you guys want non-storage articles?

[honold]

the domain name is a bit, uh, limiting

anyway i wrote this overclocking piece up for the overclockers and hardocp forums, haven't posted it yet. if you want me to html-ize it and host the pics here, i'm game. i'm only proposing this in response to the 'how do we get more people on here' chatter in the threads.

so, do you want overclockers here?

test system:
abit ic7 motherboard, intel 875 chipset, 1.4 release bios, no gaming accelleration
2x512mb kingston hyperx pc3500 memory (2.7v 5:4 for 3448/400, 2.8v 5:4 for 2225/400)
2x512mb corsair twinx pc3700 memory (2.8v 1:1 for 3448/500)
p4 2.4c cpu (1.50v)
abit siluro geforce4 ti4200 video card (stock speed), latest reference drivers
western digital raptor 36g 10k rpm ide hard drive (connected to ich5 sata controller)
turtle beach santa cruz sound card, latest stable drivers

NOTE: 3448 timing actually resulted in 2.5-4-4-8 on an abit ic7 motherboard (read by cpu-z)

2225/400 3448/400 3448/500

3dmark2001 13252 12923 13250
aidaread 5226 5041 5697
aidawrite 1917 1874 1873
quake3 324.9 308.6 319.9
superpi 224 234 226
winrar 176 207 183

3dmark2001 +2.545% base +2.530% = -0.015% (compared to 2225/400)
aidaread +3.669% base +13.013% = +9.344% (compared to 2225/400)
aidawrite +2.294% base -0.053% = -2.347% (compared to 2225/400)
quake3 +5.281% base +3.661% = -1.620% (compared to 2225/400)
superpi +4.464% base +3.539% = -0.925% (compared to 2225/400)
winrar +17.613% base +13.114% = -4.499% (compared to 2225/400)

total +5.977% base +5.967% = -0.010% (compared to 2225/400)

IN CONCLUSION:

aida, winrar, and superpi are dubious benchmarks because they're not necessarily indicative of

real-world use. memory pages are read and written, but that is only one piece of a whole

application. files are compressed, but not terribly often. pi is seldom calculated i would

like to introduce more 'real-world benchmarks' like pcmark, business winstone, content creation

winstone, and more games before coming to a total conclusion.

as it stands here on this board and chipset, 25% ddr clock improvement nets the same average speed

increase (about 6%) as a 71.25% (25%+100%+100%+60%/4) timing improvement. this would seem to

indicate that, percent for percent, clock gains net greater performance than timing gains.

furthermore, clocks can go higher, while 2225 is the most aggressive timing possible.

it would also seem that 6%, an optimistic average considering the 'lab room' results of tests such

as aida32, is not worth chasing aggressively. 1x512mb of bh-6 winbond buffalo costs $89 with free

shipping at newegg. 2x256mb of ocz pc3700 gold costs $215 with free shipping. with the price

difference between the two, one could buy a 36g 10,000rpm western digital raptor, which should

translate to a real-world gain of more than 6% (as compared to a typical 7200rpm drive).

regardless of how much memory is purchased (2x256mb, 4x256mb, 2x512mb, etc) the price difference

remains just as stark.

i consider these results a work-in-progress as i would like to add more tests to the average. i'm

also interested in testing specific timings to determine which matter the most in general use.

screenshots of all of my tests including cpu-z windows showing both cpu and memory information (as

well as my latest results comments) are available at http://www.badbeat.com/ram/ - please only use

these for spot-checking validity. if they get hammered, i will be forced to take them down or

find a free host.

i welcome your questions/comments/corrections. i am only interested in finding the best bang for

the buck, for everyone.

 

[honold]

note: the activity, especially the jpegs, could increase traffic quite a bit. the jpegs need to be available just to provide validity.

 

[blakerwry]

note: the activity, especially the jpegs, could increase traffic quite a bit. the jpegs need to be available just to provide validity.

what about making them gifs? full size (resizing adds blurryness which hurts gif compression) gif screenshots are usually quite a bit smaller than jpg for me... 64 colors is usually more than enough...

 

[CityK]

NOTE: 3448 timing actually resulted in 2.5-4-4-8 on an abit ic7 motherboard (read by cpu-z)

In order for a DDR chip to support CL3, the necessary pipeline stages have to be available. As far as I'm aware, only Micron and Samsung chips support it (i.e. are hardwired for it). So, even if the option is available in the BIOS, functionally it defaults to 2.5. Also, IIRC, Intel chipset doesn't support CAS 3 functionality.

also interested in testing specific timings to determine which matter the most in general use.

