Sorry to hear that Dan has passed away. I worked with Dan in the early days of the old DoD HPCMP Programming Environments and Training (PET) program. I helped him teach a couple of parallel programming classes for some Air Force people as well as some of the submarine folks at Electric Boat Company in Connecticut. Very nice guy.
I’m old enough to remember when it became apparent that Cray was going to make a lot of money with vector processing that companies like Floating Point Systems (if I remember correctly) would sell you hardware and software that you could scab onto a VAX 11-780 to do vector processing. Later in the 80s, IBM would sell you a vector processing box you could add to a 3090. Now we have processors that have “vector processing” built into them and other than some compiler options and some knowledge of the best way to structure loops to achieve optimum vectorization its mostly transparent to users. I think (hope) this will be the path for GPUs. Using them should NOT have hardware specific instruction other than something like do concurrent. The reason we have compilers in the first place is to relieve programmers of the burden of writing everything in machine code by hand.