Raskt minne øker ikke ytelsen på A64!

[M4c]

Hvor mye forskjell er det mellom forskjellige minnehastigeter og minneinstillinger på Athlon64? Leste dette innlegget på forumet på Anandtech, tenkte det var interessant her (link til original her).

 

Konklusjonen er; ikke bruk pengene dine på raskt minne hvis du har A64!

Intro

 

I have hinted in the past of doing some memory tests, since it's my belief, given the Athlon 64 architecture and it's integrated memory controller, there is no significant advantage to buying expensive ram over the cheap stuff. All the time I'll say get the cheapest PC3200 value ram @ $140 for 1gig instead of overclockers RAM costing twice as much. ANAND and just about everyones else is guilty of hyping the expensive ram and never show comparos with cheap ram. Are you getting the best bang for the buck buying expensive ram? Well here it is.

 

I'll use the same processor speed thoughout, a realistic overclock of 2610Mhz, only adjusting timings, speed and/or swapping for cheaper parts. I use a overclock, because higher processor speed, the more pronouced the benefits of expensive ram*should show. My tests are not compatable with Anands because he uses different processor speeds thoughout his A64 memory tests which is apples to oranges. Keeping processor speed the same ensures seeing *only* the benefit of differnet memory configuations.

 

 

 

Test Configuration:

 

-Video card is BFG6800GT@ultra.

-CPU Winch 939 1800@2610 Mhz, 290Mhz HTT, 3x LDT

-Operating system Windows 2000 service pack 4 and all updates.

-Dx version- 9c

-Lastest *official* Nvidia video card and Nvidia motherboard drivers.

-All memory is 2 x 512 MB.

 

Test 1: Super budget crucial PC2100 ram Henceforth refered to as “old ram”

290 bus

100 memsetting yeilds PC2300 (145Mhz)

3 3 3 1T

 

Test2: PC3200 Buffalo Value ram w/ CH-5 chips Henceforth refered to as “value ram”

290 Bus

133 mem setting Yeilds ~PC3000 (186Mhz)

3-3-3 1T

 

Test 3: Low Latency (LL) and PC3000, many ram can do this!! I used the TCCD here, but "value" buffalo can handle this speed and latency too!!

290bus

133 mem setting Yeilds ~PC3000 (186Mhz)

2-2-2 1T

 

Test 4: Highend: PQI PC3200 w/ Samsung TCCD

290Bus 1:1

200 memsetting Yeilds ~PC4700 (290Mhz)

3-4-3 1T

 

Test 5: Highend: Crucial Ballistix PC3200 (uses Micron G)

290Bus

166 memsetting Yeilds ~PC3800 (237Mhz)

2.5-2-2 1T

 

Testing software:

 

All real world except Sandra mem/cpu just to show how erronious it is for A64:

 

-SiSoftware Sandra Lite 2005, mem and CPU

-Far Cry 1.3 HardwareOC Far Cry Benchmark Demo , max details, AF = 1, AA = 0, 800x600 and 1600 x 1200

-Unreal Tournament 2003 @ 10x7 default settings

-Doom3- All High Quality Settings, "timedemo demo1 1", AA4x, 1600x1200 and 1024x768

-Half Life 2- All High Quality Settings, Route Kanal custom demo features indoor/outdoor, AA6x, AF4x, 1600x1200 and 1024x768

-LAME 3.96.1 (build from mitiok.cjb.net)Default settings 163MB CBS Masterworks Beethoven Symph #9 op.125 Allegro non troppo, un poco maestoso 16:44m

-Cinebench 2003

-ScienceMark 2.0 beta (23SEP03 build) - Molecular Dynamics

-PiFast Parameters: 0,0,10m,1024,1

-Superpi 1M

-Winrar 3.41 internal benchmark

-POV-Ray v3.6 Chess2.pov 1024x768 AA.03

-DVDShrink 3.2 Lawrence Of Arabia Superbit Encode Chap 1-6, 863MB to 500 MB = 58% compression (sharp quality setting)

 

Results

 

Sandr mem INT ALU

PC2300 - 3-3-3 1T- 4285

PC3000 - 3-3-3 1T- 5556

PC3000 - 2-2-2 1T- 5602

PC4700 - 3-4-3 1T- 7272

PC3800 - 2.5-2-2-1T - 6977

 

Sandra mem FLOAT FPU

PC2300 - 3-3-3 1T- 4288

PC3000 - 3-3-3 1T- 5550

PC3000 - 2-2-2 1T- 5603

PC4700 - 3-4-3 1T- 7267

PC3800 - 2.5-2-2-1T - 6904

As you can see, TCCD Sanda mem is showing a whopping ~70% faster vs. the cheapest PC2100 ram ever sold. And 30% more than value ram, PC3000 ram @3-3-3. Will this translate into real performance gains? Well lets see.

