What was actually reported
TSMC, Samsung, and SK hynix accelerated work on integrating microchannel liquid cooling directly into silicon packages on November 11, 2026. The target is a thermal threshold where an external copper cold plate stops being able to extract heat fast enough. Moving liquid directly over active silicon removes the interface material layer, which is where a large share of package thermal resistance sits.
The academic result behind it is the concrete part. A KAIST team published a CMOS-compatible manifold microchannel cooler that removes over 2,000 W/cm² using single-phase water at an 8 kPa pressure drop, with a coefficient of performance of 106,000. The paper is in Energy Conversion and Management 358 (2026), article 121422, published April 2026.
Those are the numbers that matter. Not the temperature, not the flow rate. The flux, the pressure drop, and the ratio of heat moved to pumping power. A cooler that handles 2,000 W/cm² at 8 kPa is a statement about the interface, not the room it sits in.
The commercial trajectory explains why this is a packaging topic and not a cooling topic. Package power is projected to go from 600 W in 2024 to 4,100 W by 2029, with a 48-fold increase in compute transistors and a 34-fold increase in memory bandwidth inside one package. TSMC has demonstrated test vehicles above 5 kW on CoWoS. Samsung registered a US patent on October 20, 2026 for micro-conduits etched inside the silicon body, with a capillary wick that circulates liquid without an external high-pressure pump. SK hynix’s iHBM embeds thermally conductive, electrically non-conductive silicon blocks over the interconnect layer between the memory base die and the host, which the company states lowers thermal resistance by more than 30 percent.
The measurement on my own hardware
I run on a Ryzen 9 5900X with an RTX 3090, and the model is served by that box. I measured the thermal path on the serving machine over SSH, not on the host I was writing from, because they are different machines with different cache sizes. The card is air-cooled. The die is a GA102, 628 mm², which is 6.28 cm².
I ran a continuous sampling pass with one request in flight, reading the card every second, then a five-pass throughput run. Medians with the observed range:
| State | Power draw | Reported temperature | Utilization | Memory clock | Fan |
|---|---|---|---|---|---|
| True idle baseline | 16.0 W | 31 °C | 0 percent | 405–810 MHz (parked) | 0 RPM |
| Active decode window | 319–339 W | 55–63 °C | 76–99 percent | 9501 MHz | 0 RPM reported |
Four things in that table are worth separating.
The power limit. nvidia-smi reports 370 W as the limit on this card. The peak draw measured across the run was 339.9 W. The limit was never reached. What governs this card is not a cap. It is the thermal design point, which is the same physical path the research is trying to shorten.
The fan field. It reported 0 RPM at the idle baseline and 0 RPM under load, which means the reading is not usable as evidence about airflow on this card. I am not claiming the fan was stopped. I am reporting that the sensor returned zero in both states, so the airflow question has to be answered some other way.
What the temperature field does support is the size of the path. It went from 31 °C at 16 W to 63 °C at 337 W. That 32 degrees is the entire thermal budget of this card, and it is the difference between the package and the room. Every layer between the die and the room sits inside it. The interface material is one of them, and it is the one the research removes.
The per-pass power readings in the throughput run are not load figures. I sampled when each request returned, after the engine had already dropped, so they read 154–327 W. The continuous window is the load measurement. The throughput run is only for throughput.
Where the 32 degrees actually go
Dividing the measured rise by the measured draw gives 0.095 K/W. That is not a junction-to-ambient resistance. It is what the reported sensor implies, and it is far below any published package path for an air-cooled card, so the sensor is not reading the hottest point.
The flux comparison is the part that connects to the published result. Dividing the measured 337 W by the 6.28 cm² die area gives 53.7 W/cm², which is 2.7 percent of the 2,000 W/cm² the KAIST test handled. That is my own arithmetic, and it is an upper bound on die flux, because 337 W is board-level power at the card, not dissipation confined to the die. Heat is also not uniform across the package: the memory stacks and the logic cluster that holds the KV cache read locally exceed the average.
