optimizing for itanium

Here’s a really interesting paper from USENIX 2005: Itanium — A System Implementor’s Tale.

It’s got a good overview of the Itanium architecture in general, and has details for the geeks out there. The money paragraph:

…we were able to eliminate all but one of the 15 data-load stalls, resulting in only 3 bubbles and a final execution time of 36 cycles, or 24ns on a 1.5GHz Itanium 2. This is extremely fast, in fact unrivalled on any other architecture. In terms of cycle times this is about a factor of two faster than the fastest RISC architecture (Alpha 21264) to which the [Linux] kernel has been ported so far, and in terms of absolute time it is well beyond anything we have seen so far. This is a clear indication of the excellent performance potential of the Itanium architecture.