View previous topic | View next topic

Moores law

Page 1 of 1

Robertmonk
1313366.  Tue Feb 12, 2019 1:22 pm Reply with quote

In the episode about a byte, kilobyte, etc, stephen says moores law increases memory every two years. I am fairly certain that moores law refers to CPU processing power. Which would be in hertz.

 
PDR
1313369.  Tue Feb 12, 2019 1:47 pm Reply with quote

Moore's original observation related to achievable transistor density (presumably in junctions per unit area) so neither would be strictly correct, although memory capacity per unit area and unit cost is probably more closely coupled to junction density (that's a guess based on my understanding of integrated circuit design parameters, not a claimed fact) than CPU processing power.

On a point of order - while CPU clock speed may be defined in Hz this is not a reliable indicator of CPU "power". These days disparate processor architectures make this more difficult to define (how much to allow for cached operations, multiple cores, raw clock speed, microcode reliance etc etc is open to debate), but it would more probably be defined in terms of something like MIPS (Millions of Instructions Per Second) or MegaFLOPS (Millions of Floating Point Operations Per Second).

PDR

 
PDR
1313432.  Wed Feb 13, 2019 9:08 am Reply with quote

OK, time to hold my hand up again because I didn't get this one quite right either - apols.

Having just spent lunchtime finding and reading Gordon Moore's original paper what he actually said is subtley different to what I said above.

His original point was that at any given state-of-the-art there would be a transistor density* which gave the lowest cost per junction. He then went on to predict that this optimum cost point would move with time such that every two years the optimum point would be at double the junction density.

This is similar to what I previously said, but it is not the same thing - Moore was clearly looking at practicable manufacturing points rather than the current limits of technology. ie he is looking at the density at which the yeild (the proportion of the produced dies that actually work) hasn't become too small to be viable. This is important because there are SOME devices which we currently use where the actual manufacturing yeild is below 20%, but the performance of the ones that do work is deemed worth the scrap wastage. Moore wast't interested in that sort of product.

PDR

* He just talks about number of junctions on a single die, but I infer that to be reasonably interchangeable with junctions per unit area because he deals with increasing die sizes separately.

 
dr.bob
1313438.  Wed Feb 13, 2019 10:04 am Reply with quote

PDR wrote:
On a point of order - while CPU clock speed may be defined in Hz this is not a reliable indicator of CPU "power". These days disparate processor architectures make this more difficult to define (how much to allow for cached operations, multiple cores, raw clock speed, microcode reliance etc etc is open to debate), but it would more probably be defined in terms of something like MIPS (Millions of Instructions Per Second) or MegaFLOPS (Millions of Floating Point Operations Per Second).


Depending, of course, on what you're doing with the device. If you're a scientist doing lots of calculations, then you'll mostly be interested in FLOPS. For most other people, MIPS is a more useful metric.

Unless you're using the device for graphics-heavy usage, such as gaming. In that case you're probably more interested in the number of graphics processing units (GPUs) that the device is equipped with.

Unless you're a scientist who wants to do lots of very simple but very parallelisable transactions, in which case you're probably better off buying a device with lots of GPUs and using them, rather than the CPU, to do the bulk of the calculations.

 
PDR
1313447.  Wed Feb 13, 2019 11:10 am Reply with quote

Absolutely - these days both hardware and software architecture tends to be tailored to the application (or class of applications) rather than just being a generic source of processing grunt, so comparison via a single metric is (as you point out) rather problematic. But one thing I suspect we'd agree on would be that CPU clock speed isn't a particularly good universal "computer power" metric.

PDR

 
dr.bob
1313539.  Thu Feb 14, 2019 10:48 am Reply with quote

Very true. There may have been a time when that was a reliable metric, but those days are long gone.

 

Page 1 of 1

All times are GMT - 5 Hours


Display posts from previous:   

Search Search Forums

Powered by phpBB © 2001, 2002 phpBB Group