Stardate 20020128.1206

(On Screen): In a relatively low-key announcement, Apple admits that the rumors about the new PowerMacs were correct: the new top-end system is a dual 1GHz G4. Characteristically, Mac advocates and the more gullible technical press announce that the gigahertz gap with PCs has been closed.

A closer look at the system specifications, and the grandiose claims being made for performance, demonstrates that this is baloney. For one thing, Apple yet again tries to claim that it runs Photoshop vastly faster than a P4, and yet again provides no details which would permit anyone else to duplicate the test. They do not say what test they did, nor do they say how the P4 was configured. And it is virtually certain that they are using Photoshop 6.0, which is known to be crippled on x86's; and not Photoshop 6.1 where the bug was fixed, or the later release which contained SSE2 optimizations for the P4.

Fact is, even the single processor 933MHz and 800MHz Power Mac G4 systems outperform the Pentium 4 by 51 percent and 26 percent, respectively. How, you might ask, can even the entry-level Power Mac G4 outperform a PC with a Pentium 4 processor running at more than twice the megahertz? It’s further proof that megahertz alone is a poor indicator of real-world system performance, particularly when comparing different overall system architectures.

I'm afraid not. What it does prove, however, is that Apple is guilty of false advertising, and I'd like to see the FTC on their case.

A detailed examination of the system specifications of the new dual-processor system reveals a terrible flaw in the design: the system memory overall is ordinary old 133 MHz SDRAM. That is nothing like fast enough to keep two processors at those speeds fed; they are going to be spending a lot of time tossing wait-states.

To compensate for that, Apple yet again relies on L3 cache. Interestingly, the L3 cache that they boast about is the same memory that all new Athlons use as their main memory: 266 MHz DDR. But where the new PowerMacs have 2 megabytes of cache at those speeds, all of the Athlon memory runs that fast. And with the release of the 845 chipset, it's also possible to use that memory with the P4.

Cache is a mixed blessing. While it can speed access to commonly used memory locations, it actually slows access to the rest. There are two ways it can be handled.

First, an access to the main memory can be deferred until after the cache determines that it does not have the required data; and that adds latency to the access (i.e. it slows it down). Or, the main memory can be asked for the memory in parallel with the cache probe, and then the request cancelled if it turns out that the cache actually had it.

The P4 does the second one when it is used with RDRAM, but that's because RDRAM has the bandwidth to support it. A P4 running with RDRAM is constantly making unnecessary accesses to RDRAM and throwing away most of the data which is returned.

Apple doesn't have that luxury, because its main memory is too slow; if both processors are making speculative fetches to memory, it's going to choke even worse. So they must necessarily wait until the L3 cache announces a miss before going to main memory -- and that means that accesses to main memory will be even slower than on a system which does not use L3 cache.

L3 cache is a desperation move; it's what you do when you cannot make main memory run fast enough. The big question is why it is that Apple didn't make the switch to DDR for main memory and kiss off L3 cache. The answer has to be that Apple's chipset group has fucked up again.

Apple designs its own chipsets, and they've been inferior for years. They have always had lousy latency and they've always been behind the curve on memory speed. Apple only switched from 100 MHz to 133 MHz SDRAM about the time that the PC world was abandoning 133 MHz SDRAM for being too slow. The PC world switched to DDR-SDRAM and RDRAM more than a year ago and Apple still doesn't have the ability to use it for main memory.

Which means that Apple's claims about speed must be cooked again, in a very special way. In order to get the results that they are (unless they simply made them up) Apple must have carefully crafted the test so that the data being manipulated fit into the L3 cache.

If Apple is so proud of their performance, why won't they reveal the details of the tests they did so that people can examine them to see whether the tests were honest?

It's because they wouldn't survive scrutiny.