I seem to recall the earliest "real-time" Visible Human volume rendering demo being run on either a Cray or IBM (?) supercomputer back in the late 1990s. But, I couldn't remember enough keywords to find a reference and confirm it.

What I recall was that it was a distributed (clustered) machine type, not a shared memory model like the Origins and not having significant GPU hardware. The central hack was recognizing that the total RAM of the multi-node supercomputer was large enough to hold the large volume data in a chunked, distributed fashion. An MPI job ran a software renderer in parallel on all these chunks, with a 2D gather+compose to produce the final 2D image for viewing.