ps- at that time, let us know what components your using..or even in advance..that'll help you get going a bit faster when the time comes if you need help.
graysky, nice to see you again!
I noticed that Kris used your x264 benchmark in the qx9650 overclocking article that was posted on the front page of anandtech a few weeks back. congratulations buddy you made the front page
I noticed that Kris used your x264 benchmark in the qx9650 overclocking article that was posted on the front page of anandtech a few weeks back. congratulations buddy you made the front page
In support of Graysky's endeavor I ran the following 3-way benchmark test:
FSB @ 1333 with DDR2-667 (1:1)
FSB @ 1333 with DDR2-800 (5:6)
FSB @ 1600 with DDR2-800 (1:1)
I used my multiplier unlocked QX6700 and vapophase ls cooling to run the benchmarks at 4.0GHz for all tests. (A 4GHz CPU is more likely to deplete its cache during processing and become dependant on accessing the ram than a processor operating at lower speeds.)
Ram was optimized for tightest timings possible for the ram at each given speed (per memtest+ 1.70).
I used synthetic benchmarks and compared to a couple real-app programs that I use. (my life is dominated by computational chemistry (gaussian98) and financial modeling (metatrader 4.0))
I have a spreadsheet of the results, but no idea how or where to upload it and the post formating in these forums is little to be desired so I just did a screen capture of the spreadsheet:
Benchmark results in spreadsheet (screen capture)
The results are essentially that I saw ~4% improvement in synthetic benches by increasing the memory bandwidth from 1:1 to 5:6 while keeping the FSB at 1333, but only 0.8% improvment in my apps of interest (no games, sorry).
I saw an additional ~4% improvement in synthetic benches by increasing the FSB bandwidth from 1333 to 1600 while keeping the ram at DDR2-800, but only 0.4% improvment in my apps of interest (no games, sorry).
Overall, in taking my system from a 1333MHz FSB with DDR2-667 ram to a 1600MHz FSB with DDR2-800 ram the synthetic benchmarks improve 8.3% (that's marketing material folks) but my applications of interest are clearly CPU/cache bound and only show an improvement of 1.2%.
Obviously in my case further system bus increases (or ram bandwidth increases) are pointless unless they are transient effects of me wringing more CPU speed out of the system.
Notes: Originally I intended to also test 800, 1066 and 2000 FSB speeds, but my crappy Super Talent ram would not post at DDR2-1000 no matter how loose the timings or how high the voltage and for some reason my QX6700 on the Asus P5E WS Pro would not post with any CPU multiplier >14...needed a 15x for 1066 FSB and 20x for 800 FSB.
FSB @ 1333 with DDR2-667 (1:1)
FSB @ 1333 with DDR2-800 (5:6)
FSB @ 1600 with DDR2-800 (1:1)
I used my multiplier unlocked QX6700 and vapophase ls cooling to run the benchmarks at 4.0GHz for all tests. (A 4GHz CPU is more likely to deplete its cache during processing and become dependant on accessing the ram than a processor operating at lower speeds.)
Ram was optimized for tightest timings possible for the ram at each given speed (per memtest+ 1.70).
I used synthetic benchmarks and compared to a couple real-app programs that I use. (my life is dominated by computational chemistry (gaussian98) and financial modeling (metatrader 4.0))
I have a spreadsheet of the results, but no idea how or where to upload it and the post formating in these forums is little to be desired so I just did a screen capture of the spreadsheet:
Benchmark results in spreadsheet (screen capture)
The results are essentially that I saw ~4% improvement in synthetic benches by increasing the memory bandwidth from 1:1 to 5:6 while keeping the FSB at 1333, but only 0.8% improvment in my apps of interest (no games, sorry).
I saw an additional ~4% improvement in synthetic benches by increasing the FSB bandwidth from 1333 to 1600 while keeping the ram at DDR2-800, but only 0.4% improvment in my apps of interest (no games, sorry).
Overall, in taking my system from a 1333MHz FSB with DDR2-667 ram to a 1600MHz FSB with DDR2-800 ram the synthetic benchmarks improve 8.3% (that's marketing material folks) but my applications of interest are clearly CPU/cache bound and only show an improvement of 1.2%.
Obviously in my case further system bus increases (or ram bandwidth increases) are pointless unless they are transient effects of me wringing more CPU speed out of the system.
