In this blog post, we’ll explore the background behind the emergence of multi-core processors, the limits to performance improvements, the challenges facing parallel processing technology, and future directions for development.
The Emergence of Multi-Core Processors
If you look at recent computer advertisements, you’ll easily come across terms like dual-core, quad-core, hexa-core, and octa-core—collectively referred to as “multi-core.” Here, “core” refers to the central processing unit responsible for computation and control in a computer. In the past, since a CPU contained only a single processing unit, the terms “CPU” and “processor” were often used interchangeably. However, with the widespread adoption of multi-core processors, the term “CPU” today generally refers to a single chip containing multiple cores.
If multi-core technology is so superior, why wasn’t it widely used starting in the 1990s? To understand the reason, we need to examine the evolution of computer technology around 2005, when multi-core technology began to gain widespread adoption.
The key factors determining a computer’s performance are integration density and clock frequency. Integration density refers to how many circuits and transistors can be integrated onto a single chip, and it significantly influences the ability to perform complex functions and execute a wide range of instructions. Clock frequency, on the other hand, indicates how many clock cycles a CPU generates per second and is generally closely related to computational speed. The clock acts as a reference signal for the computer to perform computations; each time the clock cycles, the CPU processes one or more tasks. For example, a 3.0 GHz CPU generates approximately 3 billion clock signals per second. Of course, actual performance is influenced by various factors, such as how many instructions can be processed per clock cycle (IPC) and memory performance.
CPU clock speeds have increased rapidly since the 1990s. CPUs, which once operated at tens of MHz, moved through the hundreds of MHz range and entered the GHz era; until the early 2000s, increasing clock speed was considered the most effective way to improve performance. At the time, it was even anticipated that CPUs operating at 4.0 GHz or higher would soon appear.
However, by the mid-2000s, the direction of CPU development had shifted significantly. This was due to the emergence of the so-called “power wall” problem, where continuing to increase clock speeds led to a sharp rise in power consumption and heat generation. As clock speeds rise, more power is required, and the resulting heat also increases significantly. Once a certain threshold is exceeded, it becomes difficult to maintain stable operation with standard air cooling alone, necessitating complex cooling systems. However, such approaches significantly increase manufacturing costs and power consumption, making them economically unfeasible.
Ultimately, the semiconductor industry shifted its development strategy away from simply increasing clock speeds toward incorporating multiple cores within a single CPU to perform tasks simultaneously.
This is easy to understand if you compare it to building a house. There are clear limits to the approach of constantly demanding that a single carpenter work faster and faster. If you try to increase the work speed beyond a certain level, fatigue sets in and efficiency gradually declines. Therefore, rather than pushing a single carpenter to work at the absolute limit, it is far more efficient to have multiple carpenters share the workload simultaneously. This is the fundamental concept behind the emergence of multi-core processors.
Challenges of Multi-Core
If the concept of multi-core is so effective, why wasn’t it widely adopted sooner? It might seem that using multiple cores before heat issues became a serious problem would have yielded higher performance. However, in reality, adding more cores does not result in a proportional increase in performance equal to the number of cores. This is because a phenomenon similar to the law of diminishing marginal returns in economics also occurs in computing. Diminishing marginal returns refers to the phenomenon where the increase in output gained from adding more production factors gradually decreases as more factors are added.
Let’s revisit the carpenter analogy we discussed earlier. In theory, hiring one more carpenter of the same skill level would seem to cut the work time in half. However, this is not the case in a real-world setting. Just because two people work together on a task that would take one person two days does not necessarily mean it will be completed in a single day. Time is needed to coordinate their work, and there are steps where one person must finish before the other can continue. Therefore, the actual work time often ends up being about a day and a half rather than just one day.
The same applies to multi-core processors. While a dual-core processor has two cores performing tasks simultaneously, its actual performance rarely doubles. A quad-core processor does not always deliver four times the performance either, and the rate of performance improvement tends to gradually decrease as the number of cores increases. This is precisely what is meant by the common saying in the computer field: “1 + 1 does not equal 2.”
Why does this happen? It’s easy to understand if we revisit the example of building a house. Even if two carpenters are working together, they cannot perform every task simultaneously. The foundation must be completed before the walls can be erected, and the roof can only be installed after the walls are up. There are also times when they must share a single tool or wait for each other’s work to be finished. Ultimately, while some tasks can be performed simultaneously, others must proceed sequentially.
