But agreeing with you, I've done big optimization stuff for multicore servers (not as many cores, but same kind of work) and my workstation was something small with not even the same os. I don't need the big machine on my desk to understand the concepts. I just need the big machine to check my work. For me, that's always been a production machine, sometimes a production machine taken out of rotation for pre-validation before running on production load. I guess I should mention, I work on applications specifically, and libraries and kernels as it relates to making whatever my application is work better. I also don't have a problem with pinning threads to cpus... but my applications are usually one big program that fills the system. Someone writing a general purpose library has a harder time.
Of course, if you want to do this kind of work and you don't have your own production load, you're going to have to borrow, rent, or buy a big machine. It doesn't need to be your workstation though. I hate working with cloud nonsense, but if your tests are short, and you do the upfront work to make your images start fast, you can probably save a lot of money by renting spot instances when testing ... I don't know if you can do spot instances of bare metal though, so you're probably stuck with vm overhead.