The Fu­ture of Moore’s Law,

Should we start get­ting used to a fu­ture of pro­ces­sor stag­na­tion, with no ex­pec­ta­tion of speed im­prove­ments? Mike Bed­ford in­ves­ti­gates

Linux User & Developer - - Welcome -

er­haps it seems like a very long time ago that we first started hear­ing about the im­pend­ing end of Moore’s Law; and in fact, a quick search re­veals such pre­dic­tions go­ing back to 1997. We’ll see later just what Moore’s Law does and doesn’t say, but the sen­ti­ment of these pre­dic­tions was that a time will come when the re­lent­less im­prove­ments in pro­ces­sor per­for­mance that we’ve grown to ex­pect will come to an end. And when that time comes, we’ll be faced with a stark re­al­i­sa­tion: this is as good as it gets.

Well, it seems you might have to start com­ing to terms with that out­look sooner rather than later be­cause, ac­cord­ing to some in­dus­try ex­perts, Moore’s Law has al­ready come to an end. We did hope to get In­tel’s view on this ques­tion; af­ter all, Gor­don Moore, who gave his name to the law, was one of In­tel’s found­ing fa­thers.

But while, his­tor­i­cally, In­tel has tended to wax lyri­cal about Moore’s Law, on this oc­ca­sion, the com­pany de­clined to an­swer our ques­tions on the sub­ject. We’ll leave you to come to your own con­clu­sions about this. In the mean­time, how­ever, we’ll present our own analysis of the fu­ture of Moore’s Law and come to a view as to whether the long-stand­ing mer­chants of doom are, for once, cor­rect in pre­dict­ing the end of an era.

What is Moore’s Law?

Moore’s Law has com­monly been mis­quoted to say, for ex­am­ple, that it pre­dicts a dou­bling in pro­ces­sor per­for­mance ev­ery cou­ple of years. While this might not be too far from the mark, this pre­dic­tion would be an im­pli­ca­tion of Moore’s Law, as op­posed to the sub­stance of the law. In fact, when Gor­don Moore made his fa­mous state­ment, the first mi­cro­pro­ces­sor was still six years away, the semi­con­duc­tor in­dus­try was mak­ing dis­crete logic de­vices, and In­tel was about to launch the world’s first mem­ory chips.

A 1965 ar­ti­cle by Gor­don Moore, pub­lished in the in­dus­try mag­a­zine Elec­tron­ics, sug­gested that “the com­plex­ity for min­i­mum com­po­nent costs has in­creased at a rate of roughly a fac­tor of two per year”. He fol­lowed this by sug­gest­ing that “cer­tainly over the short term this rate can be ex­pected to con­tinue, if not to in­crease. Over the longer term, the rate of in­crease is a bit more un­cer­tain, al­though there is no rea­son to be­lieve it will not re­main nearly con­stant for at least ten years. That means by 1975, the num­ber of com­po­nents per in­te­grated cir­cuit for min­i­mum cost will be 65,000.”

If we re­move the in­dus­try jar­gon, Moore was say­ing that the num­ber of tran­sis­tors in main­stream chips had dou­bled ev­ery year and he pre­dicted that trend would hold true for at least ten years. He later amended this pre­dic­tion to a dou­bling ev­ery two years, and that ten-year pre­dic­tion has been ex­tended over and over.

Al­though they didn’t ex­ist back in 1965, to­day it’s rea­son­able to ap­ply Moore’s Law to mi­cro­pro­ces­sors. The graph (start­ing p66) shows

the num­ber of tran­sis­tors per de­vice from the first ever mi­cro­pro­ces­sor, the In­tel 4004, to the lat­est and great­est of a cou­ple of years ago. We’ve con­cen­trated on In­tel chips, but that’s not un­rea­son­able since we’re dis­cussing an In­tel pre­dic­tion. Ex­tend­ing it to the cur­rent date re­quires a de­gree of spec­u­la­tion, be­cause In­tel no longer quotes tran­sis­tor counts. The graph has a log­a­rith­mic ver­ti­cal axis which means that any ex­po­nen­tial growth – that is a dou­bling in a fixed pe­riod of time – ap­pears as a straight line. The straight line shows the trend that would be ex­pected if that dou­bling oc­curs ev­ery two years as pre­dicted by Moore’s Law.

