November 30 2003:

Last Month we covered the current CPU landscape with an overview of the latest and emerging technologies being offered by the 3 Major players, Intel, AMD, and Apple .This month I’ll cover the differences in the architectures of the competing CPU technologies, the apparent discrepancies in clock frequencies, and its relationship to overall Performance. I’ll also discuss comparative real world performance figures and have a few words on the dreaded benchmarks and the fallout created by the marketing departments being a little over zealous with their claims, and usually doing more damage than good. I’ll also have an overview on the emerging 64 Bit landscape, and its significance in the short and long term.

Comparing not only cross platform but also comparable X-86 performance has always been a bit of a barn fight. Although in the past when the architectures of the CPU technologies were similar, the clock speeds of the processors were in some way at least a rough guide to the overall performance of the CPU’s on offer. That changed dramatically on the Apple side with the introduction of the G3, which had a vastly different RISC (Reduced Instruction Set) architecture that featured an extremely short 4-stage instruction pipeline. This allowed it to perform far more efficiently than the competing X-86 architectures of the time. Both Intel and AMD had very similar CISC (Complex Instruction Set) based CPU with similar 10 instruction stage Pipelines, so at least on the X-86 front, the race for Mhz race was still well and truly on. Now while Apple was touting the line of the “ Megahertz Myth”, both Intel and AMD realised that Mhz still sells, and went headlong into battle trying to out do each other to the landmark 1 Ghz mark. That was won by AMD, much to Intel’s embarrassment. This battle set the landscape unfortunately for the one of the most misleading advertising spins of all time on Intel’s part with the PIV. The age of the Megahertz Myth was definitely upon us, and although now a days it is clearer that Mhz is not the end all of gauging overall CPU performance, there is still a lot of confusion.

O.K, lets see if can make this a little easier to understand.

Now without going too far into techno babble, one major factor for the apparent discrepancies in Clock Speeds of the competing CPU’s is IPC: Instructions Per Clock. The efficiency of a processor is directly related to how many instructions are completed per clock cycle. This is also directly related to the length of the Instruction Pipeline that is employed by the processor. The shorter the pipeline, the higher the number of instructions that can be completed per Clock Cycle. The trade off is that the shorter the instruction pipeline, the less the capability to increase clock frequency, the longer the instruction pipeline, the greater the capability of the CPU to be clocked. Remember that IPC is only one half of the performance equation, the other half being the clock speed that the CPU operates at. O.K, are you with me so far ??

Right, lets continue then -

The G4/G5 and the PIII / Athlon /Opteron all have a much shorter pipelines than the PIV which has a huge 20 Stage instruction pipeline. Due to the length of the pipeline the PIV is far more reliant on higher clock speeds to obtain its performance than all the other architectures. This not only explains the disproportional difference in CPU clock frequencies across the different architectures, but also the dismal performance of the earlier Willamette PIV's, which when released at 1.4 / 1.5 Ghz , were actually far slower than a 1 Ghz PIII. Red Faces all around. Of course there are other factors that come into play as well as IPC and Clock frequency, including FSB, Cache size and efficiency, branch prediction, pre fetching, Instruction sets, etc, etc, all play an important role in the overall performance of the CPU.

The architecture of the PIV has been designed from the ground up to be able to scale stratospheric clock speeds, while the AMD and IBM 970 approach rely on High IPC, which allows a far higher number of instructions to be processed at lower clock speeds. Both have their advantages and disadvantages, and the argument of which is the better will no doubt rage on, but in the end it is irrelevant in regards to overall performance. It is unfortunate that due to the Megahertz Myth, a lot of end users still equate clock speed with overall performance, and although Intel have been forced to do a back turn, especially with the performance of the Banias ( Centrino ) Mobile CPU’s, the confusion still reigns ..