out over the dies, covering them in a uniform layer of fluid that aided in thermal dissipation as it moved over the dies and out the other hole in the thermal hat. These Intel test chips appear to work somewhat similarly to that one, though with an enclosed assembly that keeps the coolant separate from the dies in the package. Very neat stuff.
I wonder how they deal with dirt and contaminants which can occur. There are possibly such small structures, that it is nearly impossible to clean without destroying the chip once clogged up.
Filters which in return need a higher pressure and tight replacement intervals?
There was a time where the CPU didn’t even had a heatspreader, what changed this?
@DarkServant
The reason for processors over time requiring heat spreaders and then active liquid cooling is primarily due to the increases in TDP caused by the logic clock speeds increasing. The faster a chip is made to run, the more power it consumes, and in turn the hotter it runs. There’s only so much that reducing the process size can accomplish as far as thermals are concerned. Current chips are running as fast as 5 GHz and that takes a lot of power to do even for 3 nm process technology.
The reason old chips like the 286, 386 and 486 didn’t need the complicated heat dissipation methods we see nowadays is because those didn’t run much beyond 20 to 50 MHz on average. The 486DX4 could get up to 120 MHz, but that only really needed just the heatsink. When it comes the the earlier 486DX-33 and older, they just weren’t running fast enough or getting hot enough to require anything more than a simple heatsink and 286es and 8086es didn’t even need that because those maxed out at 10 or 12 MHz.
As for the issue of crap clogging up the tiny channels in the cold plate heat transfer block, I would imagine there’d have to be some kind of filtration system to mitigate that. It wouldn’t make sense not to unless the heat spreader can be made to come apart easily without destroying the package. I think something along the lines of IBM’s old thermal conduction modules (TCMs) would have to be implemented for this since those were able to be fully disassembled to service them. But these are all speculations on my part here.
Another reason for liquid cooling nowadays is the sheer amount of logic circuitry on a die. Back when the 286, 386 and 486 were popular, the most circuitry a chip could handle was maybe a couple million (~2 to 8-ish million) transistors with the manufacturing process sizes of that time. Today, chips have multiple cores, larger more complicated cores and they’re running at 5+ GHz. The typical processor core now has over 1 BILLION transistors, excluding caches or any inter-core or inter-die fabric logic. Include those and you have huge processors with easily a trillion or more transistors. That takes lots of power, too, increasing TDP substantially and thus heat output, hence the need for the cooling systems and techniques we have now.
What leads you to believe that the ‘new TIM’ is a liquid metal?
@Stephen Beets
Thanks for your in depth information. I see the increase in transistors is enormous, if i look at the first Geforce 256 GPU with 17 Million transistors and in contrast the Blackwell GB202 which has 92,2 Billion transistors (remembering playing Quake 3… and of course Unreal with 2x Voodoo 2 with a gigantic 1024×768 resolution on CRT before Y2K)