Moore's Law Slowing is Counterbalanced by Other Developments

It is possible to argue that Moore’s Law (which suggests a doubling of transistor density about every 12 to 18 months, or about every two years in practice) has only slowed, and not stopped, since the mid-1960s. 

Time Period	Doubling Time (Years)
1965-1975	1
1975-2000	2
2000-2010	2-3
2010-Present	3+

On the other hand, consider the transistor densities we now have to “double.” As we push the boundaries of how closely together transistors and pathways can be spaced, it becomes more difficult to manufacture the chips. 

Year	Processor Model	Transistor Count	Doubling Time (Years)
1971	Intel 4004	2,300	-
1974	Intel 8080	6,000	3
1978	Intel 8086	29,000	4
1982	Intel 80286	134,000	4
1985	Intel 80386	275,000	3
1989	Intel 80486	1,200,000	4
1993	Pentium	3,100,000	4
1997	Pentium II	7,500,000	4
1999	Pentium III	9,500,000	2
2000	Pentium 4	42,000,000	1
2006	Core 2 Duo	291,000,000	6
2010	Core i7	1,170,000,000	4
2014	Core i7 (Haswell)	1,400,000,000	4
2018	Core i9	2,000,000,000	4
2022	Apple M1	16,000,000,000	4

Similar slowing can be seen in accelerator and graphics processing chips. 

Year	GPU Model	Transistor Count	Performance Improvement	Doubling Time (Years)
2012	Kepler (GK110)	7.1 billion	Baseline	-
2016	Pascal (GP100)	15.3 billion	2x	4
2018	Turing (TU102)	18.6 billion	1.2x	2
2020	Ampere (GA102)	28.3 billion	1.5x	2
2024	Blackwell (B200)	208 billion	30x	4

The other issue is that transistor counts are not the only important variables. Parallel processing is an architectural shift that prioritizes throughput over raw clock speed.

Accelerator chips are designed for specific tasks like AI or video processing and their task-specific metrics arguably are more important than simple clock speed.

Heterogeneous computing combines CPUs, GPUs, and accelerators for optimal performance across different workloads, meaning overall system performance is more relevant than individual component speeds.