4 ms·

> to asymptotically approach zero over the course of a few years Koomey's Law says compute costs will be ~1/16th / ~7% in a decade. IIUC, storage prices & ban

> to asymptotically approach zero over the course of a few years Koomey's Law says compute costs will be ~1/16th / ~7% in a decade. IIUC, storage prices & bandwidth reduce at slower rates. "Approaching 0" and "a few years" seem a little hyperbolic.

tomxor 4y ago
> Koomey's Law says compute costs will be ~1/16th / ~7% in a decade. Probably worse, Koomey's law is an observation based on the effects of Dennard scaling (constant power density while doubling transistor count due to scaling), and this pretty much ended over a decade ago due to no longer being possible to scale transistors in a uniform way. Power improvements will continue due to other one off process improvements, and due to other factors such as architectural improvements or moving more things into hardware, but the trend is not going to be as pronounced or consistent as it was before.
- tester756 4y ago
  Why focusing on transistors count then? Compare top CPUs from 2012 and current top CPUs I've opened cpubenchmark and compared those CPUs: AMD Ryzen Threadripper PRO 5995WX and Intel Core i7-3970X @ 3.50GHz (I didnt know how to find something better) First one has 98k points, second one has 8k points. It's giant gap. even pretty normal (consumer) CPU like Intel Core i9-13900K has 60k~ points.
  - tomxor 4y ago
    > Why focusing on transistors count then? I did not say transistor count, I said power density due to transistor scaling. There is a difference, transistor counts have continued to increase since 10 years ago for other reasons, but it no longer comes with the implicit power density and switching speed improvements of dennard scaling underlying the original moors law projection. It also doesn't have the same road ahead of it since transistors are no longer uniformly scaling, and dies can't grow exponentially. > First one has 98k points, second one has 8k points. Why do you think that is? What are "points", how have CPUs changed since 2012? and How did 2012 CPUs change since 2002? i.e what are the qualitative differences between these periods that are made opaque by "points", and which are key in indicating improvement projections.
coldtea 4y ago
Even 7% sounds very optimistic given the current climate. A couple of piled on downturns, and it could even be the same or higher...
YetAnotherNick 4y ago
But Koomey's Law say nothing for the cost. It just states energy usage for the same computation would be ~7% in a decade. There is also hardware acquisition cost or datacenter building cost, which is definitely not decreasing at that rate.

tomxor 4y ago

> Koomey's Law says compute costs will be ~1/16th / ~7% in a decade. Probably worse, Koomey's law is an observation based on the effects of Dennard scaling (constant power density while doubling transistor count due to scaling), and this pretty much ended over a decade ago due to no longer being possible to scale transistors in a uniform way. Power improvements will continue due to other one off process improvements, and due to other factors such as architectural improvements or moving more things into hardware, but the trend is not going to be as pronounced or consistent as it was before.

tester756 4y ago
Why focusing on transistors count then? Compare top CPUs from 2012 and current top CPUs I've opened cpubenchmark and compared those CPUs: AMD Ryzen Threadripper PRO 5995WX and Intel Core i7-3970X @ 3.50GHz (I didnt know how to find something better) First one has 98k points, second one has 8k points. It's giant gap. even pretty normal (consumer) CPU like Intel Core i9-13900K has 60k~ points.
- tomxor 4y ago
  > Why focusing on transistors count then? I did not say transistor count, I said power density due to transistor scaling. There is a difference, transistor counts have continued to increase since 10 years ago for other reasons, but it no longer comes with the implicit power density and switching speed improvements of dennard scaling underlying the original moors law projection. It also doesn't have the same road ahead of it since transistors are no longer uniformly scaling, and dies can't grow exponentially. > First one has 98k points, second one has 8k points. Why do you think that is? What are "points", how have CPUs changed since 2012? and How did 2012 CPUs change since 2002? i.e what are the qualitative differences between these periods that are made opaque by "points", and which are key in indicating improvement projections.

coldtea 4y ago

Even 7% sounds very optimistic given the current climate. A couple of piled on downturns, and it could even be the same or higher...

YetAnotherNick 4y ago

But Koomey's Law say nothing for the cost. It just states energy usage for the same computation would be ~7% in a decade. There is also hardware acquisition cost or datacenter building cost, which is definitely not decreasing at that rate.