How many CPU cores do you need?

IS A DUAL-CORE CPU SUF­FI­CIENT TH­ESE DAYS? OR IS QUAD-CORE OVERKILL FOR EV­ERY­DAY COM­PUT­ING? DARIEN GRAHAMSMITH TACK­LES A THORNY QUES­TION

PC & Tech Authority - - CONTENTS -

We put it to the test through a va­ri­ety of sce­nar­ios ...........................................................

How many CPU cores do you need? It’s a ques­tion with no easy an­swer. Even the cheap­est Celeron CPUs of­fer at least two cores, and for a light­weight lap­top that may be ab­so­lutely fine. At the other end of the scale, if you’re work­ing on a high-pres­sure data-anal­y­sis project where time is money, more cores are al­ways bet­ter.

For most of us, though, it’s not al­ways clear whether the ben­e­fit of hav­ing ex­tra cores will be worth the cost. A desk­top Core i3 pro­ces­sor with two phys­i­cal cores (plus two vir­tual cores, cour­tesy of In­tel’s Hyper­Thread­ing tech­nol­ogy) costs around $160 (i3 7100), while you’ll pay at least twice that for a quad-core Core i7, or an eight-core AMD Ryzen 7 chip. And if you move up into “Ex­treme” ter­ri­tory, you can pay well over a grand for a mon­ster such as In­tel’s 18-core Core i9-7980XE. It’s hard to know where to draw the line.

Don’t give up hope, though. We may not be able to tell you how many cores you per­son­ally need, but we can dig into the ben­e­fits of ex­tra cores, and of Hyper­Thread­ing tech­nol­ogy, to help you fig­ure out whether it’s worth in­vest­ing in an eight­core beast, or whether a nim­ble lit­tle du­al­core pro­ces­sor is all you need.

WHAT WE’RE TEST­ING FOR

For the pur­poses of this ex­er­cise, I ran a se­ries of bench­mark tests on a known sys­tem, vary­ing the num­ber of CPU cores avail­able while leav­ing ev­ery­thing else ex­actly the same. This isn’t di­rectly rep­re­sen­ta­tive of the range of CPUs in the mar­ket: dual-core models typ­i­cally come with slower clock speeds and smaller on-board caches than quad-core models. How­ever, this ap­proach does iso­late the ben­e­fit of more cores, which is nor­mally ob­scured by those other vari­ables.

Hap­pily, the test is very straight­for­ward to set up on a mod­ern PC: most BIOS in­ter­faces let you se­lec­tively dis­able in­di­vid­ual cores, and turn Hyper-Thread­ing on and off›at will. So a Core i7 desk­top pro­ces­sor can be tested in sin­gle-core, dual-core, quad-core and vir­tual eight-core con­fig­u­ra­tions.

The spe­cific plat­form I chose for the ex­per­i­ment was, con­ve­niently enough, my own desk­top PC – a Win­dows 10 sys­tem based on a Core i7-3770K. That CPU is a few years old – it’s a 22nm Ivy Bridge model, which lacks some of the so­phis­ti­ca­tions of the cur­rent 14nm Kaby Lake se­ries – but

this isn’t a prob­lem. The ba­sic prin­ci­ple of multi-core, multi-threaded com­put­ing hasn’t changed across the gen­er­a­tions. Sim­ply put, all cores are elec­tron­i­cally iden­ti­cal, and Win­dows and the CPU work to­gether to as­sign threads to pro­cess­ing cores as needed. If mov­ing from two cores tofour makes a pro­gram run twice as fast on Ivy Bridge, you can ex­pect to see the same ben­e­fit on Kaby Lake, and all the gen­er­a­tions in-be­tween.

Bar­ring any ma­jor sur­prises, it should also hold true for fu­ture gen­er­a­tions. Con­ceiv­ably, AMD or In­tel might at some point switch to a het­ero­ge­neous ar­chi­tec­ture, sim­i­lar toARM’s “big.LIT­TLE” de­sign for smart­phone pro­ces­sors. This fea­tures sev­eral light­weight en­er­gy­ef­fi­cient cores to keep things tick­ing over, part­nered with heavy-duty, power-hun­gry sil­i­con that only wakes up when there’s hard work to be done. The switch would be a big up­heaval, though: we can be pretty con­fi­dent that Win­dows plat­forms will be stick­ing with a ho­mo­ge­neous de­sign for the fore­see­able fu­ture.

