Jump to content
Sign in to follow this  
jubbin2001

Mem Bandwidth Debunk

Recommended Posts

Hey all. I am new to the forums, but have read quite a bit here. One thing that interested me was one of the over-clocking posts (now a sticky) by Angry_Games (yes it is over a year now, I realize that) about memory bandwidth no affecting the outcome of your benchmarks/system performance. Well me being a gamer (and a broke one at that) I decided to really try and look at this. What I found was rather….interesting shall we say. In no way do am I contradicting the previous post, but perhaps suggesting a revamp may be in order. Right now this is just a preliminary finding, and I will be updating as more benchmarks are run.

 

As we know the 8800 GPU has been taking the world by storm. In the rush for the best gaming experience, Vista being released, and the promise of grand things from DX10, people have been upgrading in preparation for games like: Crysis and UT3. I am one of those people. I can’t really afford spending $700 on a GTX, and the difference between the 640mb and 320mb versions of the GS really didn’t seem that big of a deal to me so I went with the 320mb version. I figure by the time I see the benefit of the 640mb there will be something better anyway….it always seems to work that way. In my preliminary findings it would seem that these cards are directly affected on how it uses your RAM Memory bandwidth.

 

On to my system. My current specs are as follows:

 

Asus A8N-SLI Premium

OCZ 550 PSU

AMD A64 x2 4400+ @ 2.684

2GB (4x512) Mushkin Enhanced Performance PC3200 3-3-3-5

BFG Tech 8800GS OC

1x WD Caviar 80gig Ultra ATA-100

1x WD Caviar 240gig Ultra ATA-100

(YES I still rock IDE, since, it has been proven that no FPS jump can be attributed to SATA in any RAID configuration at this time. The performance you get from running SATA in RAID 0 is faster load times…that’s about it. I still have plenty of space left to fill, so no need to upgrade there. Untill I see what these next-gen games are doing and how they tax the system, I will probably keep my drives)

 

Benchmarks:

Sisoft Sandra 2007

Aquamark3

3DMark06 (still more to come on these…ran out of time)

 

I have posted the CPUz screens as well to show that no changes were made to the anything other than the memory divider.

 

I have posted as well the Aquamark screens showing the performance difference between the dividers

 

Also have the 3DMark06 score off the last benchmark running 2.684 with the 1:1 CPU/DRAM config.

 

266:

 

Aquamark: http://www.geocities.com/get_n_jiggy2001/B...arks/266_AQ.jpg

 

CPUz: http://www.geocities.com/get_n_jiggy2001/Benchmarks/CPUz.jpg

http://www.geocities.com/get_n_jiggy2001/Benchmarks/CPUz2

 

Mem Bandwidth: http://www.geocities.com/get_n_jiggy2001/B...m_Bandwidth.jpg

 

333:

 

Aquamark: http://www.geocities.com/get_n_jiggy2001/B...arks/333_AQ.jpg

 

CPUz: http://www.geocities.com/get_n_jiggy2001/B...ks/333_CPUz.jpg

http://www.geocities.com/get_n_jiggy2001/B...s/333_CPUz2.jpg

 

Mem Bandwidth: http://www.geocities.com/get_n_jiggy2001/B...m_Bandwidth.jpg

 

400 or 1:1:

 

Aquamark: http://www.geocities.com/get_n_jiggy2001/B...arks/400_AQ.jpg

 

CPUz: http://www.geocities.com/get_n_jiggy2001/B...ks/400_CPUz.jpg

http://www.geocities.com/get_n_jiggy2001/B...s/400_CPUz2.jpg

 

Mem Badwidth: http://www.geocities.com/get_n_jiggy2001/B...m_Bandwidth.jpg

 

3DMark06: http://www.geocities.com/get_n_jiggy2001/B...00_2.68GHz_.xls

 

Now from looking at these, it would appear as if the 8800GS is really using everything it can, and the Memory Bandwidth is playing a role in how fast this card can go. Like I said earlier this is preliminary, and I would as other users of an 8800 to check their results as well, and see what happens. I will be looking farther into this as well.

 

I had a bud of mine check on his 7800GT, and yes there was not really a significant different, but here we see that going from 266 - 400 there was a difference. I know my RAM isn't helping me any, and I want to retest once I get my OCZ 2x1GB...and see where I can go from there.

