[pjmlp]

Their design is completly different, it isn't shared execution units like it happens on CPUs.

[juggertao]

But it is. GPUs have many more threads in flight than execution units.

[pjmlp]

Threads groups get exclusive resources in SIMT execution pipelines.

[juggertao]

And at memory stall they are exchanged with other waiting thread groups.

Just like HT.

[RealHusky]

Scheduling algorithm is different.

CPUs target low latency (they switch often). GPUs target high troughput (they switch rarely, only when needed).

High troughput algorithms dont have problem with a lot of threads. Low latency algorithms have problem with a lot of threads (they need lot of cache memory because of constant switching).