> I really fail to see what people see even in Nvidia/Cuda They're a nice middle ground between a GP CPU and ASICs that is well suited to the kinds of workload

> I really fail to see what people see even in Nvidia/Cuda They're a nice middle ground between a GP CPU and ASICs that is well suited to the kinds of workloads that venture capitalists are, for the moment, interested in investing in. And it's possible that there are enough applications that benefit from GPU architecture that it's more efficient to force a GPU to pretend to be a CPU instead of having a CPU. Most likely it's just another cycle of the RISC vs CISC pendulum. Like cloud vs desktop, physical vs virtual, serial vs parallel, monolithic vs micro, etc. We'll stick with CUDA and use it to solve as many problems as we can with that method, and then as we discover new ways to create or organize GP CPUs, we'll shift back to those until we find that problems get hard to solve again.