With the hard­ware in place, I chose three tests to ex­plore how core avail­abil­ity im­pacts per­for­mance. The first was sim­ple: I con­fig­ured the Win­dows Per­for­mance Mon­i­tor to track to­tal CPU us­age for the first five min­utes af­ter sys­tem startup. Dur­ing this time I didn’t in­ter­act with the com­puter at all; the idea was to gain abase­line pic­ture of how much pro­cess­ing power is eaten up by startup tasks and back­ground ac­tiv­ity.

Af­ter that, I ran the stan­dard im­ageed­it­ing, video-edit­ing and mul­ti­task­ing bench­marks that we use in re­view­ing all PCs and lap­tops. And fi­nally, for some ex­tra in­sights, I turned to PCMark 10, the lat­est edi­tion of the ven­er­a­ble bench­mark suite.

TEST 1 STARTUP AND CPU UTIL­I­SA­TION

For bet­ter or worse, Win­dows keeps alarge num­ber of pro­cesses and ser­vices tick­ing over in the back­ground. When you have lots of cores to play with, though, this doesn’t no­tice­ably in­ter­fere with youruser ex­pe­ri­ence.

Even so, th­ese pro­cesses all con­sume re­sources. That was brought home to me very quickly when, for myfirst test, I switched my reg­u­lar desk­top PC down to a sin­gle-core, sin­gle-threaded con­fig­u­ra­tion. Af­ter re­boot­ing, the Win­dows desk­top ap­peared fairly quickly – but the sys­tem it­self was com­pletely un­re­spon­sive. I was able to move the mouse around, but click­ing and press­ing keys did noth­ing.

When fi­nally I man­aged to open theTask Man­ager, the ex­pla­na­tion was clear: all of Win­dows’ pro­cesses were try­ing to start at once, and the CPU was jammed at 100%. As the Per­for­mance Mon­i­tor traces right

in­di­cate, with no in­ter­ac­tion from me at all it still took more than two min­utes for things to calm down. Even then the CPU still hov­ered above the 50% mark: if you tried to get any work done on a sin­gle-core sys­tem your ap­pli­ca­tions would al­ways be fight­ing for re­sources.

En­abling Hyper-Thread­ing helped. With a sec­ond vir­tual CPU on hand, the log­jam cleared in around half the time – al­though that still meant a minute of thumbtwid­dling be­fore the com­puter be­came us­able. Af­ter that, Win­dows re­ported much lower on­go­ing CPU us­age, al­though this must be taken with a pinch of salt: Win­dows treats vir­tual cores the same as real ones, mean­ing if you were to com­pletely ham­mer the core with a heavy sin­gle-threaded work­load it would show as only 50% util­i­sa­tion, even though there would be lit­tle or no ca­pac­ity for Hyper-Thread­ing.

The dif­fer­ence be­tween real and vir­tual and cores is neatly il­lus­trated by our du­al­core, non-Hyper-Threaded CPU trace. If Hyper-Thread­ing were as good as a sec­ond phys­i­cal core then you’d ex­pect to see iden­ti­cal per­for­mance here; in fact, CPU us­age drops more quickly, and set­tles at a lower level.

That’s not to say that Hyper-Thread­ing is of no value. In dual-core, Hyper-Threaded mode, the CPU was fi­nally able to keep up with all of Win­dows’ startup tasks: pro­ces­sor util­i­sa­tion only mo­men­tar­ily hit 100%, in­di­cat­ing that the sys­tem was al­most fully re­spon­sive as soon as the desk­top ap­peared. We would call this a min­i­mum us­able spec, and it’s prob­a­bly no co­in­ci­dence that a two-core, four-thread con­fig­u­ra­tion is the stan­dard for Core i3 pro­ces­sors.

