- Joined
- Sep 27, 2002
- Location
- Alberta
Edit: You don't really need to read this whole thing, all the important conclusions are at the end in red and green. Read the couple of lines above that and the colorful stuff.
I'm sure all of you have come to a point where you had to choose between a higher FSB or tighter timings and many people passionately argue in favour of one or the other. I decided to test it out for myself and decide what was best for me. Keep in mind, these are my results and may or may not be representative of what happens to everyone, but at least I've provided some data.
Some things to note (you don't really have to read these unless you're interested):
1) I ran several tests including PCMark, all the 3D Marks and Aquamark. All the latest versions.
2) The baseline was set as the average of 3 runs of default timings explained below, the rest of the results are the percentage of the default. That is, I took the results divided them by the default average and multiplied by 100.
3) Each batch of tests were run off a fresh reboot in the exact same order and same amount of time between each individual test.
4) Because each run took upwards of half an hour an there were over 20 of them, I only ran each on once. You may argue that they may not be statistically accurate, but I wasn't about to double or triple the time it took to make this. I think it represents real life well enough to give you a picture of what's happening.
5) You'll notice that most of these differences are very small and I've started the graph at 90% to exaggerate the difference. If the graph started at 0% they would not even be noticible because most difference are less than 1%
6) Noting the above, a difference of 1% may just be due to random variance and not to any real differences in performance.
7) While I was running the 3D marks, I didn’t set up my system to get the highest possible 3D mark. Instead I set the settings at what I usually play them at. 8X anisotropic filtering and performance as opposed to quality based rendering. This may create a video memory bottleneck, but I'd rather have results representative of what I actually play at.
8) The graph is long and complicated, but gives a complete set of results. I think one of the most important bars is the average bar at the top of each set. This is essentially an average difference off all the results (a combination of PCmark and gaming marks). Some games go up, some games go down, but if you look at the average it should give you a good overall view of things.
9) Noting the above, the average score did not include the grammar check scores of PCmark because I found them to be too randomly variable. They would fluctuate independently of what other scores were doing and I felt that they were throwing off the average. The total PCmark however still incorporates the grammar check, an I've also included it in the bar graphs, just not in the average.
With that being said:
System Specs:
Asus A8N SLI Deluxe Bios 1006
Athlon64 3000+ @ stock speed (1.8Ghz)
512 Megs DDR Kingston HyperX (default timings defined below)
MSI NX6600 GT
480 Watt Antec True Blue
36 Gig raptor
200 Gig Barracuda PATA
Sound Blaster Audigy 2
And a bunch of other peripherals attached.
System settings:
Windows XP SP2 (with all the latest patches)
Forceware Drivers 71.84 (8X aniso, performance rendering, no AA, no v-sync, most other things were defaults)
nForce Drivers 6.39
Audigy Drivers 5.12.2.444
No Anti-virus running at the time of the tests
All the latest versions of the tests that I ran.
Default memory timings:
1T timing
Tcl=2
Trc=11
Trfc=13
Trcd=2
Trrd=2
Tras=5
Trp=2
Twr=3
Trwt=4
Since I couldn't find a reference for the RAM that I had that actually listed the default advanced timings, I just went with what A64tweaker told me were the defaults. It runs stably at these settings.
Interpreting the results. This is best done by example. If result is titled Trwt_6, this means that it was changed from the default setting of 4 to 6. Most of the tests were ran with at least a change of 2, unless it was impossible, eg. Tcl can never be changed to 4 so I went with the highest settings. Most tests were changed to a higher number so that the timings were slower, unless it was already at the highest setting, in which case it was changed down. This was done for stability reasons.
Now without further ado, here are the results (click on the pictures to get a full res version, only 70kb):
Conclusions:
Well as you can see for yourself (look at the data chart for the real numbers), most of these differences were negligible by themselves. This is both in terms of application and game performance. Most differences were less than 1%, however, if you combine a few of them and get the tightest possible timings, you'd be lucky to get a 2-3% difference. In almost all cases, raising the FSB by 5 MHZ (which is a 45MHz increase in CPU frequency, and 10MHz increase in RAM frequency), was better than getting tighter timings. The only exception is going from 1T timing to 2T timing which gave an average difference of almost 3%!
Totally surprising are the results of overclocking the PCI express bus. People have reported that this gives a performance increase, but in my results, overclocking it from 100MHz to 120MHz gave the biggest drop in performance, with almost 4%! I guess this is to be strongly avoided. I'm not sure why this happened but it's a huge difference.
Other important timings are Trcd, Trp and Tcl.
Also included are some other settings that may be unique to the Asus A8N, such as PEG buffer length and speed. I'm not sure what these do exactly so I decide to test them out as well. I think they have something to do with SLI or the PCIE bus, but since I'm not running SLI I didn't think they would matter. The only one that seemed to make a difference was the PEG faster mode. It increased performance, unsurpsingly mostly in games, and mostly in 3Dmark 2003 and 2005. If people only look at those results it would look like it makes a huge difference but if you factor in the other real work game tests, it only makes a minor difference. Another setting that made a big difference in the 3Dmark03 and 05 was the video memory overclock. While it didn't make a huge difference in other games, these synthetic benchmarks improved substantially, suggesting that there was a memory bottleneck, but also demonstrating that these benchmarks exaggerate performance gains. As always, take caution when interpreting synthetic benchmarks.
Interestingly, none of the settings seemed to change the aquamark minimum framerate, but rather only affected the average, which kind of sucks because minimum framerates are what you notice most.
As a final conclusion, I will list in descending order, the factors that made the largest difference in performance. Those that increased performance are in green and those that decreased performance are in red. Beside each one is the absolute value of the difference from the default (average values only). Once again the color shows the increase or decrease.
PCIe_120MHz 4.33080537 (Unexplainable how overclocking this bus lowered results so much)
Timing_2T 2.91497771 (Biggest difference, if going to 2T is going to give you less than 10MHz (extrapolating), then it may not be worth it)
FSB_5MHz 1.6649575 (Loosen any one timing if it means getting more than a 5MHz increase in FSB)
Trcd_4 1.42644412 (Surprisingly, this made the biggest difference, more than twice that of Tcl which everyone seems to think makes the biggest difference)
Trc_13 0.7977236
Trp_4 0.71380812
Tcl_3 0.64054618 (Looking at only timings, this falls in 4th place of importance, so go ahead, loosen it if you have to)
PEG_Faster 0.430825287 (This had quite a large effect on performance, even more so than the core or the memory combined)
Trwt_6 0.38369667
Twr_2 0.23116357 (Remember, I lowered the timing to this value because the default is 3)
PEG_BufferLong 0.200149743 (Normally it's set to default, which I guess defaults to short, because when set to short it had no significant difference).
Memory_50MHz 0.1536281 (A small impact, and this average is largely influenced by the 3D mark scores, so I would say that this had essentially no effect on real game performance. Surprising but those are the results)
Trfc_15 0.11591604 (At this point of less than 0.2% difference I would say that it is not a statistically significant change, although I didn't do the statistics to prove it. That would explain why we see an increase in performance by loosening this timing)
Core_5MHz 0.0821403 (Once again, statistically insignificant, and once again surprising. I guess there is a bottleneck somewhere in my system and it doesn't seem to be the video core)
PEG_Buffer_Short 0.06031109 (This is probably what it defaults to because there is almost no difference)
Well I guess that's all. I will address questions and issues as they come up. I didn’t comment on stability because that's far too variable to make generalization based on two sticks of ram I tested. People will have to test that for themselves, but I just hope that this will help people who have to make the hard decision choosing higher FSB or tighter timings.
I just thought I'd make my contribution to the forums and I've found them to be a great resource and full of helpful people and tips.
Enjoy.
I'm sure all of you have come to a point where you had to choose between a higher FSB or tighter timings and many people passionately argue in favour of one or the other. I decided to test it out for myself and decide what was best for me. Keep in mind, these are my results and may or may not be representative of what happens to everyone, but at least I've provided some data.
Some things to note (you don't really have to read these unless you're interested):
1) I ran several tests including PCMark, all the 3D Marks and Aquamark. All the latest versions.
2) The baseline was set as the average of 3 runs of default timings explained below, the rest of the results are the percentage of the default. That is, I took the results divided them by the default average and multiplied by 100.
3) Each batch of tests were run off a fresh reboot in the exact same order and same amount of time between each individual test.
4) Because each run took upwards of half an hour an there were over 20 of them, I only ran each on once. You may argue that they may not be statistically accurate, but I wasn't about to double or triple the time it took to make this. I think it represents real life well enough to give you a picture of what's happening.
5) You'll notice that most of these differences are very small and I've started the graph at 90% to exaggerate the difference. If the graph started at 0% they would not even be noticible because most difference are less than 1%
6) Noting the above, a difference of 1% may just be due to random variance and not to any real differences in performance.
7) While I was running the 3D marks, I didn’t set up my system to get the highest possible 3D mark. Instead I set the settings at what I usually play them at. 8X anisotropic filtering and performance as opposed to quality based rendering. This may create a video memory bottleneck, but I'd rather have results representative of what I actually play at.
8) The graph is long and complicated, but gives a complete set of results. I think one of the most important bars is the average bar at the top of each set. This is essentially an average difference off all the results (a combination of PCmark and gaming marks). Some games go up, some games go down, but if you look at the average it should give you a good overall view of things.
9) Noting the above, the average score did not include the grammar check scores of PCmark because I found them to be too randomly variable. They would fluctuate independently of what other scores were doing and I felt that they were throwing off the average. The total PCmark however still incorporates the grammar check, an I've also included it in the bar graphs, just not in the average.
With that being said:
System Specs:
Asus A8N SLI Deluxe Bios 1006
Athlon64 3000+ @ stock speed (1.8Ghz)
512 Megs DDR Kingston HyperX (default timings defined below)
MSI NX6600 GT
480 Watt Antec True Blue
36 Gig raptor
200 Gig Barracuda PATA
Sound Blaster Audigy 2
And a bunch of other peripherals attached.
