[SOLVED] How are modern cpus and gpus designed?

[chetan85]

Hello all, currently doing a digital circuits course in university. CPUs and GPUs in modern times have billions of transistors. How is something on that scale improved or even iterated on? I understand being able to tweak individual adders or multiplexers to increase efficiency, but how is such a thing possible for people to do on a chip that has billions of them? Do they use supercomputers to simulate all the transistors, or is there some level of similarity which is then scaled up to billions?

 

[InvalidError]

Most non-trivial digital chips these days are designed using some form of hardware description language and then software handles the bulk of lower-level details much like it does for software. Transistor-level details are tucked away in the fab's primitive library and your design tools do their best to express your HDL as a function of those for manufacturing or programming an FPGA.

Performance gains at a more macroscopic level come from studying how the architecture performs when running typical code, identifying the most significant weaknesses and finding ways to fix them.