From a CFD perspective, I think this analysis is flawed because of two things:<p>1) any organization thinking about buying a 1000-core cluster will have at least one in-house code. The reason being that none of the commercial ones really scale well to much beyond 50 cores. And that changes the comparison a lot, since any external support for an in-house code will be really expensive. (If you've ever looked at an in-house CFD code, you'd generally be mortified at the coding.)<p>2) they say that hardware costs are only 1/3 of total costs, but they are assuming a slow-ass interconnect that no-one in their right mind is considering today. Once you go up to a real HPC cluster with 1:1 non-blocking FDR or dual-rail QDR, hardware costs are more like 3/4. And really, you need that kind of interconnect to be able to scale around the memory bandwidth bottleneck, which any code that does stencil operations is bound by.<p>But if your organization is, say, a Gromacs shop working on chemical engineering, it could make sense.