Joe, Re: "While Itanium can perform well as a database server, where there is a lot of repetition/parallelism in sorting and comparisons, there is very little parallelism in business logic."
Define "business logic". In terms of software that small, medium, or large businesses use, there are many different genres that software can be divided into. Financial software, for example, can have a lot of parallelism in it. So can any software that produces graphs, vectors, or analytical output. Applications that create content, that refine images, sound, or multimedia, applications that render 3D, transform data, or display graphical formats can all have greater levels of extracted parallelism. I think you are underestimating just how many applications Itanium can be good at.
Re: "Is this a compiler generated execution and prediction? I didn't know Itanium had any built in capacity for speculative execution and prediction."
Check out this early presentation. It will explain a lot.
developer.intel.com
Re: "I think McKinley has more execution and address calculation blocks than Merced. Are these explicitly addressed by the compiler, that is, does the compiler know how many there are, or is the executable in sort of intermediate form, and the processor assembles the intermediate instructions to what gets actually executed?"
Unlike x86 code, IA-64 has the compiler dispatch data sets in terms of several predefined templates. See this slide.
developer.intel.com
McKinley has two more integer / memory execution blocks as Merced, but it does not add any more templates to the library, so the compiler does not need to be updated. In Merced, if there are no available execution blocks to service a given template, the data is suspended for a clock cycle until more resources are available. Adding more execution blocks decreases the probability of this happening. With two bundles filled with three symbols each, an IA-64 CPU can service as many as six instructions every clock cycle. To be fair, most software doesn't have enough parallelism to fill all six slots on every clock cycle; however, new compiler breakthroughs for Itanium architecture are appearing all the time, and each has new algorithms for keeping these filled.
"Or how else will existing code take advantage of additional resources provided by McKinley? Or does the software need to be recompiled for each processor?"
At MPF last fall, Intel mentioned that McKinley would offer 1.5x or greater performance gains over Merced by using the same binaries, and recompiled binaries will offer as much as 2x performance in certain situations. Clearly, there are some features that only improve performance in specifically optimized code, but it seems like there will still be plenty of performance, even without having to recompile.
wbmw |
