The program´s results can be almost practically directly compared to other PC setups regardless the chipset/Mobo/Cpu or type of Memory & drivers/OS used to see the performance differences between them! ( Bios [revisions] applied on MoBo´s (or badly programmed sh*itty ones) may have some impact on final results just like detailed Mem & some other settings there! ) Benching should be done in Safe mode (or at last pressing Shift key on a keyboard while Win is booting ) . Also any power saving settings in Bios should be disabled and not running stuff like that too [Cool´n´Quiet for example], to prevent ram frequency change during testing! ... I hope I do not have to mention that system if OC-ed, has to be completelly Rock Solid Stable in any circumstances ... ( my way of testing here) ! Explanation how this bench works is at the bottom of the article! The benching result is impacted mostly by a mem.sub-system performance & (very) little by other system settings & with these days powerful setups, CPUs clock has almost negligible bench impact; especially with 1GHz/100Fsb/mem & upper ones ...

And how does better memory performance impacts on better PC performance? A bit! [for example: 50 % better result with this bench [mem_index 1.5], speeds the whole system up for 10% in average comparing to another same frequency or similar performance CPU, if run Mem in Sync with Fsb; if not only for 8% approx., so for to say so a 2GHz machine in average can perform just like a 2,2G one in every-days usual work with it!] .

The math formula for such system performance boost in % could be like this: calculated / compared mem_index twice square rooted [or 4th Root of it] minus 1 multiplied with 100 ; for upper assumption follows like this : {4th Root of 1,50 - 1} x 100 = 10,6 % !}. Simple as that! And how did I calculate the yellow column results ? Simple by dividing the final benching result (aqua colored column -2nd number) with real memory clock (1st number) & than multiplying with 100 [for 100MHz result] !

The result depends practically almost proportional to true memory clock used (frequency) of it on the same machine & practically almost does NOT depend of CPU clock-frequency or Ram type used. For AMDs latest CPUs sometimes is not so easy to determine that true memory clock, so you better visit & read before certain forums, for example thread at XtremeSystems and maybe use Cpu-Z program too, to be shure about that clock (well, IMHO Cpu-Z is not reliable showing that true mem.clock for "K8" platform CALCULATION & Bios numbers also not, as some "my" tests revealed!). ... { IMHO, sometimes is good not to trust blindly even to certain statements inside manufacturers datasheets & white papers; not even speaking of about certain statements on some web_sites/forums/UseNet, (including some mines past ones) ... } ! ... Because of problems determining real mem.clocks with "K8" family CPUs when runnig it out of defaults, mainly those results should be revisited giving around IMHO 10% better (yellow column) results than shown downthere !!! ... Maybe we have a new memory speed winner ... and mine result should be thrown off the "throne"

This Benchmark does not depend on theoretical Mem bandwith at all (like SiSoft Sandra does - good for IT marketing purposes!), but can be used also to see differences while Tweaking Bios setting & if knowing how (having an appropriate tool-program) even hidden ones of the chipsets registers! (the success on tweaking machine can be seen with Sandra too).

Standard memory response can be worse or better even more than -/+10% than normal when playing with memory/chipset settings & latencies on the same FSB/mem clock while tweaking mem-subsystem. Lower the latencies, better ram sticks & Northbridge design, better the results normally! Also higher mem clock, better the result - logical, isn´t it? But take care, if the frequencies & tweaking is pushed too far over "sweet spot", you may be experiencing some degradation in linearity of rising the results even before PC crashes or Bench program quits with illegal operation, because the machine is struggling to make it (see OC try on RDram machine down there! - serial technology does not like pushing it to far...)

So I proposed some testing to some friends on the Net to do the job for me mainly with results as follows:

Tester:	Ram_MHz_real_clock	MoBo or Chipset or /&	100MHz_MHz real clock=score (indexed)	remarks
(owner/poster)	& score :	memCntrl-integrated:	KB/sec RE-CALCULATED	§ - I tested ; ¤ - same

.§§§§§§§§§.

§ Spajky	135 / 380	. . . . i443_BX . . . .	281 *****	(Win98seLite)
CellyTualatin	[{66/185 ¤ 100/281}]	¤	¤	SDRAM

Spare P200mmx PC 7yOld

66 / 60

i430_TX

91

(sdram-60ns)

.§§§§§§§§§.

from Usenet thread : hr.comp.hardver
Re: Zamolba, Anketa, Ram REAL life performance! & Re: Moze mala ankete/test po WinRARu ?

