Investigations into Athlon X2 Overclocking
by Jarred Walton on December 21, 2005 12:00 PM EST- Posted in
- CPUs
RAM Latency
In the Venice article, we talked quite a bit about memory latencies and how they affect performance. Since we're using four different types of RAM this time, we will most likely see more variation in RAM latencies. However, even a relatively large difference of 20% will often have less than a 5% impact on real world application performance. Another item to pay attention to is the difference between higher bandwidth vs. lower latencies that the OCZ VX and EL Platinum represent.
CPU-Z includes a latency.exe program that we used to get these values, and we selected the score in the bottom-right of the table. This position represents a 32MB data set with a 512 KB stride size, and the results are reported in CPU cycles. We have two charts again, the first in CPU cycles and the second in nanoseconds. Since cycle time decreases as clock speed increases, we would pay a bit more attention to the results in nanoseconds. However, the elapsed CPU cycles are also important as they represent wasted CPU time. If cache misses are relatively constant - and they are - then the higher the RAM latency is in CPU cycles, the less efficient the processor becomes. Here are the results.
Since all of the various configurations ran successfully with 1T timings, latencies stay relatively close - even the value RAM isn't that far behind the fastest RAM used. We couldn't get results for 4x512MB except at stock clock speeds, but comparing the scores at 10x200 is still possible. You can clearly see how much the 2T command rate impacts performance, which brings up another point: a lot of value RAM is not guaranteed to run at 1T command rates. This same value RAM required 2T command rate with the Venice chip for the 9x300 setting, even though it was still running below its rated DDR400 speed. Your mileage may vary.
Something else that we failed to mention last time was the numerous in-between options for RAM. 2.5-3-3-7/8 RAM can be purchased for around $80, give or take. In the $100 to $130 price range, there are many sets of 2x512MB DIMMs rated for 2-3-3-6-1T or 2-3-2-5-1T timings. The difference between CL2 and CL2.5 isn't huge, but it may warrant spending an extra $20. The true performance/overclocking RAM starts at around $120 (Crucial Ballistix - really an awesome deal if you ask me!) and goes up from there to over $200. That's a dramatic price increase on the high-end for what is likely to be a small performance difference, but there is something to be said for the ease of use that expensive RAM offers. Where we had to do some work finding the optimal performance settings on value RAM at the various overclocked speeds (and we really only scratched the surface), the OCZ EL Platinum could usually be left at a DDR-400 setting in the BIOS with 2.5-3-3-7-1T timings. It wasn't optimal in all cases, but it required very little effort to find a stable overclock.
The OCZ VX is definitely held back by the Infinity motherboard. 3.2V is the minimum required voltage for stable 2-2-2-8-1T timings. We were able to reach as high as DDR-452 speeds while maintaining 2-2-2-8 timings, but DDR-460 required 2-3-2-8-1T and DDR-500 required 2-3-3-8-1T timings. Beyond DDR-500, even raising CL to 2.5 didn't help much. OCZ VX is still very fast RAM, but it really needs the best in terms of motherboards to reach its full potential. In the RAM area, we now have two strikes against our selected motherboard.
In the Venice article, we talked quite a bit about memory latencies and how they affect performance. Since we're using four different types of RAM this time, we will most likely see more variation in RAM latencies. However, even a relatively large difference of 20% will often have less than a 5% impact on real world application performance. Another item to pay attention to is the difference between higher bandwidth vs. lower latencies that the OCZ VX and EL Platinum represent.
CPU-Z includes a latency.exe program that we used to get these values, and we selected the score in the bottom-right of the table. This position represents a 32MB data set with a 512 KB stride size, and the results are reported in CPU cycles. We have two charts again, the first in CPU cycles and the second in nanoseconds. Since cycle time decreases as clock speed increases, we would pay a bit more attention to the results in nanoseconds. However, the elapsed CPU cycles are also important as they represent wasted CPU time. If cache misses are relatively constant - and they are - then the higher the RAM latency is in CPU cycles, the less efficient the processor becomes. Here are the results.
