No real Benchmarks?

Apples to Oranges?

How's AltiVec compare to MMX, 3D-Now, and SSE? Well, MMX is a joke. It always been considered more hype than substance, so I won't even compare them. SSE and 3D-Now are pretty much comparable to each other, although 128-bit vs. 64-bit registers give SSE the advantage. AltiVec kicks both their butts however. Why? Partly because the AltiVec is a newer implementation with more and better instructions. However, it is mostly because of the number of available SIMD registers. SSE has 8 registers available for it, while AltiVec has 32.

For the uninitiated, registers are the locations on the processor that hold the numbers being played around with. For most instructions, data has to be brought in from the memory or cache to the register before being manipulated. This is true with integer and floating-point operations as well as SIMD. Less registers means that a Pentium III will run out of data to process faster than a G4. It also makes it more difficult to do complex calculations that involve more than 8 x 128-bits of data. This problem isn't limited to the SIMD unit, developers have struggled with the x86's lack of registers for years in the integer and floating-point arenas.

At Seybold, Steve Jobs flashed the chart below up on the screen to demonstrate AltiVec's superiority to SSE.

Misleading numbers?

The figures above represent the theoretical performance of a G4 and a PIII running several standard linear algebra functions using their respective SIMD engines. According to this chart, the G4, even with a 100Mhz deficit, is almost 3 times faster when performing linear algebra functions. You should notice several things, however. First, this is not really a test. It is a mind game. These results were obtained by engineers sitting down and writing out the most efficient possible algorithm for performing each task, and then figuring how many cycles it would take their processor to complete that algorithm.

So what you're saying is…

These are not benchmarks! They do not mean anything except that the Motorola engineers built a more efficient SIMD engine than the Intel engineers. This owes a lot to the fact that the Intel engineers are saddled with a 25 year old architecture, while Motorola completely threw their 68000 line of Apple processors out the window and started clean with the PowerPC.

Most PC enthusiasts know that theoretical performance doesn't mean anything next to real world performance, and it's here that things get a little trickier for Apple. Remember, a processor can't manipulate data that it can't get its hands on, a fact that Apple tries to mollify with a large (1MB) L2 cache. Unfortunately, Apple will not be using one of the coolest features of the G4, it's ability to use a 128-bit memory path. Until the new systems are released, it will be difficult to gauge how much of a problem this will be, but its something to keep in mind.