TalkChess.com

Not really a big surprise for people following CPU chips for the last 10 years or so, but this is another reason us chess programmers need to work harder:

http://www.economist.com/technology-qua ... moores-law

Work harder ? You mean starting and stopping the automatic tuner ?

Great article Mark, thanks.

Thanks also for all your hard work on Komodo, looking forward to 9.4

Is Pentium E5200 with Komodo 9.4 able to beat Carlsen or Nakamura

That CPU is from Q3'08.

Let me put it this way: which CPU from the 2000s decade is needed to beat reigning world champion.

Plenty of interesting ideas in that article. Thanks.

mjlef wrote:Not really a big surprise for people following CPU chips for the last 10 years or so, but this is another reason us chess programmers need to work harder:

http://www.economist.com/technology-qua ... moores-law

The real surprise is that they can't quote moose's law correctly. It does NOT say that speed doubles every two years. Never did.

It says that the number of transistors in a given area doubles about every two years. Which is NOT the same as doubling the speed.

bob wrote:

mjlef wrote:Not really a big surprise for people following CPU chips for the last 10 years or so, but this is another reason us chess programmers need to work harder:

http://www.economist.com/technology-qua ... moores-law

The real surprise is that they can't quote moose's law correctly. It does NOT say that speed doubles every two years. Never did.

It says that the number of transistors in a given area doubles about every two years. Which is NOT the same as doubling the speed.

+ at the lowest cost per component.

"The density of the lowest cost components doubles every 2 years." Stopped being true after the 28nm technology node.

Scaling capability by itself is not a criterium for the "Law". We can scale much further if there is no cost constraint, which is what happens in R&D and other special applications. And even at consumer level below "28nm" we scale at equal or even slightly higher cost per transistor just for improved power performance, which is valuable for phones and data centers. Both "minimal cost" and "every 2 years" don't hold anymore.

mvk wrote:

bob wrote:

mjlef wrote:Not really a big surprise for people following CPU chips for the last 10 years or so, but this is another reason us chess programmers need to work harder:

http://www.economist.com/technology-qua ... moores-law

The real surprise is that they can't quote moose's law correctly. It does NOT say that speed doubles every two years. Never did.

It says that the number of transistors in a given area doubles about every two years. Which is NOT the same as doubling the speed.

+ at the lowest cost per component.

"The density of the lowest cost components doubles every 2 years." Stopped being true after the 28nm technology node.

Scaling capability by itself is not a criterium for the "Law". We can scale much further if there is no cost constraint, which is what happens in R&D and other special applications. And even at consumer level below "28nm" we scale at equal or even slightly higher cost per transistor just for improved power performance, which is valuable for phones and data centers. Both "minimal cost" and "every 2 years" don't hold anymore.

They haven't held for several years now. Probably 10 at least... But intel does continue to find architectural tricks that improve performance here and there even with no density improvements.

I tend to put the end point a bit later: in the 2012-2014 range, the era where high volume manufacturing of P1270 ramped up (intel's "22nm"), or at the foundries when the apple A8, snapdragon 810 and qualcomm MDM9235 rolled out. Up to the generation before it is quite reasonable to say it was still in place. I suppose that would put the real inflection point in the 2010-2012 range.

Some seem to be believe that the end of Moore's "law" also implies the end of technical progress, or consequently that any form of progress is evidence that Moore's "law" is still in place. Such is no doubt the effect of years of marketing trying to equate the two.

mvk wrote:I tend to put the end point a bit later: in the 2012-2014 range, the era where high volume manufacturing of P1270 ramped up (intel's "22nm"), or at the foundries when the apple A8, snapdragon 810 and qualcomm MDM9235 rolled out. Up to the generation before it is quite reasonable to say it was still in place. I suppose that would put the real inflection point in the 2010-2012 range.

Some seem to be believe that the end of Moore's "law" also implies the end of technical progress, or consequently that any form of progress is evidence that Moore's "law" is still in place. Such is no doubt the effect of years of marketing trying to equate the two.

It was always clear to many of us that density doubling could not continue indefinitely when one is using basic atomic structure for conductors. Years ago IBM proved that a conductor 1 atom wide was doable, but the experiment was a manual one where a human operator drug individual atoms and deposited them. Not exactly practical for making any sort of chip.

What happens next is unknown, but at least we are not at a "bad place" today if cycle times freeze at current levels.