Since all of the various configurations ran successfully with 1T timings, latencies stay relatively close - even the value RAM isn't that far behind the fastest RAM used. We couldn't get results for 4x512MB except at stock clock speeds, but comparing the scores at 10x200 is still possible. You can clearly see how much the 2T command rate impacts performance, which brings up another point: a lot of value RAM is not guaranteed to run at 1T command rates. This same value RAM required 2T command rate with the Venice chip for the 9x300 setting, even though it was still running below its rated DDR400 speed. Your mileage may vary.
Something else that we failed to mention last time was the numerous in-between options for RAM. 2.5-3-3-7/8 RAM can be purchased for around $80, give or take. In the $100 to $130 price range, there are many sets of 2x512MB DIMMs rated for 2-3-3-6-1T or 2-3-2-5-1T timings. The difference between CL2 and CL2.5 isn't huge, but it may warrant spending an extra $20. The true performance/overclocking RAM starts at around $120 (Crucial Ballistix - really an awesome deal if you ask me!) and goes up from there to over $200. That's a dramatic price increase on the high-end for what is likely to be a small performance difference, but there is something to be said for the ease of use that expensive RAM offers. Where we had to do some work finding the optimal performance settings on value RAM at the various overclocked speeds (and we really only scratched the surface), the OCZ EL Platinum could usually be left at a DDR-400 setting in the BIOS with 2.5-3-3-7-1T timings. It wasn't optimal in all cases, but it required very little effort to find a stable overclock.
The OCZ VX is definitely held back by the Infinity motherboard. 3.2V is the minimum required voltage for stable 2-2-2-8-1T timings. We were able to reach as high as DDR-452 speeds while maintaining 2-2-2-8 timings, but DDR-460 required 2-3-2-8-1T and DDR-500 required 2-3-3-8-1T timings. Beyond DDR-500, even raising CL to 2.5 didn't help much. OCZ VX is still very fast RAM, but it really needs the best in terms of motherboards to reach its full potential. In the RAM area, we now have two strikes against our selected motherboard.
Facebook
Twitter
RSS
46 Comments
View All Comments
rekabwolrab - Friday, February 24, 2006 - link
I'm new to OC and both the articles were very nice. Good Job. I am looking forward to the next installment with HSF/Cooling.shoeish - Friday, February 24, 2006 - link
Any results to share about watercooling or stock HSF with this chip yet?mcpdigital - Wednesday, December 28, 2005 - link
This article comes in the right moment since lots of people are thinking about upgrading their PCs or just did it.In my case I have a pretty simmilar configuration with LanParty Ultra-D, 3800 X2 and OCZ EL 3200 (2x1GB).
I found the breakeven of my setup at 280x9(2520), Mem at 210 MHz CAS 2,3,3,5 1T and HT x 3. Memory is running at its best, with 1T, Fastest in BIOS and CAS2 achieving around 61000 MB/s transfer rate running Sandra 2005 Pro, a value that is a bit under the maximum bandwidth with HT @1680 MHz of 6720MB/s
Anything over this speed makes the system unstable and requires a lot of slowdowns in other settings, voltage and temps raising fast, its a bad tradeoff IMO.
So Anandtech simple of the 3800 X2 seems a little better than mine, not that I'm not happy, I'm for sure.
Marcelo
Some1ne - Sunday, December 25, 2005 - link
Re: If you have any specific requests or suggestions before then, let me know.I noticed that as you increased the clock speed, you also increased your chipset voltage in a fairly linear way. I question whether or not this is really necessary or beneficial. I have a MSI Neo4 Platinum mainboard, and I've never had to touch the chipset voltage when overclocking. In fact, some of the behavior I observed when playing with it seemed to imply that the chipset got slightly less stable with higher voltages (though I didn't do enough testing to know conclusively if the relationship holds or not). Using the stock chipset voltage, I was able to hit:
2464 MHz (352x7) on a Winchester 3000+ w/ 6.6% over-VID on the CPU
2420 MHz (242x10) on a Manchester 3800+ w/ 10% over-VID on the CPU
2400 MHz (400x6) on a Winchester 3000+ w/ 6.6% over-VID on the CPU, just to see if the board would run stably at a 400 MHz "fsb" setting...it did
So as far as I can tell, boosting the chipset voltage is not necessary in order to attain a good overclock. It might be interesting if you could do tests to see what, if any, impact it has on stability at higher clock speeds, or maybe at least re-run your 2.7 GHz tests with stock chipset voltage just to make sure that your instability wasn't coming from an overheating chipset.
