"Moore's law is dead. It's completely over." Those are the words of NVIDIA CEO Jensen Huang explained at a recent press conference, addressing complaints about the high price of GPU chipsets.

Moore's law is well-known and refers to the observation made by Gordon Moore, one of Intel's co-founders, that the number of transistors on a microchip tends to double every 18 to 24 months.

Figure 1:

Huang's helpless comments came in the wake of chips failing to expectations for doubled performance every two years at the same cost level. This reflects that improving chip performance based on current techniques presents a major challenge for the industry.

Chip Processing Has Hit Its Physical Limit

In the 50 years since Moore's Law was first proposed, chip makers have managed to keep pace by adding cores, driver IC threads, and accelerators. The shrinking of chip size has tested the limits of what is physically possible.

This slowdown was first observed in 2000. By 2018, improvements to chip performance were 15 times weaker than those predicted by Moore's Law. It's an industry consensus that this discrepancy will only continue to grow.

Figure 2:

Dennard Scaling Provides a Solid Supporting Evidence

Robert Dennard's prediction, also known Dennard Scaling, provides a solid evidence in support of how Moore's Law works in practice. Dennard Scaling refers to the fact that as the power density of transistors is constant, each transistor uses less power as transistor density increases. This scaling effect means that chips will work at lower voltages and currents as they shrink, and therefore consume less power.

Dennard Scaling started to slow down in 2007 and almost disappeared in 2012. Under more advanced chip manufacturing processes, transistor gate length has decreased, but current leakage has increased as well. This has led to chips using more power, and producing more heat to be dissipated. It is transistor current leakage that contributes to the failure of Dennard Scaling.

Since transistors are no increasing within the same power budget, doubling transistors now results in doubling the amount of power that is consumed. The Dennard Scaling law no longer applies, as chips will consume more power in proportion to the increase in the number of cores.

Figure 3:

The failing Dennard Scaling effect and slowing Moore's law signal that better chip manufacturing will produce diminishing returns in terms of performance and energy saving.

When mobile devices with 4-nm chips hit the market in early 2022, many users complained about device overheating and quick battery discharging. This shows that incremental improvements in chip manufacturing can no longer deliver the desired power saving effects.

Better Chip Manufacturing Is a Dead End

According to estimates by the Global e-Sustainability Initiative (GeSI), ICT accounted for 2.3% of global carbon emissions in 2020. Chips are heavily used in electronic devices, and will play a crucial role in creating a green ICT sector. Are more advanced processes that much helpful for the industry to significantly bring down energy consumption?

For smartphones, intensive use requires frequent battery charging. Given the dominant use of chips in smartphones, better manufacturing processes have always been key to chip production. Let's use the A-series chips in Apple's iPhone devices as an example. The 7-nm A12 uses 50% less power than the 10-nm A11, the 5-nm A14 is 30% more power efficient than the 7-nm A13, and the 4-nm A16 slashes power consumption by 20% compared with the 5-nm A15. Due to these clearly diminishing returns, it's doubtful that future chip processes will produce substantial improvements.

This is also true for PCs, for which processor chips account for more than 30% of power use. Both Intel and AMD have turned to advanced processes: Intel 10 nm for the 12th gen i9-12900HX, and AMD 6 nm for its most powerful Ryzen 9 6900HX ever. However, i9-12900HX is better performing (24259 vs. 14013, CPU performance scores) and 24% more power efficient (194 vs. 156, energy efficiency) than the Ryzen 9 6900HX. So, more advanced processes do not always ensure better performance and energy efficiency.

5G base stations are another hot topic. Operators and vendors have sought to make them so energy efficient that they can be carbon neutral. Advanced processes are considered by many vendors as a tool to achieve this. However, base station power amplifiers (PAs) for converting DC power into radio waves are the biggest consumer of power (about 90% of the total). Moreover, after one generation of upgrade, chips only consume a small proportion, likely less than 1%. This shows that advanced processes will not have a significant effect.

Advanced Processes Are Just One of the Many Ways for Energy Saving

For smartphones, screens consume the largest amount of power. LTPO screens with a minimum refresh rate as low as 1 Hz are widespread on smartphones. This lower refresh rate will result in less power consumed. For PCs, power management is the crucial approach, whereas for 5G base stations, PA efficiency is the biggest factor. Currently, PA efficiency remains below 50%, leaving much room for improvement, while manufacturing processes are already hitting a physical limit.

Therefore, it is clear that even as Moore's Law dies a slow death, and advanced processes produce diminishing returns, there are still a number of ways that we can make electronic devices more power efficient.

This content is sponsored by C114.