Investigations into Socket 939 Athlon 64 Overclocking
by Jarred Walton on October 3, 2005 4:35 PM EST- Posted in
- CPUs
System Settings
We have a bunch of screenshots from CPU-Z showing the CPU and Memory tabs, covering most of the settings that we used. Rather than linking 44 images, though, we're just going to provide a single Zip file of all the screens. One thing that became immediately clear is that the BIOS voltages were almost never reflected in the CPU-Z results. Which one is more accurate is impossible to say, short of busting out a voltmeter (and knowing where to attach it).
We did not remember to get a screenshot of every single configuration tested, since we went back to fill in the blanks on CPU performance after running the initial benchmarks. However, you can get the settings used in the following table. If you have a motherboard that doesn't support the same settings that we used, you may or may not be able to reach a specific overclock.
Disclaimer: Many of the tested voltages on the CPU are probably higher than necessary. After trying for 10x280 with up to the maximum voltage possible from the motherboard, I was probably a bit too lenient on turning voltages back to normal. These are more or less the settings I used during the testing - there may be a few errors in record keeping. If you are looking for long-term stability and you can get the system to run stable at 1.450V instead of 1.650V, that would be a wise decision. The results in the following table are merely intended as an initial reference point.
Note how CPU voltages scaled rapidly as we neared the highest overclock levels. We didn't spend a lot of time trying to get things running stably at a lower voltage level, so mostly, we went in .05V increments - again, you might be able to get better results. If we experienced a crash during our benchmarking, we would try to increase the CPU and/or chipset voltage to get the tests to run stable. If that didn't work, we resorted to tweaking memory timings, generally by increasing latencies until we found a stable setting. Once we went from CL2 to CL2.5, we didn't spend the time trying to get 2.5-2-2, 2.5-3-2, or anything other than 2.5-3-3 (or higher latencies) to run stably.
With our performance RAM, we kept it at a steady 2.8V setting. We did try 2.9V on some of the higher overclocks, particularly where we had to drop from the PC3200 to PC2700, but we couldn't get 1T timings at PC3200 above a 280 MHz CPU bus speed. The value RAM was kept at a steady 2.6V setting and 2.5-3-3-8-1T timings, except in a few cases where we had to run with 2T timings. We tried to get 3-4-4-8-1T instead, but at 9x300, we could not run the value RAM without the 2T setting.
You'll notice the "crash" and "unstable" comments on several of the highest overclock attempts. "Crash" means that we were unable to run many of the tests due to repeated lockups, reboots, etc. "Unstable" means that we were able to get benchmark results for all (or nearly all tests), but programs might crash at times. For example, Far Cry might crash at 1024x768 4xAA on the first attempt, but rebooting and starting again from that point would complete the tests. We tried to run all of the gaming benchmarks in order without rebooting, which will keep system temperatures higher than letting the GPU cool down for a couple of minutes while we reboot. We won't include the settings that crashed in our results, but we did include the unstable results. We'll be using these unstable settings for some cooling tests in the future to see if a change in HSF will help - and hopefully even allow higher overclocks.
A last comment is that we didn't fully benchmark all of the settings listed in the charts. We tested 1800, 2000, 2200, 2400, 2600, and 2700 MHz. We also tested 2800 MHz on a couple of configurations, although stability was iffy at best. In order to provide a linear scale (so that the results at 2700 aren't skewed), we interpolated the in-between scores. This is a problem with the graphing capability that we have within Excel. We did run some quick tests at each setting, though, just to verify that we could POST and complete PCMark04/PCMark05. In case you're wondering, the entire benchmark suite takes around 4 to 5 hours to complete. That will hopefully explain why we didn't run the additional tests or spend a lot of time fine-tuning each tested setting.
And now, on with the benchmarks.
We have a bunch of screenshots from CPU-Z showing the CPU and Memory tabs, covering most of the settings that we used. Rather than linking 44 images, though, we're just going to provide a single Zip file of all the screens. One thing that became immediately clear is that the BIOS voltages were almost never reflected in the CPU-Z results. Which one is more accurate is impossible to say, short of busting out a voltmeter (and knowing where to attach it).
