Interesting perspective re CXL synchronous API. Wouldn't things like OOO execution and speculation help with that? And anyway the latency is supposed to be comparable to NUMA latency, is that really such a deal breaker?