Yep, this also reflects the design target of GPUs targeting much larger working sets, so have higher main memory bandwidth and rely less on caches. CPUs rather

Yep, this also reflects the design target of GPUs targeting much larger working sets, so have higher main memory bandwidth and rely less on caches. CPUs rather have few fast threads of execution working on hot cached data than many slow ones talking to main memory (because N-way thread level parallelism often splits your cache N ways, to N working sets)