Anthony Frausto-Robledo, Editor (afrausto@architosh.com)

7 Oct 00

Pipelines, Pentiums and G4's oh my!

This brief report is dedicated to all my PC friends...you know who you are ;-)

Much has been written about processors over the years...and almost all of it flys over the head of PC and Mac users alike. It seems that processors are just so damn complicated that computer users leach on to any simplification—what-so-ever—in order to provide some basis of comprehension.

Of course they do. Can you imagine walking into your local CompuUSA without any mental yardstick for evaluating a computer's processing power? It would be like buying a car without knowing the horsepower.

Imagine:

"Sir, how powerful is the engine in this new Accord?"

"Yes, it has a 3.0 liter, V-6 engine with 24 valve VTEC, with a sophisticated second-order balance system."

"Hmm...sounds like it might be powerful."

Indeed. However, that Accord has 200 horse power @ 5500rpm, though it sounds more impressive. Thankfully, "Horsepower" is a good shortcut measure of a car engine' performance potential. And we can use that.

If only MHz could be more like Horsepower

Horsepower is a more useful yardstick of performance (speed potential) than MHz is essentially because it 'approximates the qualities of a sum'. The sum of benefits or larger items. However, MHz does not work like this at all.

MHz does NOT sum up the benefits of chip cache size, pin to pin proximity, copper versus aluminum, pipeline depth and latency, number of registers, and types of processing units. It just doesn't do that. MHz is simply the rate at which the input and output signals of the millions of transitors operate. For example, a 300 MHz G4 has millions of transitors which will open and close at a rate of 300 million times per second.

However, MHz does nothing to tell us how many transitors are in the 300MHz G4. That matters immensely since the millions of on-off switches we call transitors are creatively put together to form little adding machines.

Of MHz and Men

Sometimes there comes along a fellow who can truly explain the problems of Mhz. It so happens that this fellow named Jon did a pretty good job of that. And by really talking about everything but MHz. [Original MacWEEK article]

I have summarized Jon's comments below because they were written with the type of care people usually employ in online message boards (not much). Jon was discussing the pipeline length in the upcoming G4e chip, which may take the G4 from 4 stages to 7:

I wish people who don't really know about things like the pipeline system of chips wouldn't talk about it.

Look folks, the K6-III's pipe was six stages. It reached 450Mhz. Maybe 500Mhz with the K6-III+. The G4 has four stages, it reaches the same [MHz].

Jon discusses how the pipelines of chips generally set the MHz, rate at which transitors switch on-off. The Pentium III's and Athlons have 12 and 10 stages, respectively, and reach up to 1.12GHz.

Basically once a chip is designed it wont go any faster unless you change the design (rumored G4+) or change how it is manufactured.

Jon means faster as in put info into the chip (input) and get that info back out (output). He then goes on to describe three ways to increase this process: (1) design a new chip core (as in Pentium P6 architecture changing over to Pentium 4), (2) tweak the design of the core (as in PentiumPro, Pentium 2, and Pentium 3 chips, all tweaked P6 architectures), or (3) improve the manufacturing process.

In discussing this Jon describes,

To speed the getting the signal from the input to the output pins you can either speed it up (ie switch to Copper from Aluminum) or make the distance shorter.

To make the distance shorter you just bring the pins closer together by making everything between them smaller. This is done by the .25 --> .18 --> .13 etc. switches in manufacturing technology. You'll notice as you shrink the speed, increase is porpotional. This is regardless of the chip [architecture (core)].

Any serious computer user has read at least one article about the Pentium or G3 moving over to a smaller micron (millionths of a meter) manufacturing process, effectively placing the millions of transitors closer together. Shrink the chip through this process and you get a faster chip, proportional to the size decrease. Jon notes that all chips work this way, even graphics chips.

For example the [Nvidia] TNT was introduced at 90MHz, and became the TNT2 which ran at 125Mhz when it was made on a .18micron process instead of .25 micron process. Funny how .25 is 39% bigger than .18 and 125Mhz is 39% faster than 90Mhz, isn't it? :)

Some Closing Points About Pentiums and PowerPC chips

There have been four generations of architectures for PowerPC chip designs over the years. PowerPC 601 (G1) through PowerPC7400 (G4). With each new architecture the chips got faster MHz for MHz because there cores changed. The Pentium has been with the P6 architecture from the beginning, modified or tweaked over the years for better performance.

Second, over the last few years in particular, the G3 and G4 have used smaller micron technology than the Pentiums and have moved to copper. Both of these have contributed to making small microprocessors that use less energy, run cooler and speed up the overall process of calculating. The Pentiums have yet to employ copper which adds greatly in the performance of the chip.

MHz is interesting but to use it as the basis of chip performance is like using RPM potential alone as the basis of car engine/car speed performance. And we still didn't talk about things like vector processing units (AltiVec) on the G4 or cache sizes of the L1 and L2 cache. Cache size is much smaller on Pentiums and Pentiums don't have vector processing units.

Regardless of all this, wouldn't it be nice if MHz did summarize chip performance?

[Editor's note: More discussions in the Small Pipeline Got You Down? thread are quite excellent in discussing MHz in chips and pipeline depth. Recommended!]