( in croatian language ! )

.§§§§§§§§§.

RaneZ	233 / 351	nforce2U	150
¤	220 / 552	on_Die Cpu / nForce3	249 !!!!	(A64 s.939 10:11 async-)
Zowo	200 / 337	nForce2	168
calypso	166 / 311	i865PE	187
Vinjoza	181 / 286	nForce2_Ultra_400	158
Bleki	216 / 368	nForce2_Ultra	170
®ohar	200 / 290	¤	145
D J	166 / 308	nForce2	185
¤	200 / 361	nForce2	181
ZYXY	133 / 202	i815	151	(Sdram)
MarS	200 / 277	i865G	139
djuro	166 / 244	nForce2_Ultra	147
Ivo	168 / 234	KT400A	139
Sandro	240 / 466	i865PE	194
Delboy	166 / 202	KT600 (KT400A)	122	(async +)
manic	267 / 553	on_Die Cpu !	207	(Athlon64 s.754)

XLR8R	190 / 303	nForce2 400	159
ZZ	200 / 311	nForce2 ?	155

Sanjin

166 / 311

Via ? (Epox 8RDA3+r.3.2)

187

¤	200 / 343	¤	172
f1	166 / 309	Via ? (Soltek mrn-l)	186
Yossarian	133 / 318	i855 ( centrino pl.)	239 ! !	(Banias, 3:4 async +)
N!NO	220 / 603	nForce4 / on_Die Cpu !	274 ***	(A64 s.939 ) in_sync Cas2 2ch

.§§§§§§§§§.

... if some "astonishing" results would re-appear somewhere, I will update this tables ...

.§§§§§§§§§.

from local site´s Forum : Slo-Tech.com :     ( in slovenian language ! ) .§§§§§§§§§.

jest10

165 / 225

SIS 746FX

136

¤	137 / 225	kt400a	131	(async -)
OwcA	218 / 469	i875	215	( 272Fsb 5:4 async -) 2ch
¤	245 / 490	¤	200	(in_sync) 2ch
¤	233 / 500	¤	214	( 291Fsb 5:4 async -) 2ch
¤	166 / 395	i915GM (centrino2 pl.)	238 ! !	(Sonoma) (3:5 async +)
JIM22	240 / 441	nForce2 ???	184
¤	250 / 473	¤	189
flisko	225 / 605	on_Die_Cpu!	268 !!!!	(A64 s.939 )
Moonman	255 / 696	nForce3 / ¤ -Winchester-core!	273 **	(A64 s.939 10:11 async -)
GUFY	273 / 717	nForce3 / ¤ -Winchester-core!	263 !!!	(A64 s.939 10:11 async -)
Fury	200 / 473	nForce3 / ¤ !	236 !!	(A64 s.754)
eDDi	245 / 570	nForce3 / ¤ !	232 !	(A64 s.754 10:12 async -)
mysly	265 / 600	nForce3 / ¤ !	226	(Athlon64 s.754)
Lith	229 / 504	nForce3 / ¤ !	220	(Athlon64 s.754)
Tic	245 / 531	nForce3 ? / ¤ !	218	(Athlon64 s.754)
§ OskarT	133 / 240	nForce2	180
§ MiranP	133 / 222	nForce2	167

§ DamjanK	133 / 308	i815ep(T)	232 !	(sdram) (Asus_TUSL-2C)
§ SandiP	100 / 118	Ali (magic)-sh*t	118	(sdram)

Boss	138 / 239	ApolloPro133a	173	(sdram)
§ LeonB	100 / 160	i845	160	(sdram)

Azrael	100 / 270	- i850 -	* 270 ! !!! *	RDram /defaults*
¤	130 / 299 (struggling!)	¤	230	RDram /conserv.sett.

TribesMan

200 / 467

Intel ?

233 !

2ch

ReD_ScorpioN

124 / 250

ApolloPro133a

201

(sdram Cas222)

_Dejan_

255 / 599

on_Die_Cpu!

235 !

(Athlon64 s.754 , in_sync) (DFI LANParty nF3 250Gb)

zee

166 / 377

i855PM ( centrino pl.)