 

Sandra CPU Drystone ALU

PC2300 - 3-3-3 1T- 11064

PC3000 - 3-3-3 1T- 11068

PC3000 - 2-2-2 1T- 11066

PC4700 - 3-4-3 1T- 11071

PC3800 - 2.5-2-2-1T - 11068

no difference

 

Sandra CPU FPU

PC2300 - 3-3-3 1T- 4253

PC3000 - 3-3-3 1T- 4253

PC3000 - 2-2-2 1T- 4253

PC4700 - 3-4-3 1T- 4252

PC3800 - 2.5-2-2-1T -4353

no difference

 

Sandra CPU SSE2

PC2300 - 3-3-3 1T- 5527

PC3000 - 3-3-3 1T- 5536

PC3000 - 2-2-2 1T- 5528

PC4700 - 3-4-3 1T- 5551

PC3800 - 2.5-2-2-1T - 5446

meh, now lets get to the real world:

 

Farcry 800x600 in FPS

PC2300 - 3-3-3 1T- 61.46

PC3000 - 3-3-3 1T- 66.49

PC3000 - 2-2-2 1T- 72.67

PC4700 - 3-4-3 1T- 74.97

PC3800 - 2.5-2-2-1T - 77.08

 

Farcry 1600x1200 in FPS

PC2300 - 3-3-3 1T- 62.25

PC3000 - 3-3-3 1T- 68.14

PC3000 - 2-2-2 1T- 71.31

PC4700 - 3-4-3 1T- 73.52

PC3800 - 2.5-2-2-1T - 74.42

On farcry, we see ~18% faster performance from the old ram. But only ~8% faster than value ram. hint: spend the xtra $100 on a better video card. Notice async low latency Ballistix owning all even 1:1 TCCD.

 

 

Doom3 1024x768 in FPS

PC2300 - 3-3-3 1T- 77.8

PC3000 - 3-3-3 1T- 80.3

PC3000 - 2-2-2 1T- 81.4

PC4700 - 3-4-3 1T- 82.3

PC3800 - 2.5-2-2-1T - 82.4

No signifigant difference above PC3000

 

Doom3 1600x1200 in FPS

PC2300 - 3-3-3 1T- 44.6

PC3000 - 3-3-3 1T- 44.6

PC3000 - 2-2-2 1T- 44.6

PC4700 - 3-4-3 1T- 44.5

PC3800 - 2.5-2-2-1T - 44.6

Video card limitation here...

 

Half-life 2 1024x768 in FPS

PC2300 - 3-3-3 1T- 113.3

PC3000 - 3-3-3 1T- 120.9

PC3000 - 2-2-2 1T- 131.6

PC4700 - 3-4-3 1T- 133.3

PC3800 - 2.5-2-2-1T - 134.4

We see low timings here credited with speed since PC3000LL is basically equivalent to PC4700. Overall nothing big above PC 3000.

 

Half-life 2 1600x1200 in FPS

PC2300 - 3-3-3 1T- 98.77

PC3000 - 3-3-3 1T- 102.5

PC3000 - 2-2-2 1T- 106.0

PC4700 - 3-4-3 1T- 107.9

PC3800 - 2.5-2-2-1T - 107.3

Same as above. Save your money.

 

Pifast in seconds

PC2300 - 3-3-3 1T- 38.91

PC3000 - 3-3-3 1T- 36.53

PC3000 - 2-2-2 1T- 35.48

PC4700 - 3-4-3 1T- 34.67

PC3800 - 2.5-2-2-1T - 34.45

We see bandwitdh have *some* effect here..nothing to write home about. PC2300 vs PC4700 is 10.9% faster, but very small difference vs. value ram.

 

Superpi 1M in seconds

PC2300 - 3-3-3 1T- 37

PC3000 - 3-3-3 1T- 34

PC3000 - 2-2-2 1T- 33

PC4700 - 3-4-3 1T- 32

PC3800 - 2.5-2-2-1T -32

Similar as above. Nothing drastic at anything above PC3000

 

Science Mark- Molecular Dynamics time in seconds

PC2300 - 3-3-3 1T- 67.03

PC3000 - 3-3-3 1T- 66.92

PC3000 - 2-2-2 1T- 66.68

PC4700 - 3-4-3 1T- 66.69

PC3800 - 2.5-2-2-1T - 66.47

Seems bandwidth has no effect here.

 

CineBench 2003 Rendering

PC2300 - 3-3-3 1T- 363

PC3000 - 3-3-3 1T- 365

PC3000 - 2-2-2 1T- 366

PC4700 - 3-4-3 1T- 368

PC3800 - 2.5-2-2-1T - 367

Seems bandwidth has no effect here.

