NVIDIA GOES 12nm
Much of Turing’s advancements are thanks to a die shrink, part of the never-ending quest to be smaller and faster. Turing GPUs are manufactured using TSMC’s 12nm FinNET process. In reality, it is more of a refinement of the previous 16nm process than a meaningful reduction of feature sizes. Witness the fact that the TU104 is only slightly smaller than the GP102 used in the GTX 1080 Ti, and the TU106 has 10.8 billion transistors, against a GP102’s 12 billion. Despite moving to 12nm, they really aren’t much smaller at all. However, it is a mature process, with excellent yields, and it should readily lend itself to faster clock speeds.
The change that will help boost things is a move to 7nm. TSMC is close to full production with this, and AMD’s next GPU is supposed to be out and about before the end of the year. If this goes well, we hope to see a more effective shrink of Turing to 7nm. This will get die sizes down to more manageable levels, and enable more cores to be added, and maybe more RT units to really push the ray tracing.