227

(Banias)

.§§§§§§§§§.

from other Usenet NGs : it.comp.hardware.overclock : Re: Ho in prova P5GD1 e Pentium 4 530;... from Re: ddr2 533 o 667 ?..., from it.comp.hardware.cpu : Re: Bestia Dothan!!! & from Re: Cpu 64 bit alternative     ( in italian language ! ) .§§§§§§§§§.

Gti

200 / 430

i915P

215

2ch Adata 500mhz DDR

Maurizio

166 / 353

i855gm ( centrino pl.)

213

P-M 715 (Dothan) LapTop

Fred	133 / 199	AMD-761	150	sdram
Belfalas	280 / 552	i925XE	197	DDR-II 2ch (Cas3-2-2-4) async +

.§§§§§§§§§.

( .... and so the winner @ defaults used to be for years : RDram + i850 ... in efficiency of using memory ... )

Looks like that simple SDram paired with old trusty BX´s Northbridge (legendary!) was the most per clock performant (efficient in practical handling mem. transactions) combination made ever widely spread on consumer market (w/o any tweaks = default values, I still get around 255 KB/s @ 100MHz Fsb!) It could be easily revived with todays technology & manufacturing process to achieve high clocks (like non-existing PC200/cas2 stick or non-existing/updated & modified/enhanced 0,13micron BX) since having lower general latencies & simpler design than DDRam combinations, even if 2nd one has almost double higher bandwith at same clock than Sdram, but could not be (looks like) properly used in practical world till these days! Nobody in consumer world later did not repeat for long a long term the success of "good old trusty BX" performance ! [except better i850+RDram (bravo engineers! - but f**king Rambus marketing policy lead their type of memory to premature market syndrome of "death", instead of success!) & newly AMDs onCPU 2ch. mem_controller on s.939/940 platform, good job done AMD !] Yeah, I know BX became obsolete these days unfortunatelly ...

But with these benching results, am I stupid thinking that I "smell a smoke" of some past secret/silent "plot" in the IT hardware industry to try to force us to change PC platforms so often? ...Dun´no .. Certain numbers with some proper Bench used can be very revealing sometimes & yes, passed days best bang 4 buck King IMHO used to be an AMD s.754 platform (Sempron3100+ rocks for OC) with built-in memory controller so achieving quite high results per clock (= efficiency) just like an old i440-BX when operating with memory ; again - AMD, good job done ! ... & there are now also A64 s.939 w/enhanced cores available CPUs too and coming also so "enhanced" Semprons s.754 either .. ... From results of previous chipsets for AMD CPUs, you can clearly imagine, why AMD has moved toward a Cpu onDie mem.controller ! ... But with time, the new performance "KINGS" will come, lets be surprised ...

This WinRAR´s built_in benchmark, if I am well informed, wasn´t primary meant to be a memory benchmark, but a tool to Reviewers using it for setting different PC systems as equal as possible in terms of mem.subsystem performance (fiddling with ram clock, latencies & dividers if possible to get identical or close KB/s speeds ; if not the result should be recalculated to appropriate clock!) before proceeding with any benchmarking other stuff to get more reliable results with other benchmarks (taking out of equation mem.subsystem performance), mainly before timing/benchmarking strictly pure CPU (or HD+DMA subsystem) performance of archiving/uncompressing a bunch of data to a single archive file & viceversa using WinRAR when trying to determine practical time needed to accomplish the task! That was the main purpose of this Benchmark & hardware test built_in! ...so memory benchmarking is some kind of "non-intended feature" and quite stresses a mem.subsystem ... As I noticed, IMHO works like this this bench: algorithm allocates for test around 32Mb of ram space, than creates a few hundreds kb size chunk of highly compressed data & moves it back and forth mainly thru memory controller on/off different addresses in that allocated memory space; so that stresses mem.controller & mainly only that allocated Ram space quite a bit ! The speed result is than calculated after few MBs of data has been processed ! This Built-in "Benchmark & hardware test" has NOT to be considered as a Data Compression Bench tool for CPU or HD capabilities & could not be used for that !!! It affects & stresses mainly only memory & its controller; you can (if not having "K8" Cpu) verify that by monitoring NorthBridge_HS temperature /rises much/, Cpu usage /low/ & no IDE activity /HD-Led/ - and that is contrary to actual compressing/uncompressing/testing archive files !

© Zdenko Jerman-Spajky
Article Revisited July 2005