 

Thoughts....comments.....suggestions? I was rather interested in these findings, and want to see if this is something we need to look at as gamers. Reverting back to the days of MAX CPU with MAX FSB, instead of just trying to max out the CPU. Thanx all!

 

-Jubb out

Share this post


Link to post
Share on other sites

I didn't even think anybody read the stickies.

 

Overall system performance is comprised of every component in your system. Some things may have more of an effect than others, but if you want the most out of your system every aspect is important..

Share this post


Link to post
Share on other sites
Guest thespin

Wasn't AG's post directed more to most REAL WORLD applications than benchmarks ? There are definitely benchmarks that depend a lot on memory bandwidth e.g. Superpi.

 

But memory bandwidth may not be all that important for most real world applications with their mix of processor, memory, storage access, video ....

 

Take a look at the speed of memory included in high end desktops/workstations. It has been going up over time but it's usually not anywhere near the fastest memory available.

Share this post


Link to post
Share on other sites

Great work jubbin :)

 

Did you also run 3DMark on each divider? I would like to see how it responds to changing the mem bandwith...

 

As for AG's sticky, been a long time since I read it, but I remember one of the points he made was that you will get better performance by OCing your CPU with your RAM on a divider then if you ran your RAM 1:1 with a lower CPU OC.

 

He was also looking at tight vs. loose timings and demonstrated that timings don't have nearly as big of an effect now that the mem controller is on the cpu.

Share this post


Link to post
Share on other sites

With the processor held at a constant speed, an increase in memory bandwidth is always going to yield an increase in performance.

 

CPU megahertz offers the greatest increase in performance. If a person holds the speed of the cpu low in order to increase memory speed, performance will suffer.

 

What is explained in that post and still is true today, to achieve best performance, run the processor at the highest stable overclock and use whatever divider is necessary for the ram to be stable.

Share this post


Link to post
Share on other sites
Guest MadDias
As for AG's sticky, been a long time since I read it, but I remember one of the points he made was that you will get better performance by OCing your CPU with your RAM on a divider then if you ran your RAM 1:1 with a lower CPU OC.

 

that´s how i understood that sticky too...

why run ur cpu at 2,25ghz with ram 1:1 when u can get 3ghz with a divider.

staying at the same cpu speed and just lowering the ram speed will always decrease the performance.

Share this post


Link to post
Share on other sites

Ok so here are some new findings I have come accross. I decided to run my PC @ 2.608 with a 266 divider and running at 2.508 with a 1:1.

 

2.608 266 divider:

 

Aquamark: http://www.geocities.com/get_n_jiggy2001/B...rks/2.6_266.jpg

 

CPUz: http://www.geocities.com/get_n_jiggy2001/B....6_266_CPUz.jpg

http://www.geocities.com/get_n_jiggy2001/B...6_266_CPUz2.jpg

 

Mem Bandwidth: http://www.geocities.com/get_n_jiggy2001/B...m_Bandwidth.jpg

 

 

2.508 with 1:1 :

 

Aquamark: http://www.geocities.com/get_n_jiggy2001/B...rks/2.5_400.jpg

 

CPUz: http://www.geocities.com/get_n_jiggy2001/B....5_400_CPUz.jpg

http://www.geocities.com/get_n_jiggy2001/B...5_400_CPUz2.jpg

 

Mem Bandwidth: http://www.geocities.com/get_n_jiggy2001/B...m_Bandwidth.jpg

 

 

Now I assure you that in no way were these images doctored or "switched". Each one was taken directly after the tests completed. Each test was completed in the same order: Memory bandwidth, CPUz, Aquamark.

 

Looking at the results, the 2.508GHz nets close to a 5fps gain on the 2.6GHz. Now I know many of you are like "Yeah wow a whole 5fps...big deal" well it kinda is to me. Being the gamer at heart I want the highest fps count I can get outta my system, just like someone who does alot of encoding wants the fastest encode times they can get.

 

I will not contest that using CPU intensive applications: 3D Rendering, video encoding, ect. you that this would be a good idea for you. On the contrary it's not. Those types of applications look for raw CPU power, and you need to feed that need, so the bandwidth thing pretty much slides to the rear. This post is all about, squeezing every last fps outta the system. 5fps doesn't sound like much when you look at it, but it can make a difference between having a game run steady at 33fps or a noticable 28fps difference. I know the human eye cannot see above 32fps...and standard US broadcasts for video are at 29.97fps. But I still would like to keep my 5fps....not to mention a little less stress on the old processor :) . I could care less if I can encode a dvd 45secs faster, or render a full screen image in 3DS MAX in 6sec flat. I don't do those, I game.