Notes: Originally I intended to also test 800, 1066 and 2000 FSB speeds, but my crappy Super Talent ram would not post at DDR2-1000 no matter how loose the timings or how high the voltage and for some reason my QX6700 on the Asus P5E WS Pro would not post with any CPU multiplier >14...needed a 15x for 1066 FSB and 20x for 800 FSB.
@Idontcare - nice job with this dude. Glad to see folks expanding on what I started here and that your data trends are consistent with what I found
graysky seriously dude I was astounded. I saw your results and naturally thought "interesting, but what if you actually increased the FSB and memory but kept same clockspeed?" Surely that would bring real-performance to the table, right?
Nope, not for me it didn't, I was truly astounded by the data.
A year ago I spent a ridiculous amount of money on Mushkin redline DDR2-1066 ram so I could get extra performance from my system. Wish I had known back then that it was a waste of money for my apps perspective.
Now with the Tomhardware review showing that PCIe bandwidth does modulate gaming performance, I wonder if anyone here will take advantage of this to overclock their PCIe busspeed from 100MHz to improve FPS in games.
Nope, not for me it didn't, I was truly astounded by the data.
A year ago I spent a ridiculous amount of money on Mushkin redline DDR2-1066 ram so I could get extra performance from my system. Wish I had known back then that it was a waste of money for my apps perspective.
Now with the Tomhardware review showing that PCIe bandwidth does modulate gaming performance, I wonder if anyone here will take advantage of this to overclock their PCIe busspeed from 100MHz to improve FPS in games.
n7
Elite Member
- Jan 4, 2004
- 21,281
- 4
- 81
I benched this same kinda stuff (1:1 vs. 4:5, etc.) way back with my P5B-D + E6300 about 1.5 yrs ago, but AT lost the thread somehow, so i can't link to it.
RAM does indeed make a small difference.
One thing people seem to forget is that those small differences add up.
I saw people ask why DDR2 over DDR, & now we ask why DDR3 over DDR2?
The reality is, even though the improvements are small, they do add up.
That being said, spending double on high end RAM vs. mid range isn't usually worth it...
RAM does indeed make a small difference.
One thing people seem to forget is that those small differences add up.
I saw people ask why DDR2 over DDR, & now we ask why DDR3 over DDR2?
The reality is, even though the improvements are small, they do add up.
That being said, spending double on high end RAM vs. mid range isn't usually worth it...
BonzaiDuck
Lifer
- Jun 30, 2004
- 16,689
- 2,069
- 126
I get along with people on these forums, and I may "know some things," but I don't know all the things there are to know.
Someone had asked "how" or "why (you would want)" to run the memory at a higher bus speed than the motherboard FSB. Someone -- maybe it was Graysky(the Guru -- no joke) -- noted that the FSB has always been a bottleneck.
There are memory operations that take place without pushing a lot of data through the CPU. There is a feature that has been around a long time now, called "Direct Memory Access" [thus, motherboards have "DMA" settings.]
Assuming that data is being transferred to a device (i.e., hard disk) or processed through the CPU (requiring the bottlenecked FSB), there are going to be unused clock cycles when memory is run at a ratio different from 1:1.
On the DDR2 angle, the "skinny" is that latency settings "aren't as important as they were with DDR," but they still buy bandwidth. Also, voltage increases for memory are not only inevitable as you push up the FSB; they are likely required when tightening the latencies at any given FSB speed. I've also discovered that there are greater voltage requirements when running "1:1" than with running a ratio of 4:5. So -- at 1:1 and lower FSB with tighter latencies that you would think more accommodating to the lower FSB, the voltage requirement I've seen to be 0.025 to 0.050V higher at the low FSB, tight latency setting. With a 4:5 ratio and the "tightest" latencies I can get to achieve the same or higher synthetic bandwidth, I'm running a significantly higher memory FSB, looser latencies (true), but lower voltage. I can tweak the combination of memory FSB, latencies and voltage to get a higher bandwidth result -- possibly with 0.025+V lower than a lower bandwidth result at lower FSB and tighter latencies.
You can think of this as a big stairway. There is a range of FSB speeds with any given CPU : RAM ratio for a particular regime of latency settings. As you move up the FSB and attempt to keep the same latency settings, it requires more voltage. At some point, there is a threshold where you have to loosen the latency settings to go higher, or you push the voltage over the manufacturer maximum spec.
