Investigations into Socket 939 Athlon 64 Overclocking
by Jarred Walton on October 3, 2005 4:35 PM EST- Posted in
- CPUs
RAM Latency
We talked about memory latency a bit before, and we estimated the total latency. Using CPU-Z and the included Latency.exe utility, we can get an actual real value for memory latency - or at least, more real than our estimate. There are different ways of measuring memory latency, but we simply took the highest reported value from the table that was generated. For every system, that value was in the bottom-right corner of the table, indicating a data set of 32MB and a stride of 512. While the values may or may not be entirely accurate, they should at least be consistently measured across all the tested configurations. Here are the results, in CPU cycles as well as nanoseconds. Remember that at higher CPU speeds, each cycle is faster, so pay more attention to the nanoseconds graph.
While the value RAM is clearly slower than the expensive RAM, the big question is: how much performance do you actually lose by opting for more economical RAM? This is one of the areas that we are interested in testing for this article, and we'll comment on the results throughout the benchmarks.
RAM Possibilities
One area of the BIOS that's missing (for now) is support for additional RAM ratios. The reality is that the RAM speed is derived from the CPU clock with a divider. Ideally, we'd like to see a BIOS that gives direct access to the CPU divider rather than hiding it behind approximate memory speeds. That would allow for the selection of a larger range of options, but we're not sure if that's something that AMD controls in the CPU or if the BIOS programmers can do this. Here's a list of the standard RAM Settings, CPU multipliers and the resultant memory dividers that were available in the Infinity BIOS.
We have a LanParty UT nF3 250Gb that includes support for many in-between options from DDR200 through DDR500. Some of the interesting inclusions are DDR240, DDR280, DDR300, DDR360, DDR440, and DDR500. (CPU-Z actually failed to report the divider on a couple of those settings, but the resultant RAM speed was still read properly.) While additional memory dividers on the high end won't really help tweakers looking to get the most from the TCCD, BH5, or CH5 RAM, they can be particularly useful when using value RAM. All you need to do is select the appropriate divider to get your RAM under DDR400 - assuming standard value RAM. That way, you wouldn't take as much of a performance hit by running something like 10x250 (2.5 GHz). You could select a 13X divider rather than the standard 15X divider.
Additional ratios can even be useful for tuning performance RAM. For example, OCZ VX would not run stable above DDR500 (10x250), and we had to use 2-3-3-8-1T timings even then. (The 3.2V limit of the motherboard was at least partially to blame.) Running at 10x270, we had to drop to DDR333 (CPU/12), which resulted in the RAM running at DDR450 rather than the DDR540 result that would have been required for the normal CPU/10 ratio. However, if we could have selected a CPU/11 ratio, we could have run the RAM at DDR490 and gained a bit more performance. The additional ratios aren't required, but they would be nice to have.
If the last two paragraphs didn't make sense, then you can guess why we don't get additional access to RAM dividers. Experienced users might know how to make use of the settings, but many people would simply get confused; a "Catch-22".
We talked about memory latency a bit before, and we estimated the total latency. Using CPU-Z and the included Latency.exe utility, we can get an actual real value for memory latency - or at least, more real than our estimate. There are different ways of measuring memory latency, but we simply took the highest reported value from the table that was generated. For every system, that value was in the bottom-right corner of the table, indicating a data set of 32MB and a stride of 512. While the values may or may not be entirely accurate, they should at least be consistently measured across all the tested configurations. Here are the results, in CPU cycles as well as nanoseconds. Remember that at higher CPU speeds, each cycle is faster, so pay more attention to the nanoseconds graph.
While the value RAM is clearly slower than the expensive RAM, the big question is: how much performance do you actually lose by opting for more economical RAM? This is one of the areas that we are interested in testing for this article, and we'll comment on the results throughout the benchmarks.
RAM Possibilities
One area of the BIOS that's missing (for now) is support for additional RAM ratios. The reality is that the RAM speed is derived from the CPU clock with a divider. Ideally, we'd like to see a BIOS that gives direct access to the CPU divider rather than hiding it behind approximate memory speeds. That would allow for the selection of a larger range of options, but we're not sure if that's something that AMD controls in the CPU or if the BIOS programmers can do this. Here's a list of the standard RAM Settings, CPU multipliers and the resultant memory dividers that were available in the Infinity BIOS.