 

I am not trying to turn this into a "My way is better to OC" post. This is really designed a a study into the new architecture of the 8800 series cards. I mentioned that in my first post that I knew the guide was old, and so this is more of a test with the newer hardware, just to see what if any differences may have gone un-noticed. I am in no way trying to disproved anything that Angry_Gamer was posting about, because as I found using a 7800GT, he is quite right ramping up the CPU gives you better gains with that card. With this card, maybe things are changing, and once multi-threading really hits these results will be totally off.

 

I also can't help but wonder if this phenominon might just be a fluke in my rig. I do hope others that have these cards and OC their machines will atleast try it and post their results to see if there is indeed an established pattern here. That perhaps there is a point with these cards where the bandwidth makes more of a difference than their predecessors.

 

I will be working with 3DMark06 later today, and see what those results net, and if they to mimic the Aquamark scores. As always i will post the results, and we can go from there. I also have plans for a Doom 3 bench as well as UT2004, and Fear. I figure if they are still good for the benchmarking of today, they are good enough for me. I dunno, this is all kinda fun for me, just to see the outcome :).

 

Also, the reasons I chose these clock speeds: many people can't hit a stable 3GHz on an OC with their hardware. There are alot of factors to get there. If you are one of the ppl who can, great, got an 8800 to benchmark on your highest 1:1 you can get vs a 3GHz with a 266 divider :)? I'd love to see the results!! For me, the best I have gotten: my board will post with a 3.003GHz, yet will not boot to XP....on any memory dividers/timings. That brings me to my other reason for the benchmark: to see if the new RAM I am getting may alleviate that problem....in which case the benchmarks will start all over. 2.4-2.6GHzish seems to be the norm for standard air-cooled systems on an OC. I have a bud that runs his Optron 165 @ 2.6, and that's all he can get...so really I figure its a safe generic range. I can't benchmark a theoretical 3GHz on my machine, but I can benchmark all the way up to 2.6 easily and quite stable, so that's why I am working within those ranges.

 

Thanks all!!

-Jubb out

Share this post


Link to post
Share on other sites
Guest MadDias
I decided to run my PC @ 2.608 with a 266 divider and running at 2.508 with a 1:1.

 

why a 266divider? u should try the 333divider to keep ur ram around the speed it is capable of...

50mhz difference on the ram is alot... no wonder ur 8800 is starving...

i bet that 2,6ghz on ur cpu and 217mhz(still 11mhz lower than in ur 1:1 run) on ur ram will get u better scores in aquamark (btw a very ancient benchmark).

i know that u only wanted to show that mem bandwidth really is a factor. i think noone will doubt that it is.

those mem dividers aren´t made for running ur ram at the lowest speed it can handle. i don´t think that anyone will run his ram 50mhz below of what it is capable of, just to gain 100mhz on the cpu.

 

Edit: u wouldn´t wanna see my highest 1:1 against my current 3ghz with a 300divider ;) and i´m on crossfire which means that my cards are also bandwidth hungry. i see no point in trying the 266divider since i want the most out of my system and not something that shows me that my ram is too slow, which i already know when i choose such a low divider.

Share this post


Link to post
Share on other sites

Indeed I agree with what you are saying. I am just looking at a worse case scenario, much like another guy I know who was running their A64 1.8 @ 2.6, but the only way for him to do it was with a 266 divider. I'll throw the 333 in there, and see what comes up. Thanks for the feed back and the ideas....I knew I could come here and get some great info :) .

 

-Jubb out

 

Ok Benchmark done, and yep you were right score of 100,714. Thanks for that. Well it was a fun ride while it lasted :). Thanks for the discussion guys!!

Share this post


Link to post
Share on other sites

Nice job on the write up there, bud. I'm totally with CPDMF on this one. I bought this hardware and I want every last shred of performance I can get out of it.... especially that last 5% ;)

Share this post


Link to post
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

Loading...
Sign in to follow this  

×
×
  • Create New...