Whats holding ~100% D GUI back?

Mon Dec 2 20:08:15 UTC 2019

For anyone interested in how regular GPUs work, this report is 
pretty interesting:

https://www2.eecs.berkeley.edu/Pubs/TechRpts/2016/EECS-2016-143.pdf

It explains how Single-Instruction-Multiple-Threads execution 
units hides latency, by interleaving multiple threads. So, 8-32 
execution units are working in lockstep, but the impact of stalls 
are reduced.

The typical on-die Intel GPU seems to have 8 execution units per 
slice and 2x 128-bit ALUs per execution unit.