September 29, 2020
Semiconductor chip companies cannot avoid the topic of Moore’s Law. This law was put forward by Intel co-founder Gordon Moore in 1965 and has guided the development of the semiconductor industry for more than 50 years-every 2 years the transistor density of the chip will be Double the performance and double the performance.
However, after entering the 10nm and below process, Moore's Law has always been regarded as invalid. Chip manufacturing has become more and more difficult, and the cost has become higher and higher. It is difficult to double the performance and reduce the cost at the same time. However, on this issue, Intel has always firmly defended Moore's Law and does not recognize the issue of failure.
AMD CEO Su Zifeng attended the SEMICON West conference a few days ago and delivered a keynote speech. She mentioned the issue of semiconductor technology and Moore’s Law. AMD’s point of view is very much in line with Intel’s idea that Moore’s Law will continue to be useful. But it has slowed down.
Su Zifeng took CPU and GPU as examples, and pointed out that the performance of CPU and GPU will double every 2.5 years in the past 10 years.
In the performance improvement of the past ten years, the treatment process has achieved a significant performance improvement, with energy efficiency doubling every 3.6 years and density doubling every 3 years.
Specifically, in the performance improvement of the chip in the past ten years, 40% of the factor is due to the processing technology, 8% is from the improved TDP, 12% is from the additional core area, 17% is from the architecture improvement, and 15% is from the Power management, and 8% comes from compiler improvements.
In summary, AMD believes that Moore's Law is still valid, but the speed has slowed down. From the roadmap given by AMD, it has slowed down after the 22nm node, especially after the 14nm node.
However, in terms of chip cost, the result is different. The more advanced the process, the more expensive the cost, and this trend is becoming more and more obvious.
According to AMD's data, with the 250mm2 core, the cost of the 45nm node is counted as 1, and the 32nm and 28nm nodes start to increase, and the 20nm node becomes twice the cost, and the cost of 7nm jumps to 4 times, and the future 5nm is even more exaggerated. , The cost will be 5 times the previous.