Ask HN:have moore's law slowed innovation in chip manufacturing?

https://news.ycombinator.com/item?id=7555205

by hershel • 12 years ago

4 8 12 years ago

news.ycombinator.com

There's this contrarian conversation in reddit about the possibility that moore's law has greatly slowed innovation in the chip manufacturing industry[1].

What does HN think about this ?

[1]http://www.reddit.com/r/programming/comments/229ggx/the_future_doesnt_have_to_be_incremental/cglalgg

wmf2• 12 years ago

If some company could have built 10x better transistors than the competition they could have made billions, if not tens of billions. Yet no company did this. The simplest explanation is that it wasn't possible.

hershel• 12 years ago

Technologically it does seem possible. But if they have built it, won't somebody else built similar factories, and than a race to the bottom would have begun and nobody would have made that much money , especially relative to the money they made with new generation every 18 months ?

wmf2• 12 years ago

It certainly would lead to a lot of risk if there wasn't a consensus roadmap; once company A announced 10x, company B might try to one-up them with 30x then fail, etc.

zackmorris2• 12 years ago

As far as I can tell, Moore's Law ended about 10 years ago, because 3 GHz CPUs were available then and most CPUs are slower than that now. Sure transistor counts have gone up, but the amount of computation that can be done in a single thread hasn't seen the 100x speedup that it did in previous decades. Mostly there's only been cheating by adding pipeline stages or ram latencies to get the frequencies up.

Also just before I graduated college in 1999 I remember my VLSI professors talking about how chips had passed the point where the interconnect had a square cross section and were moving to rectangular. So today chips look like skyscrapers, with tall ribbons of interconnect that suffer from crosstalk.

On top of that, there's been failure to move to distributed and multicore processing outside of GPUs, or for FPGAs to get adopted by the mainstream. I've given up on innovation on the hardware front from the big players, but I'm optimistic that approaches like Google's Go running inside hypervisors on diverse hardware is going to bring the high throughput we see in gaming and DSP to the mainstream. I’m unimpressed by OpenCL and CUDA and won’t really consider parallel computation as having arrived until more readable languages like Python and MATLAB generate accelerated code for us.

I realize there are holes in my statements here that one could fly a 747 through, but when the single fastest increase in my computing in a decade came from installing a 512 GB SSD drive, it means that something went terribly wrong. Moore’s Law ending could have more to do with computers reaching a point of being “good enough”, essentially becoming disposable appliances, than limitations with technology or cleverness. So they may not be getting any faster, but they are becoming so cheap that I think we’ll see life change in rather interesting ways in coming years.

Perdition2• 12 years ago

Moore's law is only about transistor density, the clock speed race of the 90's mirrored Moore's law but was only indirectly related (in that the smaller you make a circuit the less time a signal needs to propagate through it).

The improvement in GPUs is all in line with Moore's law.

On top of that, there's been failure to move to distributed and multicore processing outside of GPUs

Few ordinary tasks besides graphics processing benefit significant from multi-threading, and the resulting program complexity means developers don't bother with it for minor improvements.

There are other benefits from the power of modern processors, such as running programs in sandboxes to improve security, and being able to run multiple CPU intense programs on separate cores.

but when the single fastest increase in my computing in a decade came from installing a 512 GB SSD drive, it means that something went terribly wrong.

No it doesn't, it is just that SSDs cut disk access speeds more in absolute terms than going from 1Ghz to 4Ghz improves program speed. A modern 4Ghz CPU is more than 4 times faster than a 1Ghz early 2000's CPU, but in absolute terms the speed up is small (ie 500 ms to 50 ms is more noticeable than 50 ms to 5 ms).

Try running a CPU bound task (the ordinary user doesn't have these) on a early 2000's CPU compared to a modern CPU.

zackmorris2• 12 years ago

Ya I actually completely agree with everything you said, my main complaint though is that the difference between my 16 MHz LC II in the early 90s and my blue 333 Mhz G3 iMac was almost two orders of magnitude, whereas my 2.3 GHz i5 Mac Mini only feels say 5 or 10 times faster than my iMac did. The question I ask myself is if Moore's law can no longer lead to faster computing, then what good is it?

For parallelism, I have lost hope that there will be much progress in the compiler automagically parallelizing things like for loops in C++. There's some cool stuff with SSE etc, but it all feels fairly ineffectual to me. Probably people will just move to a new language like Go or Haskell/Scala and just use C for glue code. I could maybe see something like Groovy for low level languages though, where it makes global variables illegal and does some other things to eliminate side effects and then compiles to, say, Go for the 10/100/1000x speedups we'd expect for things like graphics or matrix math.

mud_dauber• 12 years ago

Moore's Law hasn't slowed innovation. The cost of fab toolsets and the cost of chip design are the culprits.

Many cutting-edge tools (etchers, deposition, others) now come with price tags of >$5M. A new 200mm fab can easily run north of $4B. The relentless trend of feature size shrinkage is still happening - some designs are underway using 16nm FinFet technology - but at the cost of increased power consumption and increased cost-per-transistor. I believe it was Nvidia that put a shot across the bow of chip manufacturers by saying they saw no economic advantage in chasing the lastest process node.

A clean-sheet-of-paper microprocessor design can easily hit $50M in expenses before the first chip is seen. Much of that expense is sheer manpower & verification processing cycles - a billion transistors' worth of real estate is worthless unless you can run a test suite against it. If you're spending >$50M on a new design, you'd better have a locked-in customer. Projecting that amount of investment for a squishy market like a new consumer electronics widge is the surest way to get your project killed.

Running a microprocessor at ever-greater frequencies doesn't help either. First, you're cooking your chip. Frequency = heat. Second, every processor vendor in the world is limited by the processor-memory chokepoint. You can't read data & code from memory into the processor(s) fast enough to keep the processor pipeline full. And the gap is widening.

Ask HN:have moore's law slowed innovation in chip manufacturing?

Related Stories