Just the other day I sat on one of the Computer History Museum's Cray I's. As much as I appreciate the fact the Cray machines come in industry-standard racks that fit seamlessly in any data center, I miss the age of exotic shapes and the immediately recognizable machines.
The old Cray-1s certainly are iconic and beautiful machines. Just from my personal experience working in HPC for the last 15-20 years, what I think is driving us away from these cool old-school designs (Thinking Machines were always my favorite) is the sheer size of modern supercomputers.
They're just staggeringly large, and there's no way you could get them built, disassembled, shipped, and reassembled without these racks. (They still do come in custom racks a bit bigger than normal, and inside they look nothing like normal racks.) It's an event when the trucks all show up to unload these things and the only thing that makes them manageable is that it's a normal form factor on a pallet you can move around with a jack. Additionally, at least in my experience, a good portion of them usually gets sent back so you're always in there taking them apart and putting them back together. Anything "creative" would just get in the way.
I can't get a good hyperlink on my phone to an artists rendition of Titan at Oak Ridge Road or Mira at Argonne (the ceiling is too low so you can't photograph all the racks at the same time, but do a Google Image search for it.) It's staggering just how big it actually is. They may not be that unique looking any more, but they're still damned impressive in person (until you have to fix one. Then the magic wears off pretty quickly.)
Edit: I do wish they would stop putting the murals on the side of them. I think a big noisy $300,000,000 black box has a certain cosmetic appeal to it that they ruin with the flashy plastic covers.
Computers used to be way more diverse and interesting. Unfortunately, that made using them a pain. So, industry converged on more similar things where possible. Economics factored in given enormous investments required into new chips. The drag of mainstream computing away from stuff like old Cray's is like a force of nature.
Doesn't mean a niche player can't come along with a neat box with good differentiators. Job's Next Cubes give me hope here. :)
Even with standard rack-mountable equipment, it can't be that difficult to do it when you have a couple hundred processor boards to arrange in space. And the vertical piles of boxes is, most likely, not the optimal arrangement.
They can still be diverse to leverage certain strengths as long as there's enough compatibility that you don't have to rewrite everything from scratch. I'm hopeful for lowRISC as the spiritual successor to BS-5000 and i960.
Potentially. Far as successors, look up crash-safe.org and Cambrige's CHERI processors. Those are cutting-edge successors to tagged (i.e Burroughs) and capability (i.e. i960, Sys38). Draper is adding SAFE's PUMP to a RISC-V corr they plan to open-source and turn into an ecosystem. CHERI is OSS plus already runs a FreeBSD port called CHERIBSD. That CHERI supports legacy, C software and runs BSD means it's ahead.
I considered porting EROS capability system, GenodeOS, or Muen to CHERI. Plus cutting it down from 256-bit to 128-bit pointers as the extra bits are just there for experimentation.
I recall watching the fancy cooling mechanism and fluids in motion. I believe one of them also functioned as a couch of sorts or at least looked like one.
Doesn't sound like the Cray Graph Engine is available separate from the Urika-GX package. Not that it necessarily would be useful without the Cray programming environment, but the open source remarks are unclear and make one wonder if they're going to release it to the public.
The Urika I got to try at CSCS was probably the easiest graph database I ever experimented with. Very single user focussed but really nicely done (single user focus might have been CSCS as a supercomputer centre decision).