 

CineBench 2003 Shading

PC2300 - 3-3-3 1T- 414

PC3000 - 3-3-3 1T- 426

PC3000 - 2-2-2 1T- 434

PC4700 - 3-4-3 1T- 439

PC3800 - 2.5-2-2-1T - 440

Infinitesmal bandwidth help

 

CineBench 2003 OpenGL SW

PC2300 - 3-3-3 1T- 1949

PC3000 - 3-3-3 1T- 1992

PC3000 - 2-2-2 1T- 2019

PC4700 - 3-4-3 1T- 2032

PC3800 - 2.5-2-2-1T - 2034

Infinitesmal bandwidth help

 

CineBench 2003 OpenGL HW

PC2300 - 3-3-3 1T- 3568

PC3000 - 3-3-3 1T- 3662

PC3000 - 2-2-2 1T- 3756

PC4700 - 3-4-3 1T- 3781

PC3800 - 2.5-2-2-1T - 3796

this is getting old, on to another game.

 

UT2003 10x7 Flyby

PC2300 - 3-3-3 1T- 297

PC3000 - 3-3-3 1T- 319

PC3000 - 2-2-2 1T- 334

PC4700 - 3-4-3 1T- 341

PC3800 - 2.5-2-2-1T - 345

14% max boost, hey but very small difference with value ram.

 

UT2003 10x7 BotMatch

PC2300 - 3-3-3 1T- 110

PC3000 - 3-3-3 1T- 118

PC3000 - 2-2-2 1T- 123

PC4700 - 3-4-3 1T- 127

PC3800 - 2.5-2-2-1T - 128

15% boost over PC2300, only 8% over value ram.

 

Winrar in KB/s

PC2300 - 3-3-3 1T- 496

PC3000 - 3-3-3 1T- 564

PC3000 - 2-2-2 1T- 636

PC4700 - 3-4-3 1T- 651

PC3800 - 2.5-2-2-1T - 703!!!

Here we see pretty signifigant gains due to bandwidth 41% diff top to bottom. This test obviously is very beneficial to have bandwidth or LL. Update: Crucial Ballistix owning all, Winrar seems the like LL even more than bandwidth. We sorta saw this already by the very close scores of PC3000 LL VS PC4700 high bandwitdh.

 

Lame in seconds

PC2300 - 3-3-3 1T- 56

PC3000 - 3-3-3 1T- 56

PC3000 - 2-2-2 1T- 56

PC4700 - 3-4-3 1T- 56

PC3800 - 2.5-2-2-1T -56

Seems bandwidth or LL has no effect here.

 

Chess2.pov in seconds

PC2300 - 3-3-3 1T- 246

PC3000 - 3-3-3 1T- 245

PC3000 - 2-2-2 1T- 246

PC4700 - 3-4-3 1T- 247

PC3800 - 2.5-2-2-1T -246

Seems bandwidth or LL has no effect here.

 

DVDShrink in seconds

PC2300 - 3-3-3 1T- 243

PC3000 - 3-3-3 1T- 235

PC3000 - 2-2-2 1T- 228

PC4700 - 3-4-3 1T- 221

PC3800 - 2.5-2-2-1T -224

DVDShrink seems to like bandwidth a little bit.

 

 

Summary of Results

Well there it was. While sandra likes to show whopping 70% increased bandwidth of expensive TCCD it really does'nt translate to real world . Only one area seemed signifigant was winrar. The couple game tests which slow slight inprovement with bandwidth or LL, I think you're much better of upgrading to the next tier of video card if on a budget instead of investing with overpriced ram. Investing in expensive Ram probably offers to poorest bang for the buck of any computer part. What's intesting was Ballistix's async 166 and LL the performance faired a little better than TCCD 1:1. Seems A64 prefers LL rather than bandwidth? Also shows that running async is'nt to be shunned.

 

Course if you want the best, go right ahead and invest in expensive ram, as shown it makes a slight difference in gaming and winrar but I'm sticking with $130 2x512 value ram and use a divider from now on because the price to performance ratio is simply not worth it. IMO You're much better off using that cash for the next level of video card which really does boost performance by ~70% in games.

 

One other thing of intrest is PC3000 running async 133 at low timings 2-2-2 was almost as good as PC4700 high bandwidth but loose timings TCCD is forced to run at 290Mhz.... many a ram can do ~200Mhz 2-2-2.

 

 

Bottom Line:

 

Some Games = very slight benefit to using expensive high bandwidth TCCD or expensive LL ram Like Ballistix over value ram. Winrar sees major benefit though, 41% top to bottom.

 

LAME, Cinebench, ScienceMark, PiFast, Superpi, POV-Ray, DVDShrink shows no or infinitesmal boost above Value ram.

