The Sandy Bridge Preview (Anand)

[qcmadness]

FMA support is already wide spread in processors and code compilers. C99 supports FMA through the fma standard math library function, nvidia gtx2XX and new gpu's support it, amd's 5XXX gpu's support it, the cell processor, fujistsu SPARC VI, itanium.... BD will support FMA4 and Haswell will support FMA3.

also, windows 7 sp1 will support AVX as does linux kernal 2.6.30

For Evergreen series, only HD5800 and HD5900 series (RV870-based) support FMA.

I hope Intel will implement FMA in Haswell in the right way, keeping a throughput of two instruction per cycle (FMA+FMA or FMA+MUL/ADD). One FMA has no advantage over separated MUL & ADD (except a bit better accuracy), but some serious disandvatages.

What disadvantages?