For DDR, tRAS is the setting. CL is no longer important like it was with vanilla SDRAM because of (a) clock forwarding and (b) the ability to send the read command early, thus CL's hit ends up being masked in the background the majority of the time

Setting tRAS to 5 is asking for corrupt data. As opposed to popular misconception, tRAS is not a latency - rather, its the minimum amount of time that a "row" in memory has to be open. It should be set to tRCD + CL + (BL/2) ... where BL is the burst length, which = 8 with DDR. A lower tRAS just allows the memory controller to terminate transfers earilier ... and you can see where trouble might develop if things were quite finished

If your chipset can negotiate a CMD of 1T with the quantity of RAM you have in your system, by all means do it. Probably require CMD = 2T though with 2 sticks and 1GB.

Will have to look over your post more closely later on.

 

[Groltz]

so, do you want overclockers here?

Well, I'll read it. Other than that, good luck.

It really feels shitty to take a few hours to write an article for SF, include pictures, etc. just to get (almost) no responses.

 

[honold]

cityk: grab results.txt from the web site - it actually formatted properly and thanks for the detailed info. fwiw, at 2225/2.8v/ddr400 the hyperx pc3500 could pass anything (memtest loops, prime95 torture test, etc) without barfing.

groltz: don't read it out of guilt i did the benchmarking for my own benefit, and i documented it for the benefit of the overclockers.com and hardocp.com forums. i'm offering to post it here instead of at those places to attract attention. none of it was in vain.

 

[Groltz]

groltz: don't read it out of guilt i did the benchmarking for my own benefit, and i documented it for the benefit of the overclockers.com and hardocp.com forums. i'm offering to post it here instead of at those places to attract attention. none of it was in vain.

It's not guilt, it's personal interest.

 

[honold]

one more thing: the corsair pc3700 uses new samsung ics (which are super-relaxed latency; made to run 3-3-8-4 pc3700, and won't even do lower than that at pc3200), and cas3 is an actual option in the ic7's bios. intel making that mean '2.5' is possible.

 

[honold]

...and for clarification, my proposal is to make an actual article out of it to be discussed in the forum, not simply post it to the forum.

 

[Prof.Wizard]

honold kudos man!
Really, great initiative.

 

[Tannin]

Absolutely, Honold. Go for it!

 

[CityK]

Just started reading. Formatting looks much better from the results.txt

First thing I notice is that the percentages in the column for the 3:4:4:8/500 compared to a base of 2:2:2:5/400 are wrong. Example:

aidaread " = +9.344% (compared to 2225/400)"

but if you normalize to the 2225/400, you get 5697/5226 = 1.090126292, or 9.012%.

Back later.

 

[CityK]

Oops, sorry, I see what you meant now. (Obviously) It wasn't clear to me at first.

 

[CityK]

Comments:
- I disagree with taking an equal weighting with all 6 benchmarks and arriving at an observation of an overall percentage increase. You should use a range, and try to discuss why the app benefits or appears not to benefit in the given situations.

Taking a look a just one of those 6 you used: 3DMark2001 is a composite of something like 20 differently weighted benchmarks...several which are graphics card limited.

25% ddr clock improvement nets the same average speed increase (about 6%) as a 71.25% (25%+100%+100%+60%/4) timing improvement

You shouldn't look at it that way in terms of performance.

There are a couple of different types of memory accesses - page hit, page miss, etc ... changing the different timing settings has different implications for each, and by extention, direct implication on the access pattern utilized by the underlying application being run. For example, after an initail access, a page (aka a row) has been opened in memory. Now if the next access is for data in that same page you get a page hit. When page hits occur, performance savings are realized because the memory controller doesn't have to go through the steps of sending the row address etc. - i.e. tRCD has been completely eliminated from the picture. So, if your app tends to have a lot of page hits, switching from a tRCD = 4 to tRCD = 2 is fairly meaningless (and vice-versa). Much of the same can be said for the other timings too.

You have to develop a model of the different types of memory accesses and then look at the performance increases/decreases that a change in timings would create. This is the real meat and potatoes of understanding memory performance....and is also where you observe how memory bandwidth is derived.

