- Apr 27, 2000
That seems to be the direction Fujitsu is taking with A64FX, and others are welcome to try, but there are reasons why dGPUs have taken over in HPC. And it's not that the programming model for external compute cards is particularly fun to use.If you build CPUs designed for HPC (as used to be done a few decades ago before hundreds then thousands and beyond of off the shelf RISC or x86 CPUs were done - i.e. Seymour Cray's stuff at CDC and then Cray) by including very wide SIMD instructions and and some fast local memory then they are going to be better than GPUs because they are designed for computation, not graphics with a side of "oh hey we'll support HPC too because we found out we can charge way more for HPC cards than GPUs".
Mostly it has to do with the availability of resources to commit to design and implementation, as well as the applicability of designs to multiple markets. It's easier to "cobble together" a supercomputer out of server CPUs based on commodity desktop core designs and dGPUs based at least loosely on commodity desktop dGPU designs than it is custom stuff like A64FX because A64FX has exactly one market: HPC. Design costs on cutting-edge nodes are pretty high. I can sell (for example) Zen3 cores all across server, workstation, desktop, and mobile markets, but I can't necessarily sell designs derivative of A64FX in the same way. I can basically say the same thing about NV's designs (some of the work they put into their high-end compute cards also goes into their consumer dGPUs serving many markets).