It’s no secret that graphics cards, especially in the server and data center sector, are becoming ever more power-hungry. With Feynman, we’re talking about fairly insane power consumption, while Blackwell is already hitting 1300W! In short, these energy constraints are prompting NVIDIA to adopt a process that will then be in its infancy: TSMC’s 16A. It’s a risky gamble, since the younger the process, the lower the yields, and when large chips are etched, this can lead to significant losses.
Feynman etched in 16A, a risky bet?
Because of significant energy constraints, NVIDIA has opted for TSMC’s 16A process. From now on, this will be a rather young process, which may have a negative impact on performance. Let’s not forget the problems with Intel’s 18A process. Well, it’s the same thing, but applied to large GPU dies.
On the other hand, with the 16A, NVIDIA can take advantage of a so-called backside power supply. Basically, power is routed through the back of the silicon, separating the signal and supply lines. As a result, we benefit from improved signal integrity and better control of the power component. The process also features Super Power Rails to reduce resistance by feeding “the source and drain terminals of the transistors”.
In contrast, with Rubin, the Greens have opted for an N3P process that will have had time to mature. However, the chip will benefit from a CoWoS-L-based package. Here, we’re expecting much higher power consumption than Blackwell, with over 2300W on the clock. One thing’s for sure: if you’re going to open a data center, given the price of energy, you’d better not set it up in France… You’ll soon be out of business.
