A Generational Garbage Collection for Processing-in-Memory
2025-07-09 23:46:58.112637+02 by Dan Lyke 0 comments
It has been a long time since I looked at alternate architectures for anything other than cost (eg: embedded), I haven't even done a lot of thinking about massively parallel stuff since my memory of RenderMan's shader language fell out of cache, but I'm finding Gray-in-Young: A Generational Garbage Collection for Processing-in-Memory's description of the UPMEM architecture (now owned by Qualcomm) fascinating.
A 32 bit RISC core with 24 hardware threads dispatched at least 11 cycles apart (which has some interesting saturation/full-utilization implications) with
... a 64-KB scratch pad memory (SPM), a 64-MB DRAM, and a 24-KB instruction memory.
So some really interesting system constraints...