dgrdsv
Member
Are you insinuating that 8 ACEs are a waste of silicon?
For graphics - yes. This is something that both AMD and Sony agree with btw. A typical async compute usage in graphics is unlikely to go higher than 5 compute queues concurrently with graphics as this will make optimizing the flow nearly impossible and may actually lead to unpredictable performance even on GCN. So that's 5 queues out of 1 ACE or to get better launch granularity - 5 ACEs with 1 queue in each. The truth will probably be somewhere in the middle. In any case - 8 ACEs are an overkill for concurrent execution and will be truly used only in a pure compute workloads.