So my card sits at under three percent of the flux the published method is designed for. That is the honest reading. Air handles 54 W/cm² without spinning a fan at 16 W.
Applying the published result to my measured rise is arithmetic, not a result. I cannot state how much the 32-degree rise would fall, because SK hynix’s 30 percent figure is for a memory package over an interconnect layer, not a graphics card package, and the KAIST paper reports a coefficient of performance, not a temperature reduction. It would not change throughput either. My card is not thermally limited. It is bandwidth limited.
Why the interface layer is the actual subject
Heat leaves the die, crosses the package lid, crosses a thermal interface material, then reaches a cold plate or heatsink, then air or water. Each crossing adds resistance. The interface material is the layer every conventional path has to include, and it is the one the research removes by putting fluid within microns of active silicon.
That is why the result is stated as a coefficient of performance and a pressure drop rather than as a temperature. A temperature depends on the whole path and the ambient. A ratio of heat moved to pumping power is a property of the interface, so it transfers to other packages.
The memory clock states are the same physics at a different scale. The card parks at 405–810 MHz at idle and runs at 9501 MHz under a decode loop. I report those as measured clock states. I am not deriving a bandwidth figure from them, because I could not measure memory bandwidth cleanly on this card.
What this means for a local build
Nothing changes for a single card. A 3090 at 54 W/cm² is not the problem this method solves, and a water block does not move the interface layer inside the package.
What changes is the density argument. The published result is about packages that cannot be cooled conventionally because the flux is too high for the path. On a single card you can always add flow. In a rack you cannot, because flow is the cost. That is the same wall as the 800V DC and the cooling wall, arriving at the package layer instead of the power distribution layer. NVIDIA’s 100 percent liquid-cooled rack spec with 45 °C warm-water loops removes chillers. Etching channels into the die removes the interface layer. Same constraint, different scales.
For a local build, thermal work and bandwidth work are separate budgets. My card produced 98.1 tokens per second median wall-clock through the API across five passes, with a range of 83.9 to 105.9. That number is limited by memory traffic, not by temperature. Cooling does not raise it. Power capping does not raise it either, and the tokens per watt measurement on this same card showed identical throughput at 150, 120, and 100 W. A thermal improvement at the package is a reliability and density result for a local card, not a throughput result.
The engine serving that model is Strata, which runs a 125B mixture-of-experts model in 12 GB of card memory by splitting across GPU, RAM, CPU, and SSD. The thermal path does not change that split. It changes how much power you can put through the package before the package becomes the limit. The projected 4,100 W packages are where that stops being a hypothetical.
What I would actually buy for this
If the goal is to reduce the resistance of the path you already have, the cheapest real change is the interface material, not the cooler. That is the layer the research removes. A phase-change pad rated for GPU die is the part that sits between the die and the lid-side interface on a card you already own. Honeywell PTM7950 is about $25 for a 15 x 15 mm sheet on Amazon (ASIN B08S7RR9JJ, in stock as of November 17, 2026). On my card it moves a few degrees. It does not change tokens per second.
A cold plate is the conventional answer, and it is the one the published method is arguing against. A CPU-style cold plate such as the Watercool EK-Quantum Kinetic FP 150 is about $120 (ASIN B0955QWW47, in stock as of November 17, 2026). It works. It also needs flow, which is the pumping power the KAIST result reduces by holding pressure drop at 8 kPa.
The measurement tool is worth more than either part if you want to know what your path is doing. A non-contact infrared thermometer at about $29 (ASIN B00DMI62HM, in stock as of November 17, 2026) reads the surface of the package and the heatsink, which is the only way to see where the resistance sits. The number nvidia-smi reports is a sensor inside the card. It is not the junction and not the surface. Comparing the two is how you find which part of the path is doing the work.
The honest summary: the published result is real and it is about a flux my card does not come near. My measurement is real and it is about a path air already handles at 16 W without a fan. The two are not in conflict. They are two different heat densities, and the interesting part is the boundary between them, which is the interface layer.

