CPU Architecture War : Clock Speed v.s. Instruction-Level Parallelism

An interesting history to look at,where each CPU makers made their way to deliver fastest CPU at that time.

* At first, Intel creates their CPU to have faster and faster clock frequency. They introduce simpler design but deeper pipeline with higher clock frequency to improve CPU performance with the introduction of Pentium 4 and Pentium 4 Extreme Edition. This method soon hit power wall and thermal limitation which make they realize that increasing clock speed will no longer a feasible way to improve CPU performance. Intel look back to their Pentium III architecture, leaves their Pentium 4, and begin to create smarter CPU using more hardware Out-of-Order (OOO) execution engine, trying to squeeze every last drop of instruction-level parallelism (ILP) out of the code, even if it costs millions of logic transistors and years of design effort to do it. From Pentium III, Intel paves their way through Core 2 until current generation of Intel Core family, the Broadwell even the Skylake, with significant improvement of instruction-level parallelism.

* AMD, on the other hand, gives a very nice starting point by introducing their Athlon and Athlon 64 architecture. I adore these two CPUs so much because it can give better performance-per-price ratio and better performance-per-watt ratio compared to the Intel Pentium 4 and even the Intel Core 2. Infact, I have used it since 2005 (Athlon 64 / K8) and 2008 (Athlon II X4). After the phenomenal AMD Phenom architecture, I don't know why AMD follows Intel mistake by introducing CPU with higher and higher clock frequency while sacrificing the instruction-level parallelism. AMD CPU performance starts to fall far behind Intel's after the introduction of AMD Steamroller architecture and so on. I hope the next AMD Zen architecture will bring back the ILP path for AMD CPU because I have heard the engineer who designs the AMD K7 and K8 will join the effort to design the Zen architecture.

*The last one is ARM, the newest player in CPU architecture. ARM splits their way into three path, depends on who design it. Apple try to make the ARM their own way by improving ILP throughout their design (Apple A6 and Apple A7/A8). While Qualcomm with their Krait Architecture try to balance between Clock Speed and ILP. ARM itself create their A57 by improving the clock speed rather than ILP.

This CPU war is not yet over. We clearly see some CPU makers change their mind in designing their own CPU architecture, from 'clock' oriented into 'ILP' oriented or vice versa. Maybe in the future, they will create more advanced Out-Of-Order Execution Engine with less power overhead, more sophisticated branch prediction, or even just a simple design with new material technology that can improve thermal efficiency. Not to forget, we will soon face the end of silicon era as our semiconductor process technology becomes incredibly small, 5nm is expected between 2020-2021 (FYI, first Pentium 4 used 180nm process technology while the last Pentium 4 used 65nm process technology).

(C) 2016 Bagus Hanindhito
Figure from www.lighterra.com / Jason Robert Carey Patterson