Memory bandwidth tests... any real differences (PC4300 vs. PC7100)

[graysky]

Common sense tell you that higher memory bandwidth should mean faster results, right? I set out to put this thought to the test looking at just two different memory dividers on my o/c'ed Q6600 system. At a FSB of 333 MHz, the slowest and fastest dividers I could run are:

1:1 a.k.a. PC5300 (667 MHz)
3:5 a.k.a. PC8888 (1,111 MHz)

Just for reference, as they relate to DDR2 memory:

PC4300=533 MHz
PC5300=667 MHz
PC6400=800 MHz
PC7100=900 MHz
PC8000=1,000 MHz
PC8500=1,066 MHz
PC8888=1,111 MHz
PC10600=1,333 MHz

The highest divider is 1:2 aka PC10600 (1,333 MHz) and it just wasn't stable with my hardware @ 333 MHz.

All other BIOS settings were held constant:
FSB = 333.34 MHz and multiplier = 9.0 which gives an overall core rate of 3.0 GHz.
DRAM voltage was 2.25V and timings were 5-5-5-15-4-30-10-10-10-11.

You can think of memory bandwidth as the diameter (size) of your memory's pipe. Quite often, the pipe's diameter isn't the bottle neck for a modern Intel-based system; it is usually much larger than the information flow to/from the processor. Think of it this way, if you can only flush your toilet twice per minute, it doesn't matter if the drain pipe connecting your home to the sewer is 3 inches around, or 8 inches around, or 18 inches around: the rate limiting step in removing water from your home is the toilet flushing/recycling and the pull of gravity, not the size of your drain line. The same is true for memory bandwidth.

After seeing the data I generated on a quad core @ 3.0 GHz, I concluded that this toilet analogy is pretty true: the higher memory bandwidth gave more or less no appreciable difference for real world applications. Shocked? I was.

Further, I should point out that in order for my system to run stable in PC8888 mode @ a FSB of 333, I had to boost my NB vcore two notches and raise my ICH to the max (both of which the BIOS colored red meaning "high risk.") The increased voltage means more heat production, and greater power consumption -- not worth it for small gains realized in my opinion. Anyway, the test details and results are below if you want to read on.

Relevant test hardware:

Motherboard: Asus P5B-Deluxe (BIOS 1215)
CPU: Intel C2Q - Q6600 (B3 revision)
Memory: Ballistix DDR2-1066 (PC2-8500)

"Real-World" Application Based Tests

I chose the following apps: lameenc, x264, winrar, and the trial version of Photohop CS3. I ran these tests on a freshly installed Windows XP Pro SP2 machine.

Lame version 3.97 – Encoded the same test file (about 60 MB wav) with these commandline options:

lame -V 2 --vbr-new test.wav

(which is equivalent to the old –-alt-preset fast standard) a total of 8 times and averaged play/CPU data as the benchmark.

x264 version 0.55.663 – Ran a 2-pass encode on the same MPEG-2 (720x480 DVD source) file 5 times totally and averaged the results. Without getting into too much detail, the benchmark is 1,749 frames @ 23 fps. Based on these numbers, I reported the time it would take to encode 215,784 frames (which is your average 2.5 h of video @ 23 fps). Why did I do this? The differences of just 1,749 frames were too insignificant.

Shameless promotion --> you can read more about the x264 Benchmark at this URL which contains results for hundreds of systems. You can also download the benchmark and test your own machine.

RAR version 3.62 – rar.exe ran my standard backup batch file which generated about 1.09 G of rars (1,654 files totally). Here is the commandline used:

rar a -u -m0 -md2048 -v51200 -rv5 -msjpg;mp3;tif;avi;zip;rar;gpg;jpg  "E:\Backups\Backup.rar" @list.txt

where list.txt a list of all the dirs I want it to back up. Benchmark results are an average of two runs timed with a stopwatch.

Trial of Photoshop CS3 – The batch function in PSCS3 was used to do three things to a total of twenty-nine, 10.1 MP jpeg files:

1) bicubic resize 10.1 MP to 2.2 MP (3872x2592 --> 1800x1200) which is the perfect size for a 4x6 print @ 300 dpi.
2) unsharpen mask filter (60 %, 0.8 px radius, threshold 12)
3) saved the resulting files as a quality 8 jpg.

Benchmark results are an average of two runs timed with a stopwatch.

"Synthetic" Application Based Tests

Just two of these were chosen to illustrate a point about theoretical gains vs. real world gains. Actually, I did SuperPI for the hell of it. WinRAR served to illustrate that point.

SuperPI / mod1.5 XS – The 16M test was run twice, and the average of the two are the benchmark.

WinRAR version 3.62 – If you hit alt-B in WinRAR, it'll run a synthetic benchmark. This was run twice (stopped after 100 MB) and is the average of two runs.

Raw Data - "Real-World" Apps
Lameenc play/cpu (average 8 runs) @ PC5300: 30.7935
Lameenc play/cpu (average 8 runs) @ PC8888: 30.8045
Result: PC8888 is 0.5 % faster

x264 time to encode 2.5 h DVD @ PC5300: 01:48:54
x264 time to encode 2.5 h DVD @ PC8888: 01:46:14
Result: PC8888 is 2.5 % faster

rar.exe back-up (average 2 runs) @ PC5300: 45 sec
rar.exe back-up (average 2 runs) @ PC8888: 44 sec
Result: PC8888 is 2.2 % faster

Photoshop CS3 Trial batch (average 2 runs) @ PC5300: 33 sec
Photoshop CS3 Trial batch (average 2 runs) @ PC8888: 33 sec
Result: PC8888 is 0.0 % faster

So stop right here and ask yourself if a 2-3 % gain is worth the higher voltage and heat.

