TechTuber TrashBench recently posted a video where he performed one of the most riskiest watercooler experiments to date: pumping water directly on to the die of a GPU.
TrashBench has one question: Why do we need metal at all? Why can’t we just pump water directly onto the bare silicon? The stated goal sounds simple enough, as least in practice: create a custom 3D-Printed GPU “waterblock”, sans the metal block that usually sits on top of the silicon, and make sure that the pump doesn’t leak.
Definitely Don’t Want To Spring A Leak Here
But simple sounding is obviously not as easy as it sounds. TrashBench’s first objective was to create the leak-proof waterblock. For the first trial and error phase, he used an already dead NVIDIA RTX 3060, making sure that he had a solid enough prototype that wouldn’t cause an electrical fire mid-testing.
The good news, for TrashBench at least, is that they did create a workable direct-to-die model, which allowed them to move on to a second working card which, in this case, was a GTX 980. This was still the trial and error stage; the clear plastic they initially used wasn’t suitable and began leaking everywhere the minute water began pumping through it, permeating through the plastic of the bracket.
A couple of tweaks to the 3D Printer settings, though, some reapplied nail polish, and epoxy to attach the cooler to the GPU, and they got it running. With that, they moved on to the next part.
Drops In Temperatures
To see if their custom direct-to-die watercooler would work and how it would perform against a conventional AIO cooler, TrashBench switched over to a more powerful RTX 2060. A few things to note here: For a baseline, they compared the performance of the custom loop with an AIO cooler from Arctic, attaching it to the GPU die with the help of a clamp. Yes, a clamp. Second, they tested their custom waterblock with the water in two states: room temperature, followed by water chilled by a -28Β°C freezer. Finally, for control, they ran two programs, Unigine Heaven and Cyberpunk 2077.
The difference in performance is, as you’d imagine, like the sun and moon. With the Arctic AIO cooler, the RTX 2060 averaged around 36Β°C on Heaven, and 33Β°C on Cyberpunk 2077. With their direct-to-die custom watercooler, that temperature dropped to 28Β°C across both programs. That’s actually a significant decrease in peak operating temperature, and proof that the exercise does work, assuming it is applied well.
And what of the temperatures of it running on -28Β°C water, you ask? TrashBench showed that the card ran at 13Β°C on Heaven, but didn’t show its performance in gaming. The trade-off here, as you can see from the video, was the condensation around the pipes and watercooler, which is to be expected.
TrashBench also took the opportunity to test their watercooler with a CPU and in this case, an Intel Core i5-7600K that he was using as his testbed, while also using the Arctic AIO as a baseline. The results were…less than ideal. Compared to the GPU, their custom watercooler performed worse than the AIO: it reached temperatures of 52Β°C while running on all cores, while the Arctic AIO cooler ran at 39Β°C .
(Source: TrashBench via YouTube)