I've heard a lot of peaple saying that moore's law (duplicating transistor in a computer chip every 2 years) is dead. So I decided to do a little research and found that data shows us that Moore's law is still alive and well. Even though we are reaching limits in other areas.
Moore's law dead is often used as an argument for searching alternates types of computing like Quantum or Neuromorphic, even though i really love the concept of new and innovative types of computing and they may have useful applications in the future, saying that progress in the digital realm has reached its limits it may be a bit extreme.
Posted by Gabster_68
14 Comments
Oooo now do this correctly. Look at the physical size of a cpu now compared to an old one. Because they are growing.
It’s suppose be transistor density, transistors per square mm is a better metric, obviously just making larger chips increases transistor count.
Edit Actually I guess I’m wrong, Moore’s doesn’t mention density, Dennard Scaling is about density and its what we all notices stopped working in the 2010’s. Of course there are real limits on size you can’t just stick larger and larger chips in your phone.
Ohh, I looked onto it for too long wondering how the first processors could only have, like 5 transistors, until I read the “(thousands)”
Single thread performance is what you should look at, it’s a much better measure of “pure” tech advancement and it is leveling off pretty substantially.
Widespread adoption of 6+ core processors over the last decade and the software tricks to take advantage of those extra cores have continued the steady progress, but it’s not the same kind of free lunch that Moore’s law originally described, it is a qualitative difference in the type of improvements we’re getting.
At first I thought…did they eliminate the channel size limitation or find a way to stop the quantum interaction of electrons? Did they figure out a way to stop finFET leaking.
Then I realized it’s just the number of transistors, not transistors per given area and nothing about cost. Chips are getting bigger, this doesn’t mean Moores law isn’t dead or dying its last legs.
BTW, the 18A chip from Intel does fix the finFET leaking problem with a GAAFET (gate all around FET) which is pretty cool.
data is 5 years old. Sources are [https://github.com/karlrupp/microprocessor-trend-data](https://github.com/karlrupp/microprocessor-trend-data)
ACM communications also had a tombstone with Moore’s law on the cover some years back and some interesting article [https://dl.acm.org/doi/epdf/10.1145/2976758](https://dl.acm.org/doi/epdf/10.1145/2976758)
Dennard scaling already ended and so will Moore’s law. There are all S curves.
But we are getting those transistor counts by gluing several CPUs together, so that’s very misleading. As the other comments say, Dennard scaling is dead.
I expect the Problem will be in 10 years, as you cannot scale down the transistor to be much smaller anymore. Even now the chiplet design means you don’t have one chip anymore as in the past so it’s not same in my view, as Moore’s law is transistor count on chip….will
Moor’s Law isn’t dead, it got bigger.
It’s nearly dead. I also think it’s partly what is leading to the AI boom. Hardware is no longer depreciating like it’s own fire so you can buy CPU/GPUs and in theory use them for 10 years or more.
There is a new tech that allows for smaller transistor construction, and the method of generating the light for the photolithography is mind-blowing. We’re now shooting lasers at droplets of liquid tin to vaporize them, generating the light. Not just one shot, either. It takes two shots to fully vaporize the tin–the first shot flattens the droplet into a disc, and the second shot vaporizes it to generate the EUV light.
It’s incredible.
I remember at university I saw a really neat diagram showing the distance a signal can travel between each clock cycle. It basically had circles of various sizes for various clock speeds, and somewhere between 5 and 10 ghz the circle becomes smaller than the cpu die
I think the diagram was implying that at a certain speed the signal can’t propagate through the cpu before the next one starts. It was a cool diagram but I can’t find it anymore
Anyway that might be a reason that clock speeds have plateaued, although realistically I think it has more to do with heat generation/dissipation
so the chart is until 2021… did I miss the last 5 years?
I’m actually surprised by the single thread performance increases since 2004. I would have expected something closer to double the performance since then rather than 10x.