Official AMD Barcelona Thread

[Shadowmage]

I did some digging in the AMD 10h Programmer's Manual and Intel's Programmer Manual.

Barcelona shuffle latency (PSHUFHW): 2 cycles latency
Conroe shuffle latency (PSHUFHW): 2 cycle latency
Penryn shuffle latency: 1 cycle (from super shuffle engine)

Barcelona 128bit division latency (DIV*): 16-20 cycles latency
Yonah 128bit division latency (DIVPD, DIVSD): 63 cycles latency (not a typo!)
Conroe 128bit division latency (DIVPD, DIVSD): 18 cycles latency
Penryn 128bit division latency (DIVPD, DIVSD): ~5 cycles latency (estimated from Radix-16 divider)

Barcelona 128bit square root latency (SQRT*): 21-27 cycles
Yonah 128bit square root latency (SQRT*): 118 cycles (not a typo!)
Conroe 128bit square root latency (SQRT*): 29 cycles
Penryn 128bit square root latency (SQRT*): ~9 cycles latency (estimated from Radix-16 divider)

Conclusion: Shuffle speed improvement for Penryn shouldn't make too big of a difference... 2 cycles latency was already very low. Division/sqrt isn't that common (but is common in F@H) and performance will improve accordingly.