pretty shocking 875 review

[Guest]

We knew the P4 was limited to 4 GB address space, now it seems the 875 is also limited to 1 GB RAM or 2 dimms unless you want to take a >30% (!!) performance hit. Ouch !
<A HREF="http://www.firingsquad.com/hardware/building_gaming_opteron_2003_Part2/page16.asp" target="_new">http://www.firingsquad.com/hardware/building_gaming_opteron_2003_Part2/page16.asp</A>

<A HREF="http://www.firingsquad.com/hardware/building_gaming_opteron_2003_Part2" target="_new">http://www.firingsquad.com/hardware/building_gaming_opteron_2003_Part2</A>

= The views stated herein are my personal views, and not necessarily the views of my wife. =

 

[jurians]

heh thank god i only have 1 gig total over 2 slots with my 875..... My professor who designed some of the pentiums told me about the potential problem that the next generation of chips will have with memory and told me that ecc was practially necessary for the large amount of memory that athlon can support...

dammit hes never wrong... yet...

 

[Guest]

Update, it seems the problem is limited to the ammount of DIMM's, and using 2x1 GB DIMM's would not give you this problem. This still practically limits the i875 to 2 memory slots though, which is not exactly a luxury for such a high end, workstation class chipset. Maybe AMD wasnt that crazy choosing for registered memory for the Opteron/FX after all.

= The views stated herein are my personal views, and not necessarily the views of my wife. =

 

[Ncogneto]

<font color=red>Second, it is clear that those who complained about the Registered DDR requirement/support of the Athlon64 FX did not fully understand the technology, or they did not consider pushing high-performance computing envelope. Registered DDR is critical for ensuring that performance remains stable despite a large number of installed DIMMs.</font color=red>

Isnt this what I said earlier?

It's not what they tell you, its what they don't tell you!

 

[Guest]

I don't know if you said it or not, but its an eye opener for sure (well,at least to me).

Also, I thought this wasnt that big an issue, as using 1 GB Dimm's should solve it, but then it appears I just can't seem to find any branded 1 GB PC3200 Dimms on pricewatch. All those "1 GB" entries are either 2x512 MB kits and/or generic modules. No corsair, crucial or Micron to be seen. Not a good thing...

= The views stated herein are my personal views, and not necessarily the views of my wife. =

 

[juin]

all the E series take register memory for Xeon.I875 is still for high end gaming use

I dont like french test

 

[Mephistopheles]

Indeed... doesn't 875 support registered memory? So there is no problem; if you really want more than 1GB, just go registered and no problem!...

<font color=red><b>M</b></font color=red>ephistopheles

 

[juin]

yes but my guess is opteron/E7505 make better use of reg memory that I875.

I dont like french test

 

[Mephistopheles]

that might be true... but it would be awfully difficult to test, as this would be a cross-platform comparison...

<font color=red><b>M</b></font color=red>ephistopheles

 

[ytoledano]

Will using 4x256 stick be just as bad?

My anger and my wrath will be poured on this place on man and beast on the trees of the field and the fruit of the ground it will burn and not be quenched Jeremiah 7:20
I am ALPHA and OMEGA, burn!

 

[Guest]

However, it would be simple to test a P4 or Xeon + 1 or 2 GB registered memory against an opteron/FX. Not only simple, probably also very usefull for a lot of people that use these ammounts of RAM, and I am firmly convinced that will be you and me as well in a year or two.

= The views stated herein are my personal views, and not necessarily the views of my wife. =

 

[Guest]

Yes.

= The views stated herein are my personal views, and not necessarily the views of my wife. =

 

[juin]

I875 would win as it made for desktop desktop are allwasye faster that server stuff but that dont mean a good support of it in any case intel work primary for 256 to 1 gig memory range and DDR 333 and DDR 400.

I dont like french test

 

[pjordan]

Why would using 4 x 256 be just as bad? They don't test that in their benchmarks. <A HREF="http://www.anandtech.com/memory/showdoc.html?i=1839&p=4" target="_new">http://www.anandtech.com/memory/showdoc.html?i=1839&p=4</A>
Anandtech says 4 x 256 is best, but they don't test 2 x 512 it seems.

 

[Kelledin]

Why would using 4 x 256 be just as bad? They don't test that in their benchmarks.

They've concluded that pretty much any time you populate more than two DIMM slots with unbuffered DIMMs, something will have to be sacrificed to maintain stability: either high bandwidth or low latency. AMD sacrifices high bandwidth; Intel sacrifices low latency.

Use registered DIMMs, and >2 DIMMs shouldn't be a problem--aside from the slight latency penalty incurred with registered DIMMs.