We did not remember to get a screenshot of every single configuration tested, since we went back to fill in the blanks on CPU performance after running the initial benchmarks. However, you can get the settings used in the following table. If you have a motherboard that doesn't support the same settings that we used, you may or may not be able to reach a specific overclock.
Disclaimer: Many of the tested voltages on the CPU are probably higher than necessary. After trying for 10x280 with up to the maximum voltage possible from the motherboard, I was probably a bit too lenient on turning voltages back to normal. These are more or less the settings I used during the testing - there may be a few errors in record keeping. If you are looking for long-term stability and you can get the system to run stable at 1.450V instead of 1.650V, that would be a wise decision. The results in the following table are merely intended as an initial reference point.
Click to enlarge.
Note how CPU voltages scaled rapidly as we neared the highest overclock levels. We didn't spend a lot of time trying to get things running stably at a lower voltage level, so mostly, we went in .05V increments - again, you might be able to get better results. If we experienced a crash during our benchmarking, we would try to increase the CPU and/or chipset voltage to get the tests to run stable. If that didn't work, we resorted to tweaking memory timings, generally by increasing latencies until we found a stable setting. Once we went from CL2 to CL2.5, we didn't spend the time trying to get 2.5-2-2, 2.5-3-2, or anything other than 2.5-3-3 (or higher latencies) to run stably.
With our performance RAM, we kept it at a steady 2.8V setting. We did try 2.9V on some of the higher overclocks, particularly where we had to drop from the PC3200 to PC2700, but we couldn't get 1T timings at PC3200 above a 280 MHz CPU bus speed. The value RAM was kept at a steady 2.6V setting and 2.5-3-3-8-1T timings, except in a few cases where we had to run with 2T timings. We tried to get 3-4-4-8-1T instead, but at 9x300, we could not run the value RAM without the 2T setting.
You'll notice the "crash" and "unstable" comments on several of the highest overclock attempts. "Crash" means that we were unable to run many of the tests due to repeated lockups, reboots, etc. "Unstable" means that we were able to get benchmark results for all (or nearly all tests), but programs might crash at times. For example, Far Cry might crash at 1024x768 4xAA on the first attempt, but rebooting and starting again from that point would complete the tests. We tried to run all of the gaming benchmarks in order without rebooting, which will keep system temperatures higher than letting the GPU cool down for a couple of minutes while we reboot. We won't include the settings that crashed in our results, but we did include the unstable results. We'll be using these unstable settings for some cooling tests in the future to see if a change in HSF will help - and hopefully even allow higher overclocks.
A last comment is that we didn't fully benchmark all of the settings listed in the charts. We tested 1800, 2000, 2200, 2400, 2600, and 2700 MHz. We also tested 2800 MHz on a couple of configurations, although stability was iffy at best. In order to provide a linear scale (so that the results at 2700 aren't skewed), we interpolated the in-between scores. This is a problem with the graphing capability that we have within Excel. We did run some quick tests at each setting, though, just to verify that we could POST and complete PCMark04/PCMark05. In case you're wondering, the entire benchmark suite takes around 4 to 5 hours to complete. That will hopefully explain why we didn't run the additional tests or spend a lot of time fine-tuning each tested setting.
And now, on with the benchmarks.
Facebook
Twitter
RSS
101 Comments
View All Comments
intellon - Tuesday, October 4, 2005 - link
I understand how/why the memory quality is not too imoprtant (5-9% increase for 100 bucks = not worthy)What I AM unclear about is the cpu itself. Would all the cpu's based on venice hit a same ceiling. Or would a 3800+ reach a higher, more stable, cooler overclock than the 3200+? There is one line that mentions these two cpu's on the first page but no comment on how they would perform when overclocked. Does a 12x help over 9x? Also am I wrong in assuming that you picked 3200+ over 3000+ because of a higher multiplier?
And like people are asking... how bad/good are the other chips? How'll a San Diego 3500+ fare against a Venice 3500+? They're faster as stock, but can they match or exceed overclock performance of venice?