The same principle applies to computers. Tasks that multiple cores can process simultaneously are handled in parallel, but parts that require waiting for each other’s calculation results or that must be executed sequentially cannot be parallelized. Therefore, the execution time of the entire program is determined by the proportion of tasks that can be processed in parallel, a principle well known in computer science as Amdahl’s Law.
For example, let’s assume we are calculating the expression “(3+5) × (9+1) + 3” on a dual-core computer. The first core can calculate 3+5, and the second core can simultaneously calculate 9+1. However, after both calculations are complete, the computer must calculate 8×10, and finally, perform 80+3. In other words, while the initial additions can be processed simultaneously, the multiplication and the final addition can only be performed once the results of the preceding calculations are available. Ultimately, since some calculations must wait for other cores to finish their work, the total execution time is not reduced exactly by half. In this example, while a single-core processor requires four steps to complete the calculation, a dual-core processor requires about three steps.
Examining data illustrating the effects of parallel processing makes it easier to understand this characteristic. For a program with a parallelization ratio of 50%, there is a clear limit to performance improvement no matter how many cores are added. For example, even with 16 cores, the execution speed improves by only about twofold. Conversely, for a program with a very high parallelization ratio, the performance gain increases significantly as more cores are added. Ultimately, increasing the program’s own parallelization ratio is a far more critical factor for performance improvement than simply increasing the number of cores.
However, since humans are accustomed to thinking sequentially, designing parallel programs is extremely difficult. This is because it requires not only dividing tasks into multiple parts but also resolving various issues such as data sharing, synchronization, race conditions, and deadlocks. For these reasons, parallel programming and parallel algorithms remain one of the most actively researched fields in computer engineering today.
Nevertheless, parallel processing has fundamental limitations. No matter how sophisticated the program design, it is virtually impossible to completely eliminate sections that must be executed sequentially. Therefore, multicore technology may also eventually reach its limits in terms of performance improvement.
Consequently, the semiconductor industry is actively developing new computing technologies alongside research aimed at improving the efficiency of multicore processors. Various technologies—including chiplet architectures, 3D stacked semiconductors, AI-dedicated accelerators (NPUs), heterogeneous computing utilizing GPUs, quantum computers, and neuromorphic computing—are being researched as next-generation computing platforms. Just as the computer industry experienced rapid growth during the transition from vacuum tubes to transistors, new semiconductor technologies and computing architectures are expected to play a crucial role in overcoming current limitations in the future.
Summary
In summary, single-core processors have long improved performance by increasing clock speeds. However, as they faced a power barrier characterized by a sharp increase in power consumption and heat generation, the approach of simply increasing clock speed has reached its limits. Consequently, the multicore architecture—which incorporates multiple cores into a single CPU to process tasks simultaneously—has emerged as a new alternative.
However, having more cores does not necessarily result in a proportional increase in performance. Actual performance depends largely on how effectively a program can perform parallel processing. The more tasks that must be executed sequentially, the more the performance gain diminishes even when additional cores are added; this characteristic can be explained by Amdahl’s Law. Consequently, in today’s computing landscape, parallel programming techniques and software optimization are considered more critical factors for improving performance than simply increasing the number of cores.
In fact, since the mid-2000s, the direction of CPU development has shifted from a clock-speed race to a focus on multi-core and parallel processing. Recently, heterogeneous computing—which utilizes not only CPUs but also purpose-specific processors such as GPUs and NPUs—has become widespread, and software is also evolving to actively leverage these architectures. The reason users no longer experience dramatic performance improvements even when replacing their computers every few years, as they did in the past, is that we have entered an era where the performance of the entire system—including storage devices, memory, graphics processors, and AI accelerators—has a greater impact than simple CPU performance improvements alone.
The semiconductor industry is also exploring new directions for development in line with these changes. Whereas increasing clock speed was once the key to competitiveness, technology is now advancing toward improving power efficiency, enhancing parallel processing capabilities, and organically utilizing various computational devices. Furthermore, with the proliferation of AI computing, research is actively underway on new technologies such as chiplet architectures, 3D stacked semiconductors, next-generation memory, optical computing, and quantum computing.
Ultimately, the reason “1+1=2” does not always hold true in computers is that even when using multiple cores, not all tasks can be performed simultaneously. Some tasks must be processed sequentially, and time is also required for multiple cores to exchange data and synchronize with one another. Therefore, the performance of a multicore system is determined not simply by the number of cores, but by the degree of parallelization in the program and the level of hardware and software optimization. In the future, computer technology will continue to evolve not only by adding more cores but also by overcoming current limitations through new semiconductor technologies and innovative computing architectures.
