Truth about x16 and x8 speeds [Archive]

youra6

01-20-10, 02:39 PM

3) Now, this is something that many people around here are not aware of, but PCI-Ex bandwidth is primarily important when your graphics card(s) do not have enough onboard memory as is required for the particular test at hand. If you run a test (say you run Crysis @1920x1200, 2xAA, with Very High texture quality) that uses 750MB of vram, and your cards have only 512MB of vram, this will cause your cards to 'farm out' the difference in memory usage to the system memory. To do this, large amounts of data must pass very quickly over the PCI-Ex bus. This is the scenario where the PCI-Ex bandwidth comes into play heavily. Otherwise, it really isn't terribly important.

This is written by a guy on another overclocking forum. Now, I have actually never heard of this before. I always thought faster cards took more advantage of faster bandwidth. It's just something interesting I found.

He also says: Many people think this question come down entirely to 'better card' = 'higher bandwidth requirement', but it's actually not that simple. In fact, it's actually kinda 'wrong', some of the time.

Any comments on this guy's explanation?

Thanks

EarthDog

01-20-10, 02:41 PM

http://www.techpowerup.com/reviews/AMD/HD_5870_PCI-Express_Scaling/25.html

A link to the information you provided would be nice... ;)

And isnt new news: http://forums.techpowerup.com/showpost.php?p=1701537&postcount=57

Aside from his smart ass answer of my same question... ;)

Bobnova

01-20-10, 02:42 PM

It's undeniable that a faster card requires more bandwidth. However, when you consider that a 5870 (fastest single GPU on the market at the moment) loses a mere 1-2% going from 16x to 8x, it's clearly not as important as some people think.

Like he said if the GPU doesn't have enough memory then PCIe width is tremendously important, though you don't have any business trying to do a >512mb setting on 512mb of ram in the first place IMO.

EarthDog

01-20-10, 02:45 PM

Like 99% of people, even here, know what at >512MB setting is though. Most just crank it up blindly.

youra6

01-20-10, 02:50 PM

A link to the information you provided would be nice... ;)

He never wrote down a source of where he got his info from. I do have the link of the the original post

Am i allowed to post links of other forums? I recently got yelled at by a overclock.net mod for linking to other sites... stupid if you ask me

Bobnova

01-20-10, 02:54 PM

Like 99% of people, even here, know what at >512MB setting is though. Most just crank it up blindly.

Good point.

I figure if it doesn't run for beans, it was too much :P

EarthDog

01-20-10, 02:56 PM

ROFLMAO! Yup! Good call! :)
__________

No big on the links man.. thought it wasnt a forum user but someone/thing more reliable. But as you can see, that is at least supported by some other forum user. :)

youra6

01-20-10, 03:33 PM

so when it comes down to it, do you guys think that "its wrong some of the time?"

Agree or disagree?

Many people think this question come down entirely to 'better card' = 'higher bandwidth requirement', but it's actually not that simple. In fact, it's actually kinda 'wrong', some of the time.

Bobnova

01-20-10, 04:44 PM

I do indeed disagree.
A faster card will always require more bandwidth then a slower card, assuming you are comparing apples to apples.
If you're comparing a low end card that requires a lot of CPU time (say it can't draw triangles) to a high end card that requires only a little bit (it can draw triangles) the bandwidth requirement for the high end card may be lower due to the triangles not haven't to be drawn elsewhere and sent intact.

That said, there isn't much that falls into the can't-do-it-but-the-cpu-can category these days.

Evilsizer

01-20-10, 05:06 PM

so when it comes down to it, do you guys think that "its wrong some of the time?"

Agree or disagree?

i would have to agree... in really does depend just because the card is faster doesnt always mean needs more bandwidth. to keep this post short since i rewrote it like 8 times now. the one main big area i want to point out where it doesnt matter. is that say you playing at 1920x1080x32Bit using 8xAA,8xAF. now this is where the gpu does all the work and the cpu doesnt have much to do. the role would be different if we were talking something CPU bound but for this instance we are not. now then with such a high res you do need more bandwidth as the texures are higher res,etc. now add into the fact you have the gpu doing AA and AF to the image. if you give the gpu a x16 pipeline to work with at those settings it wont need that much bandwidth. because the gpu has to apply the AA and AF effects to the image, that takes time. A faster card would be able to apply AA/AF to the image faster but wouldnt be a big performance increase using x16 over x8.

when we talk about CF or SLI it is a different matter.