You’ll no­tice that in re­cent years most new prod­ucts fell be­low that line, even though some of the more re­cent chips shown are high core-count, top-end Xeon chips. You might also no­tice that chips fell be­hind from around 1992 but, in time, the long-term trend was reestab­lished. Naively we might as­sume that the cur­rent be­low-par per­for­mance might be an­other short-term glitch, but there’s more to this than meets the eye.

Tran­sis­tor scal­ing

While the graph of tran­sis­tor counts doesn’t seem to sug­gest that Moore’s Law has hit the buf­fers quite yet, or at least not so badly that it can’t re­cover, an­other trend does ap­pear to be in trou­ble. That trend is the con­tin­ual de­crease in a chip’s fea­ture size.

Moore was say­ing that the num­ber of tran­sis­tors in main­stream chips had dou­bled ev­ery year

If you take more than a pass­ing in­ter­est in the chips that drive your PCs, you can’t fail to have no­ticed that fea­ture sizes de­crease ev­ery few years, even if you’re not en­tirely sure what that means. In­deed, if you do strug­gle to fully un­der­stand the term, you wouldn’t be alone, be­cause this has be­come a mat­ter of fierce de­bate in re­cent years. This be­ing the case, let’s use a rather vague def­i­ni­tion and say that the fea­ture size is the size of in­di­vid­ual fea­tures in a mi­cro­pro­ces­sor – with­out get­ting em­broiled in whether that’s the size of a sin­gle tran­sis­tor or a com­po­nent part of a tran­sis­tor.

When the 4004 pro­ces­sor first hit the streets in 1971, the fea­ture size was 10 mi­crons – which, if we ex­press it in the unit more com­monly used to­day – is 10,000nm. At some­where be­tween a third and a tenth of the width of a hu­man hair, this was surely con­sid­ered quite an achieve­ment al­most 50 years ago, even if it ap­pears pos­i­tively huge by to­day’s stan­dards.

Since then, fea­ture sizes have tum­bled, drop­ping to 1,000nm by 1985, and they have, more re­cently, re­duced by a fac­tor of around 1.4 to 1.5 per gen­er­a­tion, which equates to a halv­ing in terms of area. By and

large, to­day’s chips have a 14nm fea­ture size, and the tech­nol­ogy used to achieve that is re­ferred to as the 14nm process.

While a re­duc­tion in fea­ture size might have par­al­leled the in­crease in tran­sis­tor counts over the decades, it might not be im­me­di­ately ob­vi­ous that there’s a link be­tween the two trends. At one time, de­creases in fea­ture size fu­elled the in­crease in the clock fre­quency of a pro­ces­sor. So while the 4004 had a clock speed of 740kHz, by 1985 – when the fea­ture size hit 1,000nm – the high­est clock fre­quency achieved by the 80386 was 33MHz. By 2004, the 90nm process al­lowed the Pen­tium 4 to be clocked at 3.8GHz.

In the in­ter­ven­ing 14 years, clock speeds have barely in­creased be­yond this fig­ure; in­deed many of to­day’s best-sell­ers ac­tu­ally have lower clock speeds. Even es­o­teric chips pro­duced by bin­ning – se­lect­ing those de­vices that tests show are ca­pa­ble of ex­ceed­ing their de­sign speed – can man­age just 4.4GHz, with the op­tion to boost a sin­gle core to 5GHz, tem­per­a­ture and power con­sump­tion per­mit­ting. Had speeds con­tin­ued their pre­vi­ous trend, to­day’s best pro­ces­sor would be clocked at 100GHz but, of course, es­ca­lat­ing power con­sump­tion de­railed this par­tic­u­lar trend.

While clock speeds might have lev­elled out, fea­ture sizes have con­tin­ued to fall, for one very im­por­tant rea­son: it’s the driver of Moore’s Law. An ex­am­ple will help shed some light on this. The In­tel 4004 had 2,300 tran­sis­tors and the chip mea­sured 3 x 4mm. Had the fea­ture size not re­duced from 10,000nm, a high-end desk­top chip such as the forth­com­ing In­tel Core i9, with its es­ti­mated 7 bil­lion tran­sis­tors, would mea­sure around 5m by 7m.

Ad­mit­tedly, it’s rather ex­treme to con­sider a cur­rent pro­ces­sor built us­ing a 47-year-old process, even if it does lead to the bizarre pic­ture of a chip that wouldn’t come close to fit­ting in a desk­top PC, let alone a hand­held de­vice. How­ever, minia­tur­i­sa­tion is about

Newspapers in English

Newspapers from UK

© PressReader. All rights reserved.