 

 

 

 

 

 

 

*Note: I've gotten some PM complaining "well I run 2-2-2 275Mhz with BH-5 or OCZ VX or 320Mhz 3-4-3 w/TCCD" Well first of all, how many volts you push though? 3.5? 4.0? With a 120mm 120 cfm fan blowing thier way? Plus you use 2x256 which overclock much much more to begin with rather than 2x512 which most every user will use. Sorry, this test features the maximum overclocks of TCCD and Micron G with safe onboard voltages most Anandtech users and Anand/Westley themselves will use. My memory modules did as well or better than anands A64 memory tests. 2.85 max voltage. It also does'nt include OCZ booster, active ram cooling, and/or motherboard voltage modifications. We all know what's possible but I'm focusing on probable when entertaining different memory. Even so after all that work, expense, voiding your warranties, etc, you won't see much benefit over value ram running "ansyc" as I think I've shown

[C4STOR]

Godt å få bekreftet det jeg har prøvd å si ei stund nå. Dyr ram gir ikke ytelse som står i stil med prisen, en bedre cpu og/eller gpu vil gi langt mer valuta for pengene.

 

Jeg liker spesiellt denne setningen:

ANAND and just about everyones else is guilty of hyping the expensive ram and never show comparos with cheap ram.

 

Dette er egentlig essensen i hele greia. RAM blir hypet opp slik at folk tror at det faktisk finnes ytelse å hente.

 

Denne testen var virkelig bra

[ADT]

Er ikke helt fornøyd, de kunne godt tatt med forskjellen på 512MB RAM kontra 1024MB RAM også, dette jeg lurer på iallefall, er det verdt de 1350kr'ene ekstra? I så fall, er det i spill eller programmer jeg merker det?

 

edit: Og ja, jeg er vel en av de som har virkelig blitt offer for overklokkingsidenes "hyping" av minne betalte 1350kr for den ene 512MB brikken jeg har nå... klokker bra da men... *snufs*

Endret 15. februar 2005 av ADT

[Gjest Slettet+6132]

Takker for denne tråden.. og støtter C4Stor..

Har selv lurt på dette lenge.. og opprettet tråder (tidligere) som ikke ga noe mer enn synsing (tror det var C4Stor som kom med det eneste av verdi),

Vel.. jeg lurte da på P4-oppsett... og har nå testet selv...

Byttet ut Twinmos "cheapo" DDR400 til Twister DDR500...

I "bandwith-tests" så er jo DDR500 fullstendig overlegen...men i alle andre tilfellet(appliksjoner) er forskjellen *helt* neglisjerbar.

Det vi kanskje glemmer/ikke tar med i beregninger her.. er at dyr ram tillater noe mer "headroom" for klokking av FSB (selv om vi tar dividerer med i betraktning).

Endret 16. februar 2005 av Slettet+6132

[Fogga]

Jeg har aldri trodd at rask ram øker ytelsen på A64, men det øker stabiliteten siden minne kontrolleren sitter i cpu'en. Og den vil alltid være raskere uansett hvor raskt minne du setter inn.

[C4STOR]

Jeg har aldri trodd at rask ram øker ytelsen på A64, men det øker stabiliteten siden minne kontrolleren sitter i cpu'en. Og den vil alltid være raskere uansett hvor raskt minne du setter inn.

Noen spesiell grunn til at det skulle øke stabiliteten? Da tror jeg du vet noe ingen andre er klar over

[Playmofetish]

dyr ram tillater noe mer "headroom" for klokking av FSB (selv om vi tar dividerer med i betraktning).

Dette er jo det viktigste når det gjelder klokking, så der er et lite argument for å kjøpe dyrere RAM dersom man skal klokke. Men egentlig kan man likesågodt hoppe rett opp på PC4000 eller høyere, dersom man planlegger klokking. Da får man ihvertfall noe som er garantert å klare de oppgitte spesifikasjonene, og slipper å tyne minnet/hovedkort for de siste ekstra MHz´ene.

[C4STOR]

dyr ram tillater noe mer "headroom" for klokking av FSB (selv om vi tar dividerer med i betraktning).

Dette er jo det viktigste når det gjelder klokking, så der er et lite argument for å kjøpe dyrere RAM dersom man skal klokke. Men egentlig kan man likesågodt hoppe rett opp på PC4000 eller høyere, dersom man planlegger klokking. Da får man ihvertfall noe som er garantert å klare de oppgitte spesifikasjonene, og slipper å tyne minnet/hovedkort for de siste ekstra MHz´ene.

Det er stortsett bare ved overklokking at man har behov for ram som takler høye hastigheter, men hvis man ser på testen fra forumet hos Anandtech så er det ikke noen særlig reell ytelse å hente på å kjøre ram på høy hastighet. Så mao så taper man svært lite ytelse på å kjøre ram asynkront. Man kan mao klokke cpu å kjøre ram på 400MHz uten noe ytelsestap som er nevneverdig

Endret 16. februar 2005 av C4STOR