3 ms·
VLIW seems to have solidly lost to SIMD in the desktop space. Don't forget: as Intel was floundering with Itanium, AMD was just making SSE and SSE2 instruction
by dragontamer 2y ago
VLIW seems to have solidly lost to SIMD in the desktop space.
Don't forget: as Intel was floundering with Itanium, AMD was just making SSE and SSE2 instructions faster (AMD 3dNow! never became popular, but they continued to push that). It seems like SIMD was the historically correct choice over VLIW.
Today: every high-speed program in Desktop/Server is either SIMD-accelerated CPU (NEON or AVX512 or SVE), or full-blown SIMD GPU.
- bsder 2y agoSIMD didn't really win for general purpose workloads, either. It's just that execution units are so small relative to everything else that you might as well expend some extra silicon for them. SIMD is only really wins for very specific workloads (memory copy and 3D graphics, for example).
- dragontamer 2y agoVLIW seems to be in that awkward space where (at least in the 00s), compilers weren't good enough to beat traditional code + traditional CPUs. But tweaking the assembly for VLIW / Itanium was still too hard. So SIMD (which is much harder to write than VLIW) wins because if the programmer drops down to the assembly level... "might as well" go for the harder SIMD-model that achieves slightly better performance over VLIW. I do wonder if today VLIW machines with modern compilers could be a thing. Compiler technology is so much better now, maybe we are in the era where VLIW could be renewed. But on the other hand, CUDA shows that SIMD-compilers / SIMD-languages have improved as well.