Raw Data - "Synthetic" Apps

SuperPI/16M test (average 2 runs) @ PC5300: 8 m 8.546 s
SuperPI/16M test (average 2 runs) @ PC8888: 7 m 33.328 s
Result: PC8888 is 7.8 % faster

Winrar internal benchmark (average 2 runs) @ PC5300: 1,515 KB/s
Winrar internal benchmark (average 2 runs) @ PC8888: 2,079 KB/s
Result: PC8888 is 37.2 % faster

...but who uses their system exclusively running internal and synthetic benchmarks? Recall that for my 1.09 gig back up, I only gained about 2 % doing "real work" by using the higher divider. Hardrives are notorious bottle-necks in systems that serve to nullify any memory bandwidth increases. In this case the 37 % theoretical increase was translated into only a 2 % "real world" increase likely due to the hardrive/rar's ability to read/write the data. Again, this seems kinda wasteful to me.

I will admit that there might be special cases where running at high memory dividers may produce more substantial gains: apps such as folding@home or seti@home, etc. may benefit from the higher memory bandwidth since they tend to make exclusive use of the system memory bandwidth and rely much less on the hardrive. I have no data to back-up this though. Also lacking in my experiments are any game data. I'd be interested in knowing if the higher bandwidth can be leveraged by game engines such as UT3, Crysis, etc. but I also didn't look at these here.

Finally, since I held everything else constant, I didn't look at the tighter timings in 1:1 mode that people can often use which may give additional gains. For example, I can get away with 3-3-3-9 @ 1:1 vs. the slower 5-5-5-15 @ 3:5 with this memory.

Anyway, I hope you found this useful and maybe this will inspire someone else to look at the gaps pointed out above (and the gaps I haven't thought of too!)

 

[BillParrish]

appreciable difference for real world applications. Shocked? I was.

Nope, but good investigation and I hope anyone considering dumping a lot of money into very high speed memory reads it.

Keep in mind the sub timings were not optimized (left at one setting) for the different buss speeds. A major performance hit potential. However your testing is still very valid and informative as most users do not know how to effectively tune memory sub-timings. This shows one of the issues with EPP (preprogrammed performance profiles), there are not enough of them. A guick look at that memory on the Curcial site seems to indicate that memory does not use EPP. Further indications that the sub timings are not optimizied. You had a post over in the memory forum where I inquired where those sub timings came from but I must admit, I lost track of it, my bad.

The effect of buffering and wait states needed when running in other than 1:1 will reduce the performance impact/increase some.

The synthetic benchmarks which (well not so much super-pi but it still shows the winrar is more dramatic) show the POTENTIAL for a performance increase. If you were running a transaction based database you would see it. But most of us don't do that or watch/play benchmarks, you made a very good point.

It is no secret and long been established with C2D systems CPU MHz is much much more powerful than memory bandwidth. One should top out their CPU and then manipulate the FSB, CPU multiplier and the memory multiplier to get some acceptable memory speed.

That is the reason I keep pounding on people that for most users buying anything faster than PC6400 (800MHz) 4 4 4 12 memory is a waste and you will never know the difference if you use 5 5 5 15.

Now it is apparent why the new Penryn processors have much larger L2 caches, anytime the CPU can avoid going out to main memory for data, gobs of time are saved. Intel knows the memory subsystem, CPU to memory controller GTL buss based on FSB is now a major bottle neck. Hence the Nehalem with on board memory controller in late 08.

Good stuff, keep up the good work.

 

[graysky]

Thanks for reply and kind words.

 

[JBear]

Sorry for the necropost, but I need to give Graysky credit here. This thread deserves some more attention.

 

[Slade]

It's also best to minimize what variables you have when comparing FSB speeds, I did some testing a few months ago and found real world benefits to an increase in FSB that were a bit more substantial than 1%.

I compared a single change, 400FSB vs 450FSB, identical timings, CPU speed (albeit 8x450 vs 9x400). In this situation where the only real difference was FSB in which I went from 1600mhz memory bandwidth to 1800mhz bandwidth, performance showed an improvement anywhere from 7-10% on apps that could take advantage of a 12% memory bandwidth increase.

Among them were the more recent steam games (CSS Video Stress Test and Lost Coast Test showed an improvement in that range), FEAR, and h264 encoding (turns out pretty much all encoding got a good boost for me).

Since I do play quite a bit of TF2 this helps a bit. It helped in Portal with maxing things out and having a portal looking into a portal and still maintaining a minimum 50fps @19x12. It helps when I encode multiple video files for specific formats to play on my PS3 (wmvDiVX support) or my iPod touch (h264).

I'm sure there's a bit more situations where the bandwidth is beneficial to me, such as photoshop, visual studio or eclipse when compiling my programs and 3D Studio Max when rendering items as traditionally those programs will eat up good amounts of memory when used. Time wise it really adds up for me. Over time, saving a few seconds per compile, or a few mins per render really helps in the long run.