| Standard AMD Memory Ratios | |||
| RAM Setting | CPU Multiplier | RAM Divider | True RAM Speed |
| DDR200 | 9X | CPU/18 | DDR200 |
| DDR266 | 9X | CPU/14 | DDR257 |
| DDR333 | 9X | CPU/11 | DDR327 |
| DDR400 | 9X | CPU/9 | DDR400 |
| DDR200 | 10X | CPU/20 | DDR200 |
| DDR266 | 10X | CPU/15 | DDR267 |
| DDR333 | 10X | CPU/12 | DDR333 |
| DDR400 | 10X | CPU/10 | DDR400 |
| DDR200 | 11X | CPU/22 | DDR200 |
| DDR266 | 11X | CPU/17 | DDR259 |
| DDR333 | 11X | CPU/14 | DDR314 |
| DDR400 | 11X | CPU/11 | DDR400 |
| DDR200 | 12X | CPU/24 | DDR200 |
| DDR266 | 12X | CPU/18 | DDR267 |
| DDR333 | 12X | CPU/15 | DDR320 |
| DDR400 | 12X | CPU/12 | DDR400 |
We have a LanParty UT nF3 250Gb that includes support for many in-between options from DDR200 through DDR500. Some of the interesting inclusions are DDR240, DDR280, DDR300, DDR360, DDR440, and DDR500. (CPU-Z actually failed to report the divider on a couple of those settings, but the resultant RAM speed was still read properly.) While additional memory dividers on the high end won't really help tweakers looking to get the most from the TCCD, BH5, or CH5 RAM, they can be particularly useful when using value RAM. All you need to do is select the appropriate divider to get your RAM under DDR400 - assuming standard value RAM. That way, you wouldn't take as much of a performance hit by running something like 10x250 (2.5 GHz). You could select a 13X divider rather than the standard 15X divider.
Additional ratios can even be useful for tuning performance RAM. For example, OCZ VX would not run stable above DDR500 (10x250), and we had to use 2-3-3-8-1T timings even then. (The 3.2V limit of the motherboard was at least partially to blame.) Running at 10x270, we had to drop to DDR333 (CPU/12), which resulted in the RAM running at DDR450 rather than the DDR540 result that would have been required for the normal CPU/10 ratio. However, if we could have selected a CPU/11 ratio, we could have run the RAM at DDR490 and gained a bit more performance. The additional ratios aren't required, but they would be nice to have.
If the last two paragraphs didn't make sense, then you can guess why we don't get additional access to RAM dividers. Experienced users might know how to make use of the settings, but many people would simply get confused; a "Catch-22".
Facebook
Twitter
RSS
101 Comments
View All Comments
Crassus - Tuesday, October 4, 2005 - link
First of all, thank you for such a long article. I appreciate the work you put into this. What I'd really like to see in one of the planned articles would be an in-depth coverage of the options an enthusiast-grade mainboard BIOS offers nowadays for the RAM timings (and maybe PCIe) - beyond the standard timings covered in this article.PrinceGaz - Tuesday, October 4, 2005 - link
The finer memory-timings offered by enthusiast mobos are generally vendor specific so your best bet is to check a forum or other site dedicated to your motherboard. For DFI mobos for instance, you can find a thread which gives detailed coverage of memory settings on DFI-Street forums http://www.dfi-street.com/forum/showthread.php?t=2...">hereCheesePoofs - Tuesday, October 4, 2005 - link
Why stability test with 3dmark (an app that tries to stress teh CPU as little as possible) and pcmark (an ok pc-stressing app) instead of the combo of memtest86+, superpi, and prime95? Seems to me that if you want to find out whether yoru CPU really is stable, you'd want to stress it as hard as possible (which those three will do).Also, from what I've read from Zebo's thread in the CPU forums, 2T really doesn't have a significant impact on performance. Could you clarify this?
JarredWalton - Tuesday, October 4, 2005 - link
I've seen systems that run Prime95 and SuperPi 100% stable crash under 3DMark looping, as well as under PCMark. I imagine 2.80 GHz will crash under those if I run them all concurrently. My personal experience is that SuperPi and Prime95 only stress a few paths of the CPU, hence the inclusion of benchmarks with 11 different applications that can all fail with an unstable overclock. 3DMark GPU tests are not as demanding of the CPU, but the CPU tests are very demanding IMO. (That's part of why the top scores on the 3DMark ORB never include the CPU tests.)2T command rate, as you can see in quite a few instances, really killed performance. Perhaps tweaking other special timings beyond CL, tRCD, tRP, and tRAS might make the impact less, but you could likely tweak the same things with 1T at a lower memory speed. Command rate comes into play on every single memory access, so doubling that delay will certainly have an impact on performance.
fitten - Tuesday, October 4, 2005 - link
Good answer. Most have no clue as to how a CPU actually works. Ideally, a synchronous circuit is rated at a clock speed that the longest path will function properly (give correct results). There may be 1000s of pathways that can run at higher frequencies but that one can hold it back. Running the clock rate up may cause that one pathway not to be able to meet something like a data setup and hold time on one line (of the 32 or 64) in the data path and now you have an unstable setup that you may not detect. As always with overclocking, a crash is the best result you can get because you know you've pushed too far. Unless you are testing pretty much every instruction with every possible data against a control to compare against (some pathways can take longer depending on the data that it is being operated on), there are many errors that you may not detect... and all it takes is one, out of the possible billions, to make your machine not stable. Sure, it may be a rarely seen case of instruction+data but it exists.Programs like the Pi calculators and such do make your CPU work a lot, but the calculations are fairly repetitive and hardly a broad sample of the ISA.