And if the CPU or chipset can't actually handle registered memory...well, w0e is j00, either stick with just two DIMMs or sacrifice performance. :frown:

Anandtech says 4 x 256 is best, but they don't test 2 x 512 it seems.

I also didn't see a note as to whether AnandTech used registered or unbuffered DIMMs. The "UNbuffered" they keep saying on that page refers to the Sandra test, not the memory itself.

<i>I can love my fellow man...but I'm damned if I'll love yours.</i>

 

[pIII_Man]

I wouldnt even call sandra's memory benchmark a test, that thing is so inacurate. I know some people will think diffrently but as far as benchmarks go, sandra sux.


If it isn't a P6 then it isn't a procesor
110% BX fanboy

 

[Ncogneto]

anandtech was simplly measuring for bandwith, not taking into account additional latency would more than offset any gains that a slight increase in bandwith would net.

It's not what they tell you, its what they don't tell you!

 

[Ncogneto]

but as far as benchmarks go, sandra sux

AMEN! More people need to realize this. At best Sandra, HDDtach etc are nothing more than simple diagnostic tools not indicative of real world performance.

It's not what they tell you, its what they don't tell you!

 

[jurians]

lol thanks for reminding my that my 875 board supports ecc memory.... i was wondering.... would you rather have bandwidth decrease or the latency increase. I can see the benefits of either in different situations. Any thoughts on this?

 

[Guest]

> would you rather have bandwidth decrease or the latency
> increase.

Heh.. would you rather have a lower clockspeed, or a lower FSB/memory speed ? I mean, there is no clear cut answer to this, depends on the ammoung of the decrease, and it depends on the apps. But if I had to pick one either way, I'd go for good latency over good bandwith. The A64 hardly seems to break a sweat even when its forced into PC1600 (100/200 Mhz). The P4 seems to pay a much higher penalty for its increased latency using 4 unregistered DIMM's.

= The views stated herein are my personal views, and not necessarily the views of my wife. =

 

[savantu]

<A HREF="http://firingsquad.com/hardware/building_gaming_opteron_2003_Part2/page18.asp" target="_new">http://firingsquad.com/hardware/building_gaming_opteron_2003_Part2/page18.asp</A>

According to Firing Squad having all the 4 dimms filled eqauls a 30% drop in performance.

According to Intel you have optimum perfoamce with all the 4 dimms filled.Which is to trust???

<A HREF="http://www.intel.com/design/chipsets/applnots/25273001.pdf" target="_new">http://www.intel.com/design/chipsets/applnots/25273001.pdf</A>


Btw , 1Gb DDR dimms are not suported on Canterwood.

Firing Squad needs some attention...so they cranked something for people to discuss....

Long live Intel!

From the darkside...you know!

 

[Guest]

Anand's test (though mostly synthetic benches) showed some very marginal performance improvement using 4 Dimms. My guess is 4 Dimms helps bandwith but hurts latency a lot. Some apps/benches will gain a few percent (like those Anand used) others may lose substantially (firingsquad).

It would be good if someone really spit this out, also throwing in registered memory performance on the P4..

= The views stated herein are my personal views, and not necessarily the views of my wife. =

 

[Ncogneto]

According to Intel you have optimum perfoamce with all the 4 dimms filled.Which is to trust???

Well intel said it, so then it must be true!!!!!! All kneal before the mighty Intel.

Btw , 1Gb DDR dimms are not suported on Canterwood.

Oops, strike that previous remark

Firing Squad needs some attention...so they cranked something for people to discuss....

Oh Really? And what may I ask is your reason?


It's not what they tell you, its what they don't tell you!

 

[Vapor]

Anybody know if this applies to other P4 chipsets?? Namely mine, the 850E....I have 4x256. If no one knows, could I test it by removing 2 dimms...? Or would the unequal amounts of memory influence the results?

BTW, if the amount of memory makes no difference and if someone gives me the names of a few truthful benchies (aka, not latency or not bandwidth oriented benchies) i can go ahead and test it to see if the >2DIMM problem effects 850s.

RDRAM = ENEMY

 

[Guest]

I'm not sure really. I think it is related to the PAT technology on the 875, so its quite possible your 850 is immune to it.. especially since RDRAM is pretty different beast, I wouldnt draw conclusions from the 875 review. You could of course test, and report back

>I have 4x256

Oh dear, yet another user of 1 Gig; I thought those people didnt really exist, and everyone ran with ~256 Mb; looks there are just too many exceptions around here

= The views stated herein are my personal views, and not necessarily the views of my wife. =