So before I go further and get another nomination for "member most likely to write a book when a few sentences will do --" --
The synthetic bandwidths are one basis of comparison to "keep score" on how you're doing. Even though they're synthetic, they have led me to settings where I can get marginal improvements in actual game-play -- better scores -- even with CPU : RAM ratios that are not 1:1.
1:1 is an ideal. But you impose your own limits as to how high you'll push the CPU VCORE (and risk early mortality), how high you'll push the VDIMM (risking "no replacement" under manufacturer limited lifetime warranty), and how far you'll push the motherboard, chipset, HT/MCH voltage etc. -- and risk a dead circuit board with that. So at some point, you may decide to use a different ratio -- given some remarks or observations I made above, and the limitations of your parts and budget.
While I don't have the time now, I think there's a logical explanation why some CPU : RAM ratios are better than others, as mentioned in a recent Anandtech article on OC'ing the QX9650. I'm guessing that it has to do with the integer arithmetic of latencies and data-bursts and the number of bytes processed in any given memory operation -- so that some ratios optimize the movement of data more than others, even if the ratios aren't "1-to-1."
Somebody else can explore this or correct me, and I'd be interested in reading their thoughts on the matter.
Someone had asked "how" or "why (you would want)" to run the memory at a higher bus speed than the motherboard FSB. Someone -- maybe it was Graysky(the Guru -- no joke) -- noted that the FSB has always been a bottleneck.
There are memory operations that take place without pushing a lot of data through the CPU. There is a feature that has been around a long time now, called "Direct Memory Access" [thus, motherboards have "DMA" settings.]
Assuming that data is being transferred to a device (i.e., hard disk) or processed through the CPU (requiring the bottlenecked FSB), there are going to be unused clock cycles when memory is run at a ratio different from 1:1.
On the DDR2 angle, the "skinny" is that latency settings "aren't as important as they were with DDR," but they still buy bandwidth. Also, voltage increases for memory are not only inevitable as you push up the FSB; they are likely required when tightening the latencies at any given FSB speed. I've also discovered that there are greater voltage requirements when running "1:1" than with running a ratio of 4:5. So -- at 1:1 and lower FSB with tighter latencies that you would think more accommodating to the lower FSB, the voltage requirement I've seen to be 0.025 to 0.050V higher at the low FSB, tight latency setting. With a 4:5 ratio and the "tightest" latencies I can get to achieve the same or higher synthetic bandwidth, I'm running a significantly higher memory FSB, looser latencies (true), but lower voltage. I can tweak the combination of memory FSB, latencies and voltage to get a higher bandwidth result -- possibly with 0.025+V lower than a lower bandwidth result at lower FSB and tighter latencies.
You can think of this as a big stairway. There is a range of FSB speeds with any given CPU : RAM ratio for a particular regime of latency settings. As you move up the FSB and attempt to keep the same latency settings, it requires more voltage. At some point, there is a threshold where you have to loosen the latency settings to go higher, or you push the voltage over the manufacturer maximum spec.
So before I go further and get another nomination for "member most likely to write a book when a few sentences will do --" --
The synthetic bandwidths are one basis of comparison to "keep score" on how you're doing. Even though they're synthetic, they have led me to settings where I can get marginal improvements in actual game-play -- better scores -- even with CPU : RAM ratios that are not 1:1.
1:1 is an ideal. But you impose your own limits as to how high you'll push the CPU VCORE (and risk early mortality), how high you'll push the VDIMM (risking "no replacement" under manufacturer limited lifetime warranty), and how far you'll push the motherboard, chipset, HT/MCH voltage etc. -- and risk a dead circuit board with that. So at some point, you may decide to use a different ratio -- given some remarks or observations I made above, and the limitations of your parts and budget.
While I don't have the time now, I think there's a logical explanation why some CPU : RAM ratios are better than others, as mentioned in a recent Anandtech article on OC'ing the QX9650. I'm guessing that it has to do with the integer arithmetic of latencies and data-bursts and the number of bytes processed in any given memory operation -- so that some ratios optimize the movement of data more than others, even if the ratios aren't "1-to-1."
