www.enchantedage.com

Sometimes, you hear statements like "an Xbox has nnn % of the performance of a Windows PC" where nnn will vary from 25 to 600 or so, depending on context. However, I disagree with any such statement -- you simply can't make that statement without being a whole lot more specific about what you are trying to measure.

You first need to understand how performance on CPUs actually work.

Then, you need to understand the difference between managed/JIT-ed code, and native C/C++ code.

Then, you need to understand the difference between the desktop CLR (used on Windows), and the compact CLR (used on Xbox).

Then, you need to understand the difference between the Intel CPUs (out-of-order, speculative, deep pipelines) and the PPC CPUs (in-order, shallow pipelines).

Once you understand these things, you may be able to formulate your question such that someone who can't read your mind can answer it. Or, more likely, you'll already know the answer to your question, and won't need to ask it.

However, in general, the Xbox CPUs can do "simple math" that stays in cache really quickly. Similarly, if you use the extended AltiVec available on those CPUs, you can get great vector math throughput -- in aggregate, much better than that of a single Intel dual-core CPU. Unfortunately, you don't get to use those instructions when using XNA on Xbox. And, even more unfortunately, the compact CLR is not very good at generating code, so calling functions and passing parameters is a lot more expensive than it should be. It also doesn't auto-inline like the desktop CLR. Thus, for 3D math, the XNA compact CLR will not perform very well; you'll have to be careful in how you structure your program to take this into account.

However, for many algorithms, the time cost is actually in the memory access patterns (cache misses) rather than the CPU instructions used. Here's where the hyper-threaded CPU cores come in. When one thread stalls, waiting to fill a cache line, the second thread will be scheduled. If it currently has the data it needs in cache, it can continue while the other thread stalls -- filling in the blank space, as it were. Meanwhile, on the Intel chips, they will speculate ahead to attempt to both do some useful work ahead in the program stream while waiting for the cache line to fill, and in an attempt to pre-fetch data that might be needed "soon." That's a very different approach to performance, and makes it very hard to compare the CPU architectures as apples to apples.

Any statement of "One Xbox equals NNN percent of a Windows PC" is so vague as to be mostly misleading, because the math doesn't divide nicely like that. The various parameters are too different.