System settings:
Windows XP SP2 (with all the latest patches)
Forceware Drivers 71.84 (8X aniso, performance rendering, no AA, no v-sync, most other things were defaults)
nForce Drivers 6.39
Audigy Drivers 5.12.2.444
No Anti-virus running at the time of the tests
All the latest versions of the tests that I ran.
Default memory timings:
1T timing
Tcl=2
Trc=11
Trfc=13
Trcd=2
Trrd=2
Tras=5
Trp=2
Twr=3
Trwt=4
Since I couldn't find a reference for the RAM that I had that actually listed the default advanced timings, I just went with what A64tweaker told me were the defaults. It runs stably at these settings.
Interpreting the results. This is best done by example. If result is titled Trwt_6, this means that it was changed from the default setting of 4 to 6. Most of the tests were ran with at least a change of 2, unless it was impossible, eg. Tcl can never be changed to 4 so I went with the highest settings. Most tests were changed to a higher number so that the timings were slower, unless it was already at the highest setting, in which case it was changed down. This was done for stability reasons.
Now without further ado, here are the results (click on the pictures to get a full res version, only 70kb):
Conclusions:
Well as you can see for yourself (look at the data chart for the real numbers), most of these differences were negligible by themselves. This is both in terms of application and game performance. Most differences were less than 1%, however, if you combine a few of them and get the tightest possible timings, you'd be lucky to get a 2-3% difference. In almost all cases, raising the FSB by 5 MHZ (which is a 45MHz increase in CPU frequency, and 10MHz increase in RAM frequency), was better than getting tighter timings. The only exception is going from 1T timing to 2T timing which gave an average difference of almost 3%!
Totally surprising are the results of overclocking the PCI express bus. People have reported that this gives a performance increase, but in my results, overclocking it from 100MHz to 120MHz gave the biggest drop in performance, with almost 4%! I guess this is to be strongly avoided. I'm not sure why this happened but it's a huge difference.
Other important timings are Trcd, Trp and Tcl.
Also included are some other settings that may be unique to the Asus A8N, such as PEG buffer length and speed. I'm not sure what these do exactly so I decide to test them out as well. I think they have something to do with SLI or the PCIE bus, but since I'm not running SLI I didn't think they would matter. The only one that seemed to make a difference was the PEG faster mode. It increased performance, unsurpsingly mostly in games, and mostly in 3Dmark 2003 and 2005. If people only look at those results it would look like it makes a huge difference but if you factor in the other real work game tests, it only makes a minor difference. Another setting that made a big difference in the 3Dmark03 and 05 was the video memory overclock. While it didn't make a huge difference in other games, these synthetic benchmarks improved substantially, suggesting that there was a memory bottleneck, but also demonstrating that these benchmarks exaggerate performance gains. As always, take caution when interpreting synthetic benchmarks.
Interestingly, none of the settings seemed to change the aquamark minimum framerate, but rather only affected the average, which kind of sucks because minimum framerates are what you notice most.
As a final conclusion, I will list in descending order, the factors that made the largest difference in performance. Those that increased performance are in green and those that decreased performance are in red. Beside each one is the absolute value of the difference from the default (average values only). Once again the color shows the increase or decrease.
PCIe_120MHz 4.33080537 (Unexplainable how overclocking this bus lowered results so much)
Timing_2T 2.91497771 (Biggest difference, if going to 2T is going to give you less than 10MHz (extrapolating), then it may not be worth it)
FSB_5MHz 1.6649575 (Loosen any one timing if it means getting more than a 5MHz increase in FSB)
Trcd_4 1.42644412 (Surprisingly, this made the biggest difference, more than twice that of Tcl which everyone seems to think makes the biggest difference)
Trc_13 0.7977236
Trp_4 0.71380812
Tcl_3 0.64054618 (Looking at only timings, this falls in 4th place of importance, so go ahead, loosen it if you have to)
PEG_Faster 0.430825287 (This had quite a large effect on performance, even more so than the core or the memory combined)
Trwt_6 0.38369667
Twr_2 0.23116357 (Remember, I lowered the timing to this value because the default is 3)
PEG_BufferLong 0.200149743 (Normally it's set to default, which I guess defaults to short, because when set to short it had no significant difference).
Memory_50MHz 0.1536281 (A small impact, and this average is largely influenced by the 3D mark scores, so I would say that this had essentially no effect on real game performance. Surprising but those are the results)
Trfc_15 0.11591604 (At this point of less than 0.2% difference I would say that it is not a statistically significant change, although I didn't do the statistics to prove it. That would explain why we see an increase in performance by loosening this timing)
Core_5MHz 0.0821403 (Once again, statistically insignificant, and once again surprising. I guess there is a bottleneck somewhere in my system and it doesn't seem to be the video core)
PEG_Buffer_Short 0.06031109 (This is probably what it defaults to because there is almost no difference)
Well I guess that's all. I will address questions and issues as they come up. I didn’t comment on stability because that's far too variable to make generalization based on two sticks of ram I tested. People will have to test that for themselves, but I just hope that this will help people who have to make the hard decision choosing higher FSB or tighter timings.
I just thought I'd make my contribution to the forums and I've found them to be a great resource and full of helpful people and tips.
Enjoy.
Last edited:
