Much has been said about how the processing power of modern CPUs and GPUs grows exponentially. What’s often neglected is that typical software is written in a sequential fashion, where each execution step assumes the previous one has already finished. Running a single-threaded Python script on a 1000 core computer does not necessarily result in a significant speedup compared to a 1-core machine. Consequently, an exciting performance metric for comparing different CPUs is how long they need to execute a given program’s command sequence sequentially. Single thread performance captures these criteria. Here we use PassMark’s performance scores, as data for single-thread performance for an extensive collection of different CPUs is available on their website.
The image below shows a selection of high-end desktop CPUs from Intel (in blue) and AMD (in red) over the past 15 years. On the y-axis, we have the single-thread performance according to PassMark. The shaded curves in the background represent the regression curves for hypothetical CPUs that become 5% (in orange), 10% (in yellow), and 15% (in green) faster each year.
In this figure, we observe that the overall industry trend of the past 15 years is slightly more than 10% per year. However, an even more interesting observation is that this increase is not distributed uniformly, but progress happens in jumps. For instance, Intel’s Core2 Duo and Core-i 2000 series (Sandy-Bridge) brought a significant improvement compared to the older generation. A similar trend is visible with AMD’s Ryzen series, which, since its introduction in 2017, first caught-up with and then even outperformed Intels top-class products.
We can hope that with the recent increase in competition in the CPU market, e.g., Apple’s M1’s astonishing performance, that we continue to see even higher growth in single-threaded processing power.