That is just a software view provided by the CUDA compiler and the NVIDIA device driver, which take care of distributing the computation over SIMD lanes and real hardware threads (called warps in the NVIDIA obfuscating jargon).
The GPU processor is just a processor with SIMD and FGMT (fine-grained multi-threading). There is nothing special from this point of view.
For any CPU you can write a compiler that provides the "SIMT" software model and there are such compilers for the Intel/AMD CPUs.
SIMT is just one of many useless synonyms created by NVIDIA for traditional terms previously used for decades in computing publications, i.e. it is the same with what Hoare called "array of processes" (in 1978-08, later implemented in the language Occam) and actually the same with "parallel for" of OpenMP, just with a slightly different syntax (i.e. CUDA separates in the source file the header and the body of the parallel FOR).
alt Hacker News
Nope.
That is just a software view provided by the CUDA compiler and the NVIDIA device driver, which take care of distributing the computation over SIMD lanes and real hardware threads (called warps in the NVIDIA obfuscating jargon).
The GPU processor is just a processor with SIMD and FGMT (fine-grained multi-threading). There is nothing special from this point of view.
For any CPU you can write a compiler that provides the "SIMT" software model and there are such compilers for the Intel/AMD CPUs.
SIMT is just one of many useless synonyms created by NVIDIA for traditional terms previously used for decades in computing publications, i.e. it is the same with what Hoare called "array of processes" (in 1978-08, later implemented in the language Occam) and actually the same with "parallel for" of OpenMP, just with a slightly different syntax (i.e. CUDA separates in the source file the header and the body of the parallel FOR).