"Crippling" The Pentium 4, And Its Long Term Perspectives Cont.

The Pentium 4 has a 20 stage pipeline, twice that of Pentium III. Yet, it would likely be surprising to the reader to find out that two of them do virtually nothing. For emphasis, there are two stages that don’t do any useful work on the data. Anyone who knows anything about pipelining knows that, while more stages mean shorter clock cycles and thus higher frequencies, it also means a much higher penalty for a branch mispredict. So why would Intel add two stages that do nothing useful, but actually hurt performance due to penalties? Simple. Clock speed. The die of the Pentium 4 is large enough, and the target frequency high enough, that the electrons just can’t move to the next stage fast enough given the length of the targeted clock cycle. Two stages simply “move” the information to the next stage. Rather than make certain stages have a 2 cycle latency, they instead break it up into two, one stage has the function, quite literally, to do nothing other than to move whatever it contains to the next stage. This allows Intel to continue to ramp the core up to very high speeds.

At the introduction of the Pentium III in a .18 micron process, the fastest available chip was 733Mhz. At the introduction of the Pentium 4, the fastest available chip is 1.5ghz. That’s over twice the clock speed: in the same process technology. This is really the saving grace of the Pentium 4, or rather, the one thing that’s keeping it from being a complete and utter flop.

Consider that the fastest Pentium III Coppermine at introduction was 733Mhz Pentium III. The Pentium 4 has a top speed, at introduction, of 1.5ghz. In some “worst case” (SYSmark2000) scenarios, the Pentium 4 was about as the 1ghz Pentium 3 on an i840 board (both use dual channel RDRAM memory) *. In best case scenarios such as low resolution Quake 3, The Pentium 4 churned out 33% more frames (191fps P4 divided by 144fps P3), or the P3 churned out 25% less frames (144fps P3 divided by 191 P4) depending upon how one looks at it. The Pentium 4 scores anywhere between these two extremes. In either case, the consumer should not be upset. Sure, Intel’s new machine doesn’t smash up their previous design right now per clock. No, it doesn’t always beat out the 1.2 ghz Athlon either. But it does in some cases.

* Scores taken from the AnandTech review of the Pentium 4.

What Intel needed was an answer to the Athlon, as it is rolling over P3’s due to its ability to clock higher. They got it with their Pentium 4. They have a CPU which will resume where the Pentium 3 left off - just trailing the Athlon in performance. The secret is that on the same process technology, Intel can deliver more performance with the Pentium 4 than with the Pentium III, and this is without the recompilation of applications. This is because 1.5ghz is only the beginning for the Pentium 4. After recompilation, and after more software is written from the ground up with the Pentium 4 in mind, the Pentium 4 will continue to increase its performance even at the same clock speed. The Pentium III has no more headroom in the .18 micron process, and thus won’t improve without better compilers. The Pentium 4 has much more headroom on the same process. On any given process, the absolute top performance of a Pentium 4 will exceed that of a Pentium III.

The Pentium 4 is targeted towards the Internet (big surprise). While right now much of online existence is gaming, email checking, and web browsing (how much fast a machine do you need for the latter two??), it is becoming more and more stream oriented, with large graphics, videos, and sound. These are the areas in which the Pentium 4’s high bandwidth, and large cache lines will be useful, instead of detrimental. Because they are stream oriented, much more data is contiguous, making the miss rate very low, and the use of large cache lines well supported.