I spent most of my career in compute acceleration.  Originally this meant FPGAs (reconfigurable logic) but later GPUs.  As I built larger and larger GPU machines using Intel server CPU and Nvidia GPUs the idea of smaller and simpler processors became more appealing.  I’ve been a fan of the Barcelona Supercomputer Center Mount Blanc machine as well as the SpiNNaker at Manchester for some time.

While GPUs won the battle for now, the power consumption and price seem to be pointing back in the direction of smaller, lower power processors, perhaps with more tightly coupled, on-chip, parallel accelerators.

China seems to be moving in that direction with their new 2 Exaflop Lingshen machine using only low-power CPUs.  This machine has only homegrown (Chinese) processors and no GPUS.  This is quite a difference from supercomputers of the last decade or so, as well as all the new AI data centers.

China announces CPU-only exascale supercomputer with 47,000 homemade processors, record 2 Exaflops of performance without GPUs — Lingshen super said to use Huawei Kunpeng servers and no foreign-made components