Also, you need more apples to apples kind of comparisons. Look to showing results of the system running 1:1 at 2228/400, 3448/400.
Likewise, try running 3448/500 at 5:4
But the real question is what happens when we increase the DDR clock, but keep the latencies low, when 1:1

On Buffalo RAM - unless you can take advantage of one of those rare Mushkin or Corsair promos/sales, it probably is the best bang for the buck (...if only Newegg delivered here, wow is me ). However, just because its equipped with BH-5 chips doesn't necessarily mean the chips will be from the same speed demon batches. Simply, not all bins have the same great results. Companies like Corsair and Mushkin (supposedly) sample bins and pick the best lot. Melco (Buffalo Tech) probably just orders them straight in from Winbond. Also consider that the BH-5 is spec'd for 2.5:4:4....so not all chips are going to be capable of making the lower latency settings. Nonetheless, I would suspect that the majority would, since they really are great chips. But also playing a major role is the quality of the module's PCB onto which the chips are placed i.e. several signaling issues like termination, isolation, resistance etc etc. Corsair and Mushkin produce good PCBs. And Melco?? Who knows. Speaking of companies, please don't support OCZ - they are utter ripe off artists. That company has a real shady past..(if someone told me their was some underworld connection, it would not shock me in the least). Tthe only thing that has elevated them recently is (a) the overwhelming success of the Winbond chips ( and even then, they've screwed up quality wise with their modules on many occasions) and (b) a huge marketing campaign ... which reminds me of a comment in an unrelated memory pricing article I saw on the Inquirer last month - something along the lines of "nothing moves the memory market like FUD"....

Anyways, bring on some more benches, right now you have a rather limited set. I think your off to a great start, and that a solid article could be derived from your results.

CK

 

[honold]

Comments:
- I disagree with taking an equal weighting with all 6 benchmarks and arriving at an observation of an overall percentage increase. You should use a range, and try to discuss why the app benefits or appears not to benefit in the given situations.

yeah i don't like it either. i'm just trying to make it as 'quick big picture' as possible. i really don't like it because of the weighting of stuff like aida32 and winrar, how dramatic their differences are despite the fact that it's not a desktop user metric. if i had a number of other things in the fray i'd feel better about it.

Taking a look a just one of those 6 you used: 3DMark2001 is a composite of something like 20 differently weighted benchmarks...several which are graphics card limited.

3dmark2001 is something everyone apparently wants to see still. if i limit it to just a couple games, the game bias (why not ut2k3? why not xxx?) stuff starts to come out more.

You have to develop a model of the different types of memory accesses and then look at the performance increases/decreases that a change in timings would create.

that's a pretty large task i feel like this sort of raw tech information is already out there (sans benchmarks). i'm not experienced enough on memory/application memory usage to draw conclusions about specific things (this app has heavy xxx usage, so the xxx timing would be more important). this kind of stuff can be derived if i do timing-oriented benchmarks (which i intend to do soon, maybe tonight) but i'm not sure how much further past that - into the education zone - that i want to go.

Also, you need more apples to apples kind of comparisons. Look to showing results of the system running 1:1 at 2228/400, 3448/400.
Likewise, try running 3448/500 at 5:4

yeah the 1:1 400 stuff will be coming soon to illustrate what differences are found between async and sync modes on this specific chipset. the general concensus is that it has very little effect on the 865/875.

But the real question is what happens when we increase the DDR clock, but keep the latencies low, when 1:1

i hope to arrive at a good guess on this one by providing a barrage of metrics and comparing scale.

as for bufflo, i know their stuff isn't hand-picked and their pcb work is not as optimized as somebody like corsair, but the universal results seem to be remarkable for the price. it also appears that their pc3200 and pc3700 sticks are the EXACT same thing. same pcb, same chips, same everything. it's possible the pc3700 ones are hand-picked chips, but it seems more likely that they tuned into the overclockability of their own sticks and went with it.

as for ocz, they appear to have totally turned a corner on their reputation in the last few months. i mention them by name because they're presently considered the 'it' brand. corsair's pc3700 with its slack timings are scaring people off in droves.

Anyways, bring on some more benches, right now you have a rather limited set. I think your off to a great start, and that a solid article could be derived from your results.

have any recommendations on benchmarks to add? i'm throwing in serious sam 2, divx encoding, and pcmark 2002 for now. i'm trying to get as real-world as possible, but i'd like to get more into general use patterns over gaming or content creation.

if i might ask, what is your background in all this? you seem to be extremely well-versed on memory specifically.

 

[honold]

http://www.badbeat.com/ram/

looking much better

 

[CityK]

No special background, it just interests me both on a technical level and from a business perspective (a lot of fat ducuts to be made and lost in the DRAM industry).

CK

 

[CityK]

BTW, isn't DivX 3.11 the version that was based on stolen MS code? Try ver 5 .... much better quality than 3

 

[CougTek]

That's a good baseline for an article, but you need to add some meat around it (text). Only posting numbers will only be useful for people knowing how to interpret them. For noobs or for people simply not familiar with RAM benchmarking, it will be meaningless.

You could add a detailed description of each benchmarking application you used (in the article, not in the forum). Many parts posted by CityK are worth being included in the article too. Overall, it will give the impression that the author actually knows his stuff, not only how to run and copy/paste benchmark numbers.

I would have no problem with an article signed by two persons : Honold and CityK.

 

[CityK]

It should be set to tRCD + CL + (BL/2) ... where BL is the burst length, which = 8 with DDR.

honold, someone on SR this morning linked to a memory article to Tom's (which turned out to be the usual sugar coated dribble fraught with many inaccuracies that we've come to expect from that site). However, (sadly) it was this source which reminded me of the chipsets ability to issue an early precharge ....so long as tRAS isn't set to more then two clock cycles before the completion of tRCD + CL + (BL/2) there shouldn't be any possibilities of data corruption.

-----

I'm glad you guys appreciate the memory related stuff I've posted. I was thinking that perhaps I could string together a couple of relevant outlines if people are interested. I would keep it point/bullet form, but it would be full sentences and completely comprehensible....writing it out in paragraph form would likely take forever. I'm thinking offhand that a lot of what I would include in the outline(s) would simply be cut and pastes from some online articles - I simply don't have the time to rewrite all the material in my own words, and secondly, there is certainly no need to reinvent the wheel when someone else got it right the first time.

Now some may balk at this approach, but it is my feeling that ordering and making sense out of the readibly available information on the web is what stands between most computer enthusiasts from pocessing a good overview of how (SDRAM) system memory works. Let me put it this way; there is good info out there, but the problem is that the info is all over the place, comes in bits and pieces from articles written over the last several years, isn't always clearly written, is often accompanied with completely inaccurate info etc etc. I think that I could order, highlight and clarify all these important bits and aspects, and bring it to light for everyone.

The outline I would prepose would be like thus:

- SDRAM Chip and Module basics and packaging - applies to both SDR and DDR
- Introduce the a SDR "page miss" memory access, with a discussion which outlines where and how the various timings (CAS etc) and issues (setup time etc) fit into the picture
- Introduce the other types of SDR memory accesses
- Based on the different types of memory accesses, develop a performance model outlining the latency-bandwidth connection
- Discuss the difference brought in with DDR
- Look at the different memory accesses types in relation to DDR
- Develop the latency-bandwidth performance models for DDR
- Maybe talk performance of applications under different configurations...or what should be expected from changing this or that.
- Talk about new technologies (PAT, Hammer) ... [something I would have to read up on and then relate to memory, cause I'm pretty ignorant of them right now, but have a good feel for how they fit into the performance model]
- Talk about the future of DDR

One thing from the above should be clear though - it might take a while to complete the whole project ... absolutely no guarentees of how quickly I could get the pieces out....suffice it to say, it would be an "over the course of the summer" type thing.

CK

 

[honold]

BTW, isn't DivX 3.11 the version that was based on stolen MS code? Try ver 5 .... much better quality than 3

it's still probably the most common one in use. i'm not encoding for quality, just for metrics.

 

[honold]

That's a good baseline for an article, but you need to add some meat around it (text). Only posting numbers will only be useful for people knowing how to interpret them. For noobs or for people simply not familiar with RAM benchmarking, it will be meaningless.

well, that's still assuming it would be made an article here, which is not acknowledged or confirmed

i'm going to add details but that's as far as i got last night. as far as how detailed i will go, it probably won't be much, just specifics on how i ran the tests (what patch levels, what settings, etc)

I would have no problem with an article signed by two persons : Honold and CityK.

neither would i but i want to post it reasonably soon

as far as the 'education' stuff, i think it's great, but not necessary at this point. most overclockers only want to know what the benchmarks are, and what settings impact them. if i do enough to encompass general use i shouldn't have to go into explanations on what type of memory usage patterns are affected by what timings, and instruct them how to determine this and then adjust accordingly. my intent is for it to be a rough guide, and to post it while the facts are highly relevant.

 

[honold]

[One thing from the above should be clear though - it might take a while to complete the whole project ... absolutely no guarentees of how quickly I could get the pieces out....suffice it to say, it would be an "over the course of the summer" type thing.

i would love to see this

 

[CityK]

Its too bad we couldn't do something academic like Markowitz's Efficient Frontier (it looks at risk vs reward in terms of diversified portfolio's .... which directly parallel's the cost vs performance theme here in terms of Fq and settings)....just thinking how computer performance theory is still very much in its infancy, or at least we aren't looking in the right places for it

i want to post it reasonably soon...while the facts are highly relevant

Certainly don't wait for me

most overclockers only want to know what the benchmarks are, and what settings impact them.

Yep, probably little desire for the details.

CK