Is there much to be gained from mov­ing up to a quad-core chip? My test sys­tem didn’t com­plete its startup tasks any more quickly with four phys­i­cal cores than with two – as the graphs show, in both cases CPU ac­tiv­ity fell to back­ground lev­els at around the 90-sec­ond mark. The dif­fer­ence now was more con­sis­tent head­room: aside from an early peak, just af­ter the desk­top ap­pears, there’s al­ways at least an en­tire core’s worth of pro­cess­ing power sit­ting idle, which means that you can open ap­pli­ca­tions and start work with­out even re­ally notic­ing that other stuff is go­ing on in the back­ground. This is the con­fig­u­ra­tion that desk­top Core i5 pro­ces­sors use, while Core i7 models add Hyper-Thread­ing for a to­tal of eight log­i­cal cores. While the usual caveat about Hyper­Thread­ing ap­plies, in its full eight-thread mode the CPU never comes close to 100% load.

TEST 2 DESK­TOP BENCH­MARKS

The stan­dard set of PC & Tech Au­thor­ity desk­top bench­marks com­prises three tests. First, the im­age-edit­ing test uses the free ImageMag­ick tool to ap­ply var­i­ous colour and con­trast cor­rec­tions to a set of huge (35-megapixel) TIFF im­ages, be­fore ex­port­ing them as JPEGs. Then the video test uses Hand­brake to ap­ply var­i­ous trans­for­ma­tions to a 1080p movie file, ex­port­ing the re­sults as an H.264 video. Fi­nally, the mul­ti­task­ing test does both at once, while ad­di­tion­ally play­ing a 1080p video. The over­all score is a weighted av­er­age of the three.

As be­fore, I ran th­ese tests with one, two and four cores, with Hyper-Thread­ing both en­abled and dis­abled. The bench­mark re­sults are shown in the graphs op­po­site – and the clear in­di­ca­tion is that more cores mean higher per­for­mance, not just in the mul­ti­task­ing test but across the board.

A lit­tle thought re­veals what’s hap­pen­ing. It’s ob­vi­ous that the mul­ti­task­ing test will ben­e­fit from hav­ing at least three cores avail­able, since it’s run­ning three de­mand­ing pro­cesses at once. And thanks to our re­views of heavy­weight pro­ces­sors like AMD’s eight-core Ryzen 7 se­ries, we know that the ben­e­fits don’t stop there. Per­for­mance in this test keeps on get­ting bet­ter and bet­ter as you add more cores. Since the HD video play­back process is bound to a sin­gle core, the log­i­cal de­duc­tion is that one of the other two pro­cesses – the im­age test or the video test – must be tak­ing ad­van­tage of the ex­tra cores.

It’s no se­cret that this is the case. Step­ping up from a dual-core to a quad-core sys­tem yielded a mas­sive 91% per­for­mance bench­mark im­prove­ment in Hand­brake. What’s per­haps less well known is the fact that ImageMag­ick uses the OpenMP frame­work, al­low­ing it to di­vide up its work­loads across avail­able cores. Here the ben­e­fit is more mod­est: up­ping the

core count from two to four yielded a 25% im­prove­ment in this test.

Per­for­mance-wise, there­fore, it’s very much a case of the more cores the bet­ter, al­beit the mar­gin of vic­tory de­pends on the na­ture of the ap­pli­ca­tion you’re run­ning. For sin­gle- or lightly-threaded tasks, the switch from a dual-core to quad-core pro­ces­sor may de­liver a ben­e­fit that’s too small to re­ally no­tice. If you want to pum­mel your sil­i­con with mul­ti­ple heavy work­loads, the ben­e­fits are more clear-cut: in the mul­ti­task­ing test, my PC rock­eted from a score of 20 on a ba­sic dual-core sys­tem to 74 with all four cores en­abled – and up­ward to 89 with Hyper-Thread­ing ac­ti­vated. That’s a 445% per­for­mance im­prove­ment, which is cer­tainly not to be sniffed at.

One last note: while you can’t re­al­is­ti­cally run Win­dows on a sin­gle-core pro­ces­sor th­ese days, it was fas­ci­nat­ing to see how th­ese mod­ern bench­marks fared on such a sys­tem. The raw scores don’t ac­tu­ally look too bad; we’ve seen sim­i­lar per­for­mance from com­pact lit­tle Atom-pow­ered lap­tops. The catch is that the mul­ti­task­ing test wasn’t ex­actly run­ning as in­tended. Though the HD video player was run­ning, it was al­most com­pletely frozen, merely switch­ing from one bro­ken frame to an­other ev­ery few min­utes.

What’s more, turn­ing on Hyper-Thread­ing ac­tu­ally made things worse. While the im­age and video tests in­di­vid­u­ally gained aŽsmall ben­e­fit from the ex­tra vir­tual core,Žtry­ing to run ev­ery­thing at once brought the sys­tem to its knees, as our script tried to cram three de­mand­ing pro­cesses into two pro­ces­sor queues be­ing ser­viced by a sin­gle core. It ended up tak­ing around twice as long to run the test as with a sin­gle core. If you’re in­ter­ested, that’s five hours and 20 min­utes, com­pared to a run­time of 27 min­utes on the ex­act same hard­ware with all four cores en­abled.

TEST 3 PCMARK 10

Our desk­top bench­marks pro­vide a handy over­view of a sys­tem’s com­put­ing

“The clear in­di­ca­tion is that more cores mean higher per­for­mance, not just in the mul­ti­task­ing test but across the board”

ca­pa­bil­i­ties; for a more de­tailed look I switched to the pop­u­lar PCMark bench­mark, which sim­u­lates more than 40 typ­i­cal CPU-bound tasks – as well as a bunch ofŽgraph­i­cal tests.

In a few of th­ese tests, ex­tra cores made no dif­fer­ence what­so­ever to per­for­mance. Even in its min­i­mal con­fig­u­ra­tion, my CPU was able to keep up a smooth 30fps in the var­i­ous 1080p video play­back tests, and there’s noth­ing to be gained by throw­ing ad­di­tional re­sources at the task.

Very sim­ple tests such as “cut and paste” also saw no im­prove­ment; it’s hard to imag­ine how such a straight­for­ward oper­a­tion could ben­e­fit from ex­tra cores.

Un­ex­pect­edly, some of PCMark’s tax­ing “dig­i­tal con­tent cre­ation” tests were also un­af­fected by the num­ber of cores avail­able. Fur­ther in­ves­ti­ga­tion re­vealed that this mod­ule uses OpenCL, and hence th­ese tests were run­ning on the GPU rather than the CPU. The real sur­prise should haveŽbeen that other OpenCL-based tests did show a mod­est per­for­mance im­prove­ment as more cores were made avail­able. Even if you’re tar­get­ing the GPU, the CPU clearly still has an im­por­tant sup­port­ing role to play, en­sur­ing that GPU-ori­ented tasks are loaded and ex­e­cuted smoothly.

Those is­sues aside, I se­lected five PCMark tests to fo­cus on: the graphs above show re­sults, nor­malised so that 100 rep­re­sents per­for­mance on my quad-core CPU with Hyper­Thread­ing en­abled.

A few things stand out from th­ese graphs im­me­di­ately. For ex­am­ple, you will no­tice that there are no fig­ures for sin­gle-core, sin­gle-threaded oper­a­tion: with the CPU in this con­fig­u­ra­tion the bench­mark wouldn’t even open.

It’s also no­table that ev­ery one of th­ese tests shows a ben­e­fit in step­ping up from a dual-core Hyper-Threaded ar­range­ment (as used by a desk­top Core i3) to one with four phys­i­cal cores. Of course, this is partly due to the fact that I’ve cherry-picked the most in­ter­est­ing tests, but it was a ten­dency across the whole range of pro­duc­tiv­ity tests: on av­er­age, the Core i3 con­fig­u­ra­tion lagged 5% be­hind a quad-core con­fig­u­ra­tion and 7% be­hind a quad-core with Hyper­Thread­ing. The test that ben­e­fited least from ad­di­tional cores is “Copy data and

Com­pute 2”– a ba­sic Ex­cel oper­a­tion that you wouldn’t ex­pect to scale sig­nif­i­cantly. It was in­ter­est­ing to see that, with two cores, the oper­a­tion was ac­tu­ally slower with Hyper-Thread­ing en­abled than with it dis­abled. As with our mul­ti­task­ing bench­mark, In­tel’s vir­tual core tech­nol­ogy seemed merely to clog ev­ery­thing up. The dif­fer­ence was small, though – in real terms, the test took just a fifth of a sec­ond longer to ex­e­cute.

The task that ben­e­fited most was the “Un­sharp Mask 1” test – which shouldn’t be a sur­prise. Graph­i­cal op­er­a­tions are com­par­a­tively easy to dis­trib­ute across ar­bi­trary numbers of…threads, which is why pow­er­ful graph­ics cards use mas­sively par­al­lel ar­chi­tec­tures.

Even so, the “Video 2160p VP9” test gen­er­ated sur­pris­ing re­sults. This test plays a 4K video en­coded in Google’s roy­alty-free VP9 for­mat, and while my CPU breezed through PCMark’s var­i­ous 1080p video tests, it choked here, only man­ag­ing to achieve a full 30fps with all four cores en­abled. There’s a rea­son for this, though: the third-gen­er­a­tion Ivy Bridge CPU in­cludes spe­cialised hard­ware for de­cod­ing H.264 video, but VP9 has to be pro­cessed man­u­ally by the CPU. Sixth- and sev­enth-gen­er­a­tion pro­ces­sors – fa­mil­iarly known as Sky­lake and Kaby Lake – include hard­ware sup­port for VP9, so if your pro­ces­sor’s fairly up to date then you should be able to play 4K VP9 video with­out over­tax­ing the cores.

On that topic, it’s also worth not­ing that AMD Ryzen pro­ces­sors don’t have any sort of on-chip graph­ics pro­cess­ing hard­ware. How­ever, that means you will nec­es­sar­ily be us­ing a stand­alone graph­ics card, which will prob­a­bly include its own de­cod­ing en­gine.

Fail­ing that, since all Ryzen models have at least four cores, they should be…per­fectly ca­pa­ble of de­cod­ing 4K video “man­u­ally”.

CON­CLU­SION

With all of th­ese test re­sults to draw on, the is­sue of how many cores you need is prob­a­bly now as clear as mud. As we said at the out­set, it’s a ques­tion with no sim­ple, sin­gle an­swer.

Even so, we’ve high­lighted a few gen­eral prin­ci­ples that can help you choose your next sys­tem. It’s true that the ben­e­fits vary con­sid­er­ably from task to task – so, if you spend most of…your time word-pro­cess­ing and watch­ing videos, the stan­dard Core i3 con­fig­u­ra­tion hits a sweet spot of price and re­spon­sive­ness.

In gen­eral, though, hav­ing more cores does ben­e­fit per­for­mance – even…if you’re not run­ning soft­ware that specif­i­cally tar­gets a multi-threaded en­vi­ron­ment. Th­ese days, Win­dows it­self has so much go­ing on in the back­ground that pro­cess­ing head­room is of­ten at a pre­mium. And,…as with our im­age-edit­ing bench­mark, you may find that even pro­grams that don’t ob­vi­ously call for mul­ti­ple cores have been coded to gain some ad­van­tage.

What’s more, look­ing for­ward, it…seems cer­tain that devel­op­ers are only go­ing to rely more and more on multi-core re­sources. His­tory tells us that, as the spec­i­fi­ca­tions of the av­er­age com­puter creep up­wards, ap­pli­ca­tions get more de­mand­ing. With Ryzen bring­ing eight-core pro­cess­ing into the main­stream, it’s only a mat­ter of time be­fore dual-core pro­ces­sors are as ob­so­lete as sin­gle-core chips are to­day.

We’re not say­ing that you must buy as many cores as you can af­ford. But if you’re scout­ing out a new com­puter for a gen­eral desk­top role – even if you don’t need ex­treme mul­ti­task­ing ca­pa­bil­i­ties – in­vest­ing in four or more phys­i­cal cores could ex­tend its longevity con­sid­er­ably. Think of it that way and the ex­tra sil­i­con looks like a smart buy.

“Look­ing for­ward, it seems cer­tain that devel­op­ers are only go­ing to rely more and more on multi-core re­sources”

To put the cores through their paces, we ap­plied colour cor­rec­tions to a col­lec­tion of huge TIFF im­ages

Step­ping up from dual-core to quad-core yielded sharp re­turns in our video-edit­ing tests

Newspapers in English

Newspapers from Australia

© PressReader. All rights reserved.