I'm all for doing whatever you want with your own machine. Heck, I used to overclock all the time, too. I just find all of the lack of knowledge in synchronous circuits... interesting... when people talk about overclocking.
Saist - Monday, October 3, 2005 - link
for those who read this portion here :****
Because of the GPU limitation, we're going to be testing at 640x480, 800x600, and 1024x768. We'll also test many of the titles with 4xAA enabled, which should serve as a reality check. Even with a super fast CPU, many games are going to be completely GPU limited with the X800 Pro when we run 4xAA, especially at resolutions 1024x768 and above. Frankly, we wouldn't bother enabling 4xAA unless you can at least reach 1024x768 anyway.
****
Did anyone else think... okay.. lets stick a Radeon 9600, GeforceFX, or XGI Volari in there so that we actually will be limited? I mean... please. X800 alone goes above what most users have in their systems today. If we are buying "new" components, then yeah, the X800 is on my short list, but how about doing some reviews over hardware people actually have in their hands.
OvErHeAtInG - Tuesday, October 4, 2005 - link
If you're overclocking a new A64 Venice... somehow I think you're not still running your XGI Volari for games. Remember bench numbers are really only useful if they reflect framerates you would actually want to play with.JarredWalton - Tuesday, October 4, 2005 - link
The reason I used an X800 Pro is because I feel it's a good match for the chip, RAM, and motherboard. I can toss in a 7800GTX to show what the CPU on its own is capable of, but you can get cards that pretty much equal the X800 Pro for under $200. X800 GTO and GTO2 can match and even beat the X800 Pro.I view overclocking (and computer building in general) from a bang-for-the-buck perspective. It doesn't make sense to me to spend $100 upgrading from the 3000+ to the 3500+ if I'm going to be completely GPU limited. $200 on a graphics card is not that much money, when you really get down to it. 180 million transistor chip with 256MB of 980MHz RAM, all mounted on a large PCB? At least I can feel I'm getting a lot of stuff for $200. A CPU is far cheaper to produce (though more expensive to design). Profit margins on CPUs are notoriously high.... Personally, the X800 Pro is a decent card, but I really want something faster these days. Same goes for the 6800GT. But then, not everyone feels that way.
---------
Thought #2 (for Saist): If X800 is above what most people have, other than those buying new computers... well, what about the motherboard and processor? Socket 939 with nForce4 is a more recent configuration than X800/6800 cards. Not to mention Venice has only been out for something like 8 months.
If you're looking to spend $120+ on a new Venice chip and you've only got a 9600 Pro (or even a 9800 Pro), you're wasting your money on the wrong part (at least from a gaming perspective). A socket 754 Sempron with an X800 Pro would be far better for gaming than a Venice core with anything less than an X800/6800. Outside of gaming... well, graphics don't matter outside of gaming much, which is why Winstones, PCMark, and AutoGK are included.
Honestly, I'm not entirely sure if you were complaining about the use of a GPU that was too fast, or that it wasn't fast enough. For frequent gaming, I wouldn't recommend anyone go lower than about the X800 GTO these days. 6600GT is (IMO) now relegated to the budget/moderate-gaming setup, as many games are simply unplayable above 1024x768. I really don't like to drop below 1280x1024/1280x960 if I can avoid it. If I've misunderstood your complaint, let me know; if we simply have a difference of opinion... well, there's not much to do about that. :)
yanman - Tuesday, October 4, 2005 - link
any chance you can add in benches for 7800GT/GTX? after all, in your discussion you correctly asset that money is much better spent on high spec'd GPU to match the cpu speed that you've managed to overclock to - having used bargain rate ram and venice.i have a venice 3000+ clocked at 2686mhz, 7800gt and 2x1gb sticks of average ram (legend/hynix). until i upgraded the ram a few weeks ago i had it running for prehaps a month and a half totally solid with 2x512mb sticks of same type, at 2696mhz (337x8, ram at 225mhz (2:3) 2.5-3-4-7-1T)
the reason i ask for 7800GT and GTX is 2 fold, so we can see it from an nvidia side too (different cpu scaling maybe?), and also to show the scaling for a top-end card even if only as a reference point. It just seems a bit one-dimensional only using 1 card.
One last thing, well done to Zebo who made the excellent "Quick and dirty A64 overclocking guide" (used to be sticky in the forums) which I and many people I know used to overclock their venices with.. i'd be stuck without it!
JarredWalton - Tuesday, October 4, 2005 - link
I'm planning on doing 7800GTX testing with an X2 3800+ OC article. For gaming, it will perform identically to the 3200+ Venice. Hopefully, I'll be done in the next ~week or so.