A hobbyist called TrashBench epoxied a 3D-printed pipe fitting directly onto a bare GPU and CPU, then pumped water straight over the dies with no water block, no copper interface, no conventional hardware at all.
The GPU outperformed a professional water cooling setup by a measurable margin. The CPU did not. The rig also leaked and nearly killed the hardware. The gap between the GPU result and the CPU result is the actual story here, and TrashBench offers a theory in the video that is worth hearing before you form your own opinion. Direct liquid contact cooling is not new, oil immersion has existed for years, but a controlled side-by-side comparison that exposes a die-specific performance split is rare and specific enough to matter.
The experiment is unfinished. More tests are implied. The methodology is brutal and the sample size is one, but the data is real and the question it raises, why does die geometry or thermal interface material absence affect GPU and CPU outcomes so differently, does not have an obvious answer. Read and watch the original.
[READ ORIGINAL →]