3 ms·
Given that we have magic compilers that convert mostly sequential CUDA into highly parallel SIMD assembly code, I don't think that the Itanium concept was too f
by dragontamer 5y ago
Given that we have magic compilers that convert mostly sequential CUDA into highly parallel SIMD assembly code, I don't think that the Itanium concept was too far off.
But all of the details had gross amounts of complexity. More modern VLIW architectures are more obvious about their parallelism... Without any of the weird 'bundles' that made the Itanium terribly complex.
I still think that 'simple decoder' has hope, and I point to NVidia Volta and Turing as my case in point. At the assembly level, the NVidia SASS compilers generate the read write dependencies. NVidia has proven that the compiler can indeed specify the dependencies at compile time, though they require a huge amount of software bulk to be able to do so (PTX pseudo assembly that compiles into SASS later)
The issue is that the Itanium was poorly designed and impractical. But the overall concepts seem to still be possible (indeed, more possible given today's technology). I'd be interested in a modern Itanium, but instead split as follows: where the compiler determines the dependencies but the core figures out the precise scheduling.