The Moore’s law is over, in both its implication:
- keep doubling the number of transistors on a chip (in a given area) every 18 months
- keep decreasing the cost per transistor as density increases
The second implication ceased around 2014/2015 when the cost per transistor flattened out as density kept increasing to actually increase starting in 2016/2017. Now as we go to 7nm size for etching the cost per transistor is higher than it was (is) for a 10nm etching.
The first implication is also over in terms of being able to increase density twofold every 18 months. We are still increasing it a bit but it has reach the end of the line with present technology based on silicon. A boost may come by moving to monolayer materials, like graphene, but we are not there yet and an industrial, affordable production is not in sight.
However, at system level, we are seeing that the increase in performance keeps going on and in some cases it is faster than what Moore’s law would entail.
In his keynote in March 2018 -a year ago-, the CEO of Nvidia, Jensen Huang, remarked that in the last five years applications using graphic processor unitis have seen a 500x improvement in performance, and the chip itself has seen a 25x improvement whilst according to Moore’s law the improvement should have been around 10x faster.
He gave as example the training of a neural network using 15 million images. That took 6 full days of processing power 5 years ago whilst the same activity today it takes 18 minutes, a factor of 500x improvement.
This should not be surprising. When we apply the progress to our tools (in this case to chips) the progress we can derive by applying those tools may be smaller or greater, it depends on the way the tools are applied.
This is the case for the progress we have seen in the genome sequencing. The tools used were advancing at the rate of the Moore’s law but the sequencing advanced much more! At the application level there are several factors at play, As an example, in the case of the genome researchers have been able to benefit from the higher crunching power of microprocessors and at the same time of smarter software leading to the processing of longer strings. That made possible to shorten the time in sequencing and to lower the price with a multiplying factor that exceeds the evolution demanded by the Moore’s law.
The increase in the number of data, in storage capability, in transmission capacity (both locally and across networks) coupled with smarter software keeps increasing the application performances. We have learnt, as an example, to parallelise much more than in the past and this is introducing a multiplying factor in the progress.
So we can rest assured that even if Moore’s law has reached the end of the line the speed of evolution is not slowing down, actually in several areas it is accelerating.