Somebody else can explore this or correct me, and I'd be interested in reading their thoughts on the matter.
lopri
Elite Member
- Jul 27, 2002
- 13,314
- 690
- 126
Heh.. have you guys tried running memory slower than FSB.. Granted I know that there aren't many boards that even allow to do so today, but if you can - try it. Or you could try single-channel instead of dual-channel. Both cases affect the theoretical bandwidth significantly, and the picture will be quite different. I do understand the purpose of this thread (and what OP tries to prove), but what we see is how inefficient FSB and Intel chipsets in general. (especially their 'dividers')
But overall it is true that memory speed/latency doesn't mean much as long as it's >= FSB. Especially so for bigger L2s.
Or you could try single-channel instead of dual-channel. Both cases affect the theoretical bandwidth significantly, and the picture will be quite different. I do understand the purpose of this thread (and what OP tries to prove), but what we see is how inefficient FSB and Intel chipsets in general. (especially their 'dividers')But overall it is true that memory speed/latency doesn't mean much as long as it's >= FSB. Especially so for bigger L2s.
Originally posted by: lopri
Heh.. have you guys tried running memory slower than FSB.. Granted I know that there aren't many boards that even allow to do so today, but if you can - try it.
But overall it is true that memory speed/latency doesn't mean much as long as it's >= FSB. Especially so for bigger L2s.
Lopri check my post above with the link in it. Note I compared 1333 vs 1600 FSB for a 4GHz quad-core with either DDR2-667 DDR2-800 in each case.
What could possibly more FSB/bandwidth limited than a 4GHz quad-core? Sure the synthetics show an 8% improvement, but my handful of real applications show 1% improvement.
Sure I'll take a 1% improvement if its free and is simply a matter of "turning it on" but it is not something worth pursuing with money and time. (for me, and my specific apps)
That's all I am getting at. Why did Intel take the FSB to 1333 for Wolfdales and Yorfields when 1066 would have likely continued to be just fine? For all I can tell it is just to ensure the mobo and ram guys are included in the upgrade cycle. Got to wonder if Yorkfield would have had any issue running on i680 at 1066 FSB...
BonzaiDuck
Lifer
- Jun 30, 2004
- 16,689
- 2,069
- 126
Well, Idontcare, (supposedly, by the official specs) "some" 680i boards -- you know what I mean -- some of them are spec'd to run at 1,333, and we've pretty much proven that revisions to the BIOS have a degrading effect on performance of earlier processors. That is, maybe the over-clocking experience also speaks to the compatibility issue.
See, I've read where 680i die-hards bought QX9650 processors and discovered they "won't run" on their motherboards (and you know exactly which motherboard I mean, don't you?) I don't think anyone bothered to report which BIOS revision they were using.
So all this Yorkfield speculation -- I guess it's up in the air, or it is for me until I've stumbled on more definitive information (which may already "be there," but I haven't stumbled onto it. . . . )
what I mean -- some of them are spec'd to run at 1,333, and we've pretty much proven that revisions to the BIOS have a degrading effect on performance of earlier processors. That is, maybe the over-clocking experience also speaks to the compatibility issue.See, I've read where 680i die-hards bought QX9650 processors and discovered they "won't run" on their motherboards (and you know exactly which motherboard I mean, don't you?) I don't think anyone bothered to report which BIOS revision they were using.
So all this Yorkfield speculation -- I guess it's up in the air, or it is for me until I've stumbled on more definitive information (which may already "be there," but I haven't stumbled onto it. . . . )
Engineer
Elite Member
- Oct 9, 1999
- 39,230
- 701
- 126
This thread inspired me to try something on my E6300 with Asrock 775Dual-VSTA board. The maximum ram speed supported by the board is DDR2-667, which is what the memory was running at (5-5-5-15 2T at 667 with a 4:5 FSB:MEM divisor). I lowered the memory to 266MHz which automaticaly set it to 4-4-4-12 2T at 533MHz with a 1:1 ratio. The ram speed (SiSoft Sandra) went from 4,164MB/s to 4,750MB/s. I'll leave it at the lower setting. More bandwidth and lower power can't be beat!
(No difference in other tests like 3DMark06, etc).
I know that this doesn't really pertain to the OP, but since this thread was the inspiration, I posted my results here.
(No difference in other tests like 3DMark06, etc).
I know that this doesn't really pertain to the OP, but since this thread was the inspiration, I posted my results here.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 23K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter