overclocking and memory divisions

nonstop301

I would like to ask for an opinion with respect to overclocking and the memory dividers used to achieve an improved performance.

Is it more advantageous to increase the FSB frequency with a 5:6 memory divider that can only permit stablity at average memory times (eg. 3-3-3-8), or do you consider as a better alternative to increase the FSB using the highest memory divider eg. 1:2 which may be able to maintain much better memory times (eg 2-2-2-5) ?

I understand that with the 1:2 memory divider it could be possible to push the FSB frequency higher and use much better memory times albeit at a far lower memory frequency. I'm not sure however, if such an overclocking scheme is more beneficial than one that achieves a slightly lower FSB frequency but with the advantage of an improved memory frequency as well, even if the memory timings aren't so remarkable.

Thank you very much for your comments and I'm grateful for any further advice you may wish to provide.

Leonardo

Memory frequency is usually a much bigger factor in performance than memory timings. If you can overclock the memory along with FSB without lowering the timings, then that's the best of both worlds. Unless you've got expensive DRAM though, you'll reach a point where the FSB/memory dividers will have to be employed. (obviously) As you are working up a an overclock, just do some SiSoft benchmarks. Don't use benchmarks alone. The feeling of speed, although sometimes without merit, can be useful too.

nonstop301

Thank you for your reply Leonardo.

Is it correct to conclude that the aim of a sensible overclock is therefore the best combination of FSB frequency and memory frequency without using memory dividers ?

Naturally, with the memory divider the FSB frequency can be raised to much higher values but in terms of overall system performance and stability it seems that it would be more reasonable to reach a high DDR for the memory together with a decent gain GHz for the FSB without applying a memory divider.

For example, if you were to achieve a 0.3 GHz gain in the speed of the processor together with a memory gain from PC 3200 to PC 3900 without applying a memory division, would it be more satisfying than a 0.6 GHz gain for the processor but a memory rise to PC 3600 after applying a memory division ?

The benchmark testing in both instances didn't reveal any dramatic differences between them, although of course the results were a lot better than the values for the processor before overclocking.

I have to also mention that these are my first attempts at overclocking and I'm using a low-budget system with an AMD processor Socket 754 (Sempron64 2600+), nVIDIA nForce3 motherboard (ASUS K8N) and Kingston ValueRAM (1x512 MB) in order to get a better idea of what overclocking is about and whether to consider it with a more advanced system.

edcentric

When you run Sandra use the unbuffered memory test.
There are all kinds of details when you start OCing.
Some systems (this depends on the memory controler) suffer much lower emmory performance if you run anything other than 1:1. In other systems speed is all that really matters, and in some latancey (CAS) is very important.
You need to just play around some. Have fun.

nonstop301

Thank you for your comments edcentric.

I have decided to experiment with the limits of 1:1 overclocking on this computer because my original incentive was to reach the 2 GHz mark and see what benefits this offers before considering the prospect of a higher CPU clock speed using a memory divider.

Fortunately, I managed to achieve 2.0 GHz without the use of a memory divider and the memory frequency is operating at 250 MHz.

This translates into a 0.4 GHz gain from the original 1.6 GHz of the Sempron64 processor, coupled with a 50 MHz rise for the memory frequency from 200 to 250 MHz (DDR 400 to DDR 500).

I will perform some more Sandra benchmarks to see how 2 GHz at DDR 500 (1:1) compares with 2.25 GHz at DDR 450 (5:6) and better memory timings.

Leonardo

Ed's comments were good. Sorry, I painted with a broad brush. Yes, some boards don't do well with dividers other than 1:1. I do know though, that Asus motherboards with Intel 945, 955, 975, and 965 chipsets perform very well with dividers, as if the DRAM and memory were at default settings. Intel has almost always made excellent chipsets. If they CPUs from 2003 - 2005 had been as good as their chipsets, I don't think AMD would ever have taken so much of their market share.

Again, sorry, it's been a long time since I've overclocked AMD systems. With CPUs with a lot of FSB headroom, such as those in my signature, the FSB will very quickly outstrip the DRAM speed for all but the hyper expensive DDR2. There is no way I could get those FSB overclocks without CPU:memory dividers.

nonstop301

I appreciate your remarks Leonardo.

I agree with you entirely about the quality of Intel chipsets and your remark regarding Intel CPUs. I would never have considered buying AMD, if it weren't for the relatively low performance and less attractive pricing of the middle to low range Intel CPUs in particular. But now here I am, toying around with a Sempron64 and an nVIDIA motherboard that offer some overclocking capabilities I wouldn't have expected when I originally purchased them.

I do have a dual core machine as well and that really works immaculately for the purposes it has to serve, so overclocking isn't on the menu for that one at the moment.

I believe I have reached the limit of 1:1 overclocking with the Sempron64 however, and I did perform Sandra benchmarks to compare the rewards, if any, with respect to the benchmarks from the 5:6 overclock I mentioned in my previous post.

The unbuffered memory bandwidth test turned out clearly in favour of the 1:1 oveclock since the DDR 500 I achieved is no match for the DDR 450 despite its better memory timings. DDR 500 recorded 3500 MB/s in both Int and Float whereas the DDR 450 managed 2800 MB/s.

The memory latency test was marginally in favour of the 5:6 overclock which operates at (2.5-3-3-6-1T) as opposed to the 1:1 overclock which can only maintain DDR 500 with 3-4-4-8-2T memory timings.

Finally, the processor benchmarks rewarded the 2.25 GHz CPU Core Speed of the 5:6 overclock with 8100 for the Drystone and 6800 for the Whetstone. The surprising 2.02 GHz I achieved with the 1:1 overclock, scored 7200 and 6100 respectively.

All these results make me feel that it's probably more beneficial to use the 1:1 overclock unless I manage to raise the FSB frequency of the 5:6 further and it can reach a similar memory bandwidth range. The extra 0.2 GHz the 5:6 currently offers isn't making a visible difference with the applications I use on this machine and the 2 GHz at DDR 500 is just as comfortable with the daily tasks I perform with this computer.

The motherboard I'm using (ASUS K8N) only permits a maximum FSB frequency of 300 MHz which means that with the locked processor frequency multiplier of the Sempron64 at 8x I can theoretically achieve 2.4 GHz.

It remains to be seen whether a CPU Core Speed between 2.3 and 2.4 GHz can allow a memory bandwidth for the 5:6 divider that matches the 1:1. I don't have my hopes too high about that however, and it is more likely that to reach the 2.3 GHz range I would have to use a different memory divider, 2:3 or even 1:2. With those memory dividers of course there is no chance of a memory bandwidth that is anywhere near as good as 1:1.

nonstop301

I have decided to settle with the 2 GHz at DDR 500 but I would like to know if there's a way to reduce the memory frequency and retain the current CPU speed without using a divider, so that I can have a chance to improve the memory timings if possible.

As I mentioned above, I achieved 2 GHz with a memory frequency rise from 200 to 250 MHz but in order to maintain stability, it meant that I had to loosen the memory timings significantly (3-4-4-8-2T).

Thanks again for any comments and help you wish to provide.

csimon

Hello nonstop301. The only way I know of to reduce the frequency so that you can achieve tighter timings is to use the mem divider.

There is a free downloadable benchmark program that I use that I highly recommend because of it's small size and ease of use. It's called pcwizard and it is produced by the same people that produce the very popular program cpu-z. I would encourage you to test your current memory timings against some timings that you may experiment with. I think you'll find that with these AMD chips that whether you use tight timings or high frequency you should end up with near same results.

Good luck!

nonstop301

Hi csimon and thank you for your suggestions.

I used the PC Wizard that you mentioned to carry out a few more benchmark comparisons.

I was a little surprised to see that this time the unbuffered memory bandwidth of the DDR 500 overclock at 2 GHz was similar (slightly lower in fact) than the DDR 450 overclock at 2.25 GHz. Originally, the SiSoftware Sandra benchmarks produced a significantly higher bandwidth with the DDR 500 setting (3500 MB/s) and DDR 450 only managed (2800 MB/s).

I can't explain why that was the case but I repeated the SiSoftware Sandra benchmark and it gave me the same values as PC Wizard so I'll just have to settle with that. The only notable thing I did between the original SiSoftware benchmarks and the PC Wizard ones the following day, was a few more hours of Prime05.

Anyhow, this changes my perspective somewhat because I was encouraged by the DDR 500 performance, but as you pointed out it appears that a higher FSB frequency using a memory divider that permits slightly better memory timings despite a lower memory frequency, proves to be just as good if not a better option to choose.

Would you use the lower CPU Core Speed but at a higher DDR without the memory divider, or just go for the higher CPU Core Speed with the lower DDR and 5:6 memory divider if your benchmark comparisons produced similar results for both ?

I'll continue to experiment and see how far the Sempron64 goes on the 5:6 memory divider and with a bit of luck it may get to 2.3 Gz without requiring the 2:3 division.

csimon

nonstop301 wrote:

Would you use the lower CPU Core Speed but at a higher DDR without the memory divider, or just go for the higher CPU Core Speed with the lower DDR and 5:6 memory divider if your benchmark comparisons produced similar results for both ?

Me personally? I always go for the highest stable core speed.
Most applications won't be effected by a little difference in ram speed honestly ...but they mostly do show improvement with higher core speed.

What are your primary applications?

nonstop301

Hi again csimon and thank you for your response.

The software I use mainly on my computers is mostly medically oriented but I also enjoy the features of current peer to peer internet applications that provide improved video streaming and file transfer bandwidths. Perhaps the most demanding applications I have utilised, are in fact these benchmark tools and the rigorous Prime95 to assess the overclocking experiments.

I should mention that I have now implemented the 5:6 divider and this has allowed to the processor to operate at 2.33 GHz with a memory frequency equivalent to DDR 466 (PC 3700).

It feels like this Sempron64 is able to reach 2.4 GHz if I go down to the 2:3 memory divider but I have approached the limits of the recommended HTT bus frequency for Socket 754 with my current settings, so in order to venture further it would be necessary to scale down the HTT multiplier to 2x.

As it is, I'm very satisfied with 2.33 GHz at DDR 466 and admittedly, I didn't imagine that a reliable performance at 2.3 GHz was conceivable from a 1.6 GHz Sempron64 out of all processors. A concession in the memory clock department from 200 MHz to 166 MHz has proved to be beneficial as it is now possible to operate the memory at better timings (2.5-3-3-6-1T) and a much higher CPU Core Speed.

csimon

nonstop301 wrote:

Hi again csimon and thank you for your response.

The software I use mainly on my computers is mostly medically oriented but I also enjoy the features of current peer to peer internet applications that provide improved video streaming and file transfer bandwidths. Perhaps the most demanding applications I have utilised, are in fact these benchmark tools and the rigorous Prime95 to assess the overclocking experiments.

I should mention that I have now implemented the 5:6 divider and this has allowed to the processor to operate at 2.33 GHz with a memory frequency equivalent to DDR 466 (PC 3700).

It feels like this Sempron64 is able to reach 2.4 GHz if I go down to the 2:3 memory divider but I have approached the limits of the recommended HTT bus frequency for Socket 754 with my current settings, so in order to venture further it would be necessary to scale down the HTT multiplier to 2x.

As it is, I'm very satisfied with 2.33 GHz at DDR 466 and admittedly, I didn't imagine that a reliable performance at 2.3 GHz was conceivable from a 1.6 GHz Sempron64 out of all processors. A concession in the memory clock department from 200 MHz to 166 MHz has proved to be beneficial as it is now possible to operate the memory at better timings (2.5-3-3-6-1T) and a much higher CPU Core Speed.

Just out of couriosity ...what is your recommended htt bus limit? What is your current hht bus frequency? cpu mulitplier? etc ...

nonstop301

Here's an outline for you csimon, of the settings my machine is currently using

FSB Frequency: 292 MHz

CPU multiplier: 8x (locked)

HTT multiplier: 3x

HTT bus frequency: 1752 MHz

Recommended HTT range for Socket 754: 1300 - 1800 MHz

Memory divider: 5:6 (333MHz option on the BIOS menu)

Memory Frequency: 233 MHz

Memory timings: CAS (tCL) 2.5 / tRCD 3 / tRP 3 / tRAS 6 / Command Rate 1T

CPU Voltage adjustments aren't permitted on the ASUS K8N motherbaord, so the voltage remains at the default value.

DDR RAM Voltage adjustments are available but I kept the default Auto setting.

csimon

Very nice indeed. It seems you have ballroom on the htt though.
What about the cpu ...what is your current voltage and what is the default for that chip? What are you idle and load temps at that frequency? Are you running Folding@home?

nonstop301

Hello again csimon.

I have now reached the limit of the FSB rise that my motherboard permits, so these are the best values I can reach with this processor/motherboard/RAM combination.

CPU Core Speed : 2.4 GHz

FSB Frequency: 300 MHz

CPU multiplier: 8x (locked)

HTT multiplier: 3x

HTT bus frequency: 1800 MHz (maximum)

Recommended HTT range for Socket 754: 1300 - 1800 MHz

Memory divider: 5:6 (333MHz option on the BIOS menu, CPU/10)

Memory Frequency: 240 MHz

Memory timings: CAS (tCL) 2.5 / tRCD 3 / tRP 3 / tRAS 6 / Command Rate 1T

CPU Voltage (VCore): 1.4 V (default)

DDR DRAM Voltage: 2.5 V (default)

The temperatures with these settings range between 37 - 44 °C depending on what applications I am using. Idle temperature hasn't exceeded 40 °C.

These values have been achieved with the stock coooling, stock PSU and no voltage manipulations so it could be possible that with a motherboard that is considered more overclock friendly, a higher CPU Core Speed can be achieved if it is really necessary.

According to SiSoftware Sandra, the FSB is too high for the recommended the HTT bus speed so I will lower that to maintain better stability.
Similarly, the memory bus speed is reported as higher than the recommended value for the memory and chipset so even though the figures look good, I will also reduce that for a more stable operation.


You also asked me about Folding at Home and I know some details about the purpose of the project, but I haven't yet started or joined any team. I am planning to do so very soon and I will be more than happy to join Team 93.

csimon

I think your processor is very capable of handling the 2.4g/1/4v ...especially at those temps. Sisoft Sandra will report it as being too high though but that's what it's programmed to report. It may even report that at 2.0g idunno.
If your system passes prime95 stress "blend" for at least one full loop I'd say you are set - personally. If it does pass this test ...then you are definitely good to go w/ F@h since it doesn't heat the systems as much as prime95.
I wouldn't concern myself so much with the Sandra report really. Just my humble opinion.

nonstop301

Hi csimon and thanks again for your comments.

At 2 GHz SiSoftware Sandra doesn't produce a warning about the FSB frequency nor the memory frequency (200 MHz with CPU/10 divider, 5:6).

This Sempron64 machine isn't the one I was thinking of in terms of Folding at Home however. I'm just using this one to find out how overclocking is achieved and evaluate the potential risks and benefits.

I have another computer with the Athlon64 X2 and I presume that one can handle Folding at Home more comfortably.

What would you suggest as the minimum hardware requirements for running the Folding at Home software ?

csimon

nonstop301 wrote:

Hi csimon and thanks again for your comments.

At 2 GHz SiSoftware Sandra doesn't produce a warning about the FSB frequency nor the memory frequency (200 MHz with CPU/10 divider, 5:6).

This Sempron64 machine isn't the one I was thinking of in terms of Folding at Home however. I'm just using this one to find out how overclocking is achieved and evaluate the potential risks and benefits.

I have another computer with the Athlon64 X2 and I presume that one can handle Folding at Home more comfortably.

What would you suggest as the minimum hardware requirements for running the Folding at Home software ?

A processor with SSE capability is what I would list as a minimum. As far as frequency goes it depends on what processor we're talking about but I think 2.0 in general unless it can be overclocked higher and 512m of ram.
I'd say with that sempron you're well equipped to do the deed if you want to! Heck I have a 1.4ghz non-SSE right here that's just folding away but I'm somewhat concerned as to whether it is meeting the deadlines or not.

nonstop301

Hi again csimon and thanks for you quick replies.

I'll start off Folding at Home with the Sempron machine then and if I run into difficulties, I'll continue on the other one.

What is SSE capability by the way ?

Secondly, how do a register to be part of Team 93 ?

csimon

nonstop301 wrote:

Hi again csimon and thanks for you quick replies.

I'll start off Folding at Home with the Sempron machine then and if I run into difficulties, I'll continue on the other one.

What is SSE capability by the way ?

Secondly, how do a register to be part of Team 93 ?

I googled wiki for a good definition of SSE ...basically it is an optimization utilized by the gromacs core. The Semprons do have SSE so you're good to go. http://en.wikipedia.org/wiki/Streaming_SIMD_Extensions

To register all you have to do is include team "93" as you configure Folding@home. When you run Folding@home for the first time it will go through a configuration process. The first prompt is for you user name ...the second asks for a team number. Use the same user name and team number on all of your cpus to accumulate and keep track of progress. Setting up a multicore or multicpu is slightly more involved but we can cross that bridge once you master the single cpu configuration.

Here is a link to the SM Prospective Team Member Guide!

nonstop301

Thanks for the helpful information csimon.

I got it all set and it started the folding operation now. It has 500 frames to complete so it will keep churning away till Saturday the 23rd to get it done. I don't know if that is a good rate and whether it meets the deadline however.

I downloaded the Windows graphical client version 5.03 and it seems to be running without any problems. It did have a hiccup when I first launched it but I rebooted and it seems to be in order now.

csimon

nonstop301 wrote:

Thanks for the helpful information csimon.

I got it all set and it started the folding operation now. It has 500 frames to complete so it will keep churning away till Saturday to get it done.

I downloaded the Windows graphical client version 5.03 and it seems to be running without any problems. It did have a hiccup when I first launched it but I rebooted and it seems to be in order now.

Hiccups can be expected ...especially with the graphical client. When you get better aquainted with folding you may find the console version to be slightly more beneficial but you're doing fine for now.
Thanks for listening ...and let me be the first to congratulate you to Team 93!

A better link than the one above about getting started ...http://www.short-media.com/forum/showthread.php?t=29800

nonstop301

Thank you very much csimon

I was intending to contribute to this project and I'm glad to do so through Team 93.

nonstop301

I have another query with regards to overclocking and it's related to the CPU-Z values for HT link.

When I overclocked the Sempron to 2.4 GHz, the HT link frequency rose to 900 MHz. According to the recommendations for the HT bus frequency for a Socket 754 motherboard, the overclock to 300 MHz FSB I attempted produces a 1800 MHz HT bus frequency and that is the maximum recommended for the Socket 754 (1300 - 1800 MHz recommended range).
Is there a risk associated with operating the computer at such a high HT link level, or should I focus on achieving an overclock at HT link values no higher than 800 MHz ?

Currently, like I mentioned in the my previous posts, I'm running the computer at 2 GHz with an HT bus frequency of 1506 MHz (HT link 753 MHz) and Memory frequency 200 MHz.

Thrax

Your computer will eventually be unstable, but that will precede actual damage by a great amount.

csimon

unstable at 1800?

Thrax

No; I mean if he goes higher.

csimon

Thrax wrote:

No; I mean if he goes higher.

That's what I thought!:bigggrin:

nonstop301 wrote:

I have another query with regards to overclocking and it's related to the CPU-Z values for HT link.

When I overclocked the Sempron to 2.4 GHz, the HT link frequency rose to 900 MHz. According to the recommendations for the HT bus frequency for a Socket 754 motherboard, the overclock to 300 MHz FSB I attempted produces a 1800 MHz HT bus frequency and that is the maximum recommended for the Socket 754 (1300 - 1800 MHz recommended range).
Is there a risk associated with operating the computer at such a high HT link level, or should I focus on achieving an overclock at HT link values no higher than 800 MHz ?

Currently, like I mentioned in the my previous posts, I'm running the computer at 2 GHz with an HT bus frequency of 1506 MHz (HT link 753 MHz) and Memory frequency 200 MHz.

You should be good at up to 1800!

nonstop301

Thanks for your remarks Thrax and csimon.

300 MHz is the maximum my motherboard allows me to add to the FSB frequency so that's why I wanted to know about the HT link value and how concerned I should be about its level.

One other matter I wanted to ask about involves SiSoftware Sandra.I ran some benchmarks to compare the memory bandwidth values between two overclocks I achieved and it appears that if I select the buffered option and then repeat the test with the unbuffered option, I receive markedly higher results than a test that's carried out once with the unbuffered option already selected.

I don't know why that is the case and I only ask because it made the comparison I wanted to make slightly more confusing, until I thought of starting with the buffered test and then the unbuffered test when I tested the second overclock too

Thanks again for your thoughts and any suggestions you might have.

csimon

I don't use Sandra anymore but I would think you would run the unbuffered option ...I'm not certain though and I could be wrong.