AngelicPenguin

01-20-10, 05:22 PM

This article covers the PCI-Express scaling for the 5870 pretty well with lots of benchmarks: http://www.techpowerup.com/reviews/AMD/HD_5870_PCI-Express_Scaling/

This article talks about it for the 5970 on this page: http://www.hardwarecanucks.com/forum/hardware-canucks-reviews/25628-sapphire-radeon-hd-5970-2gb-oc-edition-review-22.html

Evilsizer

01-20-10, 05:25 PM

This article covers the PCI-Express scaling for the 5870 pretty well with lots of benchmarks: http://www.techpowerup.com/reviews/AMD/HD_5870_PCI-Express_Scaling/

is a good link for sure but doesnt show AA/AF at different settings vs non AA/AF per resolution. leave many blank spots for this discussion, would be nice if they did go into more detail.

EarthDog

01-20-10, 06:04 PM

Yep...... : http://www.overclockers.com/forums/showpost.php?p=6372913&postcount=2

I even asked for this but was shot down. Not by the article writer however...

AS far as the 5970 goes, that should put a hole in that guys thinking as far as a better card not need higher bandwidth requirements. Obviously that is to the extreme with a dual gpu solution, but...

Bobnova

01-20-10, 07:01 PM

i would have to agree... in really does depend just because the card is faster doesnt always mean needs more bandwidth. to keep this post short since i rewrote it like 8 times now. the one main big area i want to point out where it doesnt matter. is that say you playing at 1920x1080x32Bit using 8xAA,8xAF. now this is where the gpu does all the work and the cpu doesnt have much to do. the role would be different if we were talking something CPU bound but for this instance we are not. now then with such a high res you do need more bandwidth as the texures are higher res,etc. now add into the fact you have the gpu doing AA and AF to the image. if you give the gpu a x16 pipeline to work with at those settings it wont need that much bandwidth. because the gpu has to apply the AA and AF effects to the image, that takes time. A faster card would be able to apply AA/AF to the image faster but wouldnt be a big performance increase using x16 over x8.

when we talk about CF or SLI it is a different matter.

If you compare the same settings (5839xAA at whatever res) on different cards, the faster one will still need more bandwidth then the slower one.

720x770

01-20-10, 08:41 PM

I don't know why this is a discussion, he clearly states it all right there in common sense.

slower cards make the PCIe bus interact with the system RAM, which means more bandwidth going across the bus. a PCIe x16 bus can transfer 8GB/s. most DDR2-DDR3 RAM transfer around that or more, up to 12GB/s and beyond. So it's safe to say that RAM can fully saturate the PCIe bus. Although, with management protocols you could have the card use it's full bandwidth and leave the left over to the RAM transfers, effectively limiting the bandwidth that system RAM is allowed to use.

But then you have another problem, if the RAM isn't transferring data fast enough, you'll still see horrible performance. It's a lose lose scenario either way. It's like trying to fit a gallon of water into 1 cup slowly by waiting for it to evaporate before pouring more in.

Vram is very fast, transferring at 50-100's of GB/s give or take, so it's MUCH faster than even the fastest system RAM, which means even if the PCIe bus had loads more bandwidth the system RAM would still be a bottleneck compared to the Vram.

EarthDog

01-20-10, 08:55 PM

Its not a slower card. Its any card that goes beyond its ram usage and has to write to system ram.

youra6

01-21-10, 02:42 AM

I don't know why this is a discussion, he clearly states it all right there in common sense.

slower cards make the PCIe bus interact with the system RAM, which means
Vram is very fast, transferring at 50-100's of GB/s give or take, so it's MUCH faster than even the fastest system RAM, which means even if the PCIe bus had loads more bandwidth the system RAM would still be a bottleneck compared to the Vram.

And I dont understand why a person cannot get a second opinion? I wont automatically take someone's word for it. Always better to sure than to be sure with false knowledge. What may be obvious to one may not be obvious to another. I appreciate the help, but please keep that tone to yourself if you will.

I think you are getting confused with PCI-E bandwidth and GPU VRAM bandwidth.

They dont operate the same way. the GPU and onboard RAM bandwidth is completely independent from the GPU to system bandwidth. I am fairly certain the "50-100" gb/s refers to the card's internal memory bandwidth, not the external PCI Express bandwidth IIRC

Evilsizer

01-21-10, 02:54 AM

If you compare the same settings (5839xAA at whatever res) on different cards, the faster one will still need more bandwidth then the slower one.

in a cpu bound situation i can see that... when you place all the load on the GPU with aa/af the PCIE width wont matter much for the faster card. the card is reprocessing the same data over and over, there is nothing new to send to it.

youra6

01-21-10, 03:18 AM

in a cpu bound situation i can see that... when you place all the load on the GPU with aa/af the PCIE width wont matter much for the faster card. the card is reprocessing the same data over and over, there is nothing new to send to it.

cpu bound as in lets say for example.... 3dmark06?

Evilsizer

01-21-10, 04:10 AM

cpu bound as in lets say for example.... 3dmark06?

no cpu bound is when the resolution is low, ie 1280x1024 and lower. gpu bound is around 1600x1200 and higher. Does also include things that improve IQ like AA and AF.

720x770

01-21-10, 04:22 AM

And I dont understand why a person cannot get a second opinion? I wont automatically take someone's word for it. Always better to sure than to be sure with false knowledge. What may be obvious to one may not be obvious to another. I appreciate the help, but please keep that tone to yourself if you will.

I think you are getting confused with PCI-E bandwidth and GPU VRAM bandwidth.

They dont operate the same way. the GPU and onboard RAM bandwidth is completely independent from the GPU to system bandwidth. I am fairly certain the "50-100" gb/s refers to the card's internal memory bandwidth, not the external PCI Express bandwidth IIRC

first of all, you quoted me completely wrong. you spliced two different sections of my post together.

second, I clearly state that the PCI-E x16 bandwidth is 8GB/s. Then I clearly stated that System RAM can transfer upwards of 8GB/s and beyond. Then I finally stated Video RAM (on the card) is Even Faster than System RAM, going into the 100's of GB/s give or take (Internally). The point being, if your card went over it's memory limit and used RAM, it would take up bandwidth on the PCI bus and that even if the bus DID have enough bandwidth, System RAM would still bottleneck the card.

And obviously, the larger the resolution and quality settings and the more frames being drawn all adds up to more used bandwidth. They wouldn't make cards for PCI x16 if they required more bandwidth than was available, however PCI 2.0 is on the verge of its usable limits, and that's why there's PCI 3.0/PCI x32 and so on.

The only time people run into bandwidth situations right now is on CF/SLI setups and the performance drop is negligible compared to the gains. Granted, PCI 2.0 does lose about 15% of it's raw bandwidth to encoding overhead. So technically it's the overhead limiting the cards. PCI 3.0 lowers the overhead to 1.5% while increasing overall bus bandwidth.

youra6

01-21-10, 05:35 AM

second, I clearly state that the PCI-E x16 bandwidth is 8GB/s. Then I clearly stated that System RAM can transfer upwards of 8GB/s and beyond. Then I finally stated Video RAM (on the card) is Even Faster than System RAM, going into the 100's of GB/s give or take (Internally). The point being, if your card went over it's memory limit and used RAM, it would take up bandwidth on the PCI bus and that even if the bus DID have enough bandwidth, System RAM would still bottleneck the card.

I apologize for the splicing. I only wanted to quote the second part.

As if you are saying that PCI-E doesnt have enough bandwidth for the GPU or RAM, thus the RAM will bottleneck the GPU?

Am i getting this straight? Please correct me if I am not.

You are saying as if the PCI-E bandwidth (8 gb/s) is related to the GPU memory bandwidth (100 gb/s)

I almost positive that those two bandwidths are not related to each other at all.

720x770

01-21-10, 04:02 PM

I apologize for the splicing. I only wanted to quote the second part.

As if you are saying that PCI-E doesnt have enough bandwidth for the GPU or RAM, thus the RAM will bottleneck the GPU?

Am i getting this straight? Please correct me if I am not.

You are saying as if the PCI-E bandwidth (8 gb/s) is related to the GPU memory bandwidth (100 gb/s)

I almost positive that those two bandwidths are not related to each other at all.

Vram is not related, that bandwidth is internal. However, the Vram does get data through the PCI bus in a one-way fashion. The GPU communicates two way via the bus, receiving and sending data to/from the rest of the system.

Conumdrum

01-21-10, 09:59 PM

He never wrote down a source of where he got his info from. I do have the link of the the original post

Am i allowed to post links of other forums? I recently got yelled at by a overclock.net mod for linking to other sites... stupid if you ask me

They did the same to me once. I was posting links to other stores and contradicting the mods use of some watercooling liquids. And posting links from other sites. They make some direct money from sponsorships to the watercooling forum and don't want to lose that, or be told they are wrong. I don't go there anymore. That was over a year ago.

Here it's different, a solid useful link to another post on another forum is good. I link to Toms hardware and XS and other ALLL the time. I don't expect a mod to slam you on that. Links to the latest BBQ grill your selling, thats another story......

Bobnova

01-21-10, 10:05 PM

They make some direct money from sponsorships to the watercooling forum

Big ole fail there. Nothing guarantees falsehood like getting paid by someone supplying the parts you're talking about.

youra6

01-22-10, 11:23 AM

Vram is not related, that bandwidth is internal. However, the Vram does get data through the PCI bus in a one-way fashion. The GPU communicates two way via the bus, receiving and sending data to/from the rest of the system.

PCI-E has a switch, thought the bus was two way?

youra6

01-22-10, 11:27 AM

Bobnova

01-22-10, 11:30 AM

PCIe is two way, though the bandwidth used by the GPU sending data back to the CPU is fairly minimal.

4GHZ_or_bust

01-22-10, 12:08 PM

So it's between 2 or more video card that uses PCIe bandwidth the most?

ChaosInMind

02-04-10, 09:54 AM

for 1-2% difference, why not just OC your PCI-E by oh say, 5Mhz?

killem2

02-04-10, 09:59 AM

for 1-2% difference, why not just OC your PCI-E by oh say, 5Mhz?

If that were to be done, would it be 105 mhz total? Sorry I'm not the greatest at ocing :P

freeagent

02-04-10, 10:03 AM

http://www.mydigitallife.info/2009/08/24/pci-express-3-0-specifications-release-pushed-out-to-2010/

:beer:

Evilsizer

02-04-10, 10:34 AM

for 1-2% difference, why not just OC your PCI-E by oh say, 5Mhz?

If that were to be done, would it be 105 mhz total? Sorry I'm not the greatest at ocing :P

i use to oc the pcie buss with my NV 8800GTS cards. really did make a difference in CSS for the same CPU speed. i forget the difference but now that i have ATI, it didnt like the oc'd pcie buss. i never tried 105mhz pcie buss on ati, as i was use to using 110mhz pcie buss speed. you dont have to worry about HD corruption till 115mhz pcie buss speed.

ChaosInMind

02-04-10, 10:57 AM

i use to oc the pcie buss with my NV 8800GTS cards. really did make a difference in CSS for the same CPU speed. i forget the difference but now that i have ATI, it didnt like the oc'd pcie buss. i never tried 105mhz pcie buss on ati, as i was use to using 110mhz pcie buss speed. you dont have to worry about HD corruption till 115mhz pcie buss speed.

Yeah, 105 total.. I know 115 is generally "safe" but don't like to push any further than 10. 5-9Mhz should give just enough bandwidth to make up the 1-2% difference. A total of 105Mhz-110Mhz. The theory is this provides the PCI-E 8x-8x slots more bandwidth.

Evilsizer

02-04-10, 11:04 AM

The theory is this provides the PCI-E 8x-8x slots more bandwidth. it is not theory, it is factual. pcie has a lane width like memory does and a speed. for the same width with a increase in speed you get higher bandwidth,lower latency. that is assuming the latency setting doesn't change when you increase the clock speed at which it operates.

youra6

04-17-10, 03:55 AM

when we talk about CF or SLI it is a different matter.

sorry to revive this thread, but could you elaborate a little more on this?

Does x8 bandwidth become less capable to scale... lets say two GTX 480 or 5970?

How does this change when we talk about CF or SLI?

Evilsizer

04-17-10, 12:01 PM

sorry to revive this thread, but could you elaborate a little more on this?