JarredWalton - Monday, December 26, 2005 - link
The results reported are only after testing all of the lower voltages. I encountered instability without the increased voltage to the chipset and processor. That said, other motherboards may not behave the same. I intend to switch to a different motherboard for the cooling tests -- a DFI LanParty SLI-DR. I will be sure to comment on whether the voltage requirements change or not.AtaStrumf - Friday, December 23, 2005 - link
Just want to commend you for a really thorough article. I miss that from other AT editors as of late.I also agree that all that ultra high end memory with tight timings is an absolute overkill for all but the most rabid overclockers. This is especially true since Athlon got an on die memory controller and became Athlon64. Just get some good quality RAM that will get you to 220-233 MHz so you have some headroom with BIOS FSB/dividers settings, because generic usually craps out at 201-203 MHz (sad but true).
Visual - Thursday, December 22, 2005 - link
Fantastic article, folks!It really showed alot. Sure, as someone commented, using a better mobo might have been interesting... but after all its the CPU that is important here, and you made the differences in performace with varying oc well presented.
I have to say, this article showed a surprisingly high difference between memory types too. You did comment in the end that there wasn't much difference, but there are some cases where there is :) 3dMark05 is the extreme case i guess, and not "real world" enough to be worth the added price, but 15fps or more in a lot of games from going from generic to the PC4800 mem isn't bad too. Seriously, this article showed the importance of memory way clearer than any of your RAM roundups in the past.
What is still dissapointing is that the test didn't reach the near-3ghz ocs a lot of people are bragging with on some forums :p But this is a good thing in a way, as now there won't be any misled readers buying the chip and expecting unrealistic achievments. I'm still curious about what the chips can do at max though, so I'm looking forward to your stock/Chill tests :) Maybe comparison with both infinity and lanparty boards? Maybe trying out several chips so you can give us a somewhat more realistic max average oc? (Hehe, no, scratch that last one. I don't want AT going broke from buying out all the X2s, plus no matter how many chips you test, the readers' own luck will deviate from yours)
Visual - Thursday, December 22, 2005 - link
Oh hey, I want to add a bit but there is no edit feature. So here goes...The RAM difference is much higher than with the singlecore veince. This does match with the assumption that two cores would need (and benefit) more bandwidth. So it also brings hope that the move to AM2 and DDR2 will have an even further boost, atleast for the dualcores. I'm already drooling over an imaginary AM2 X2 oced with DDR2 800mhz ram or faster :p
JarredWalton - Thursday, December 22, 2005 - link
I would say the performance difference shown here (relative to Venice) is from two things. First, two cores can use more bandwidth, though most of these tests won't show that since they're single-threaded. Second, the faster graphics card allows the CPU to really stretch its legs.Once you're at realistic settings for this system (minimum 1280x1024 resolution), the scores get a lot closer. Also, 3DMark has a pretty large deviation between runs - probably 3% or so. I didn't run 3DMark multiple times looking for the best score, so the results may not present a completely accurate representation of performance. Still, the CPU tests do show generic RAM at a pretty major disadvantage as clock speed increases. If 3DMark05's CPU test is an accurate estimate of multithreaded game performance, we're looking at a 25% difference! But I wouldn't put too much stock in 3DMark05. :p
Visual - Friday, December 23, 2005 - link
From what I read on the futuremark forums once, even though 3dmark05 is multithreaded, vertex processing in cpu tests is singlethreaded (some dx9 functionality from MS, not developed by futuremark) so isnt taking full advantage of dualcores still.