Questions questions questions...
The article was wicked though. I was skeptical about buying a cheaper RAM... but seeing how another $50 is not going to help, I'll save that money for something else.
gplracer - Tuesday, October 4, 2005 - link
Very nice article. It appears to be well thought out. Thanks for the time you spent on it. I would also be nice to have an article of this type with some of the more popular power supplies.I to have had several chips that would overclock such as:
P166 @ 200mhz lol
Celeron 300a @ 450mhz
Duron 600 @ 950mhz
Athlon 1700+ (DLTC3) @ 2374mhz
2600+ at 250x10= 2500mhz
There is no way you could add all of the cpus to the review. I look forward to overclocking a dual core athlon64.
PaBlooD - Tuesday, October 4, 2005 - link
Great Articule.. thanks for that great work.I actually have a A64 3200+ Winchester core with an Epox 9NDA3+ + 512 x2 ocz premier (crap ) and i only can get the procesor to 2150 mhz... i tried with safe memo times.. but nothing..are that bad overclockers the Winchester cores? :S
(excuse my poor english ^_^)
RaulAssis - Wednesday, December 21, 2005 - link
Didi you try memory deviders like 5/6 ?yacoub - Tuesday, October 4, 2005 - link
I definitely appreciate all the walk-through of overclocking an A64 system. Very good article. One thing though - the last few pages with the test result charts... the charts make it look like the entire notion of overclocking is rather pointless since all four colored lines are nearly identical in all but a couple tests. You might want to consider a different type of chart next time that gives a -visual- impression of the benefit to better support the written descriptive increases in performance. Maybe some sort of bar chart would have worked better.JarredWalton - Tuesday, October 4, 2005 - link
I felt the visual impression conveyed exactly what I saw: the difference between the 3000+ and 3200+ in overclocking combined with value and performance RAM is, at best, small. I understand what you're saying, and trust me: I played around with the Excel graphs for many hours. None of the graphs really gave a clear picture, unfortunately. Getting four setups with about 9 settings each into a single chart is messy. Having 80 charts is even worse. Heheh.If someone can show me a preferred chart style, I'll be happy to change the graph for the next installment. The AnandTech graphing engine really wasn't capable of dealing with this type of data set, unfortunately... but Excel was only marginally better.
intellon - Tuesday, October 4, 2005 - link
I guess you could "ZOOM IN" onto the y-axis. For instance: on the last graph HL2 1024x768 4xAA, since the minimum was above 80 and max was below 140, you could set the min and max ranges of y-axis accordingly. or go GNU plot way for a sharper graph.JarredWalton - Tuesday, October 4, 2005 - link
Like the 3DMark GPU scores? I really dislike graphs that don't start at 0, because it hides the reality. (That's why I put the extra paragraph on the 3DMark scores noting specifically that they don't start at 0.) I can blow up a graph so that everyone can see the 1 or 2% margin of victory, but what does that really say? Margin of error on several benchmarks is at least 1 or 2%, and in actual use I don't think anyone will really notice even a 5% difference - I know I don't.Some people will be annoyed by this, but too many people worry about the last 1% of performance. Not because they can notice a difference, but because they want meaningless bragging rights. Sitting in the top positions in an online game requires skill. Getting 1% higher FPS usually just involves throwing more money at your PC than the next guy. Some people like to do that - sort of like some people like muscle cars. I want a fast computer, but I'm not going to lose sleep because my PC is marginally slower than my friend's, you know?
Anyway, I may look into a separate graphing tool. Excel looks fine internally, but getting the graphs into image form didn't work perfectly. The text alignment got a little tweaked when I cut and pasted the data into Photoshop.
Regards,
Jarred Walton
RupertS - Wednesday, October 19, 2005 - link
Be careful, I think Muscle Car owners are a protected class.probedb - Tuesday, October 4, 2005 - link
I'd just like to say cheers for this. It's made me finally get round to trying to OC my system. I purposely bought a 3000+ and Crucial Ballistix for this but have never got round to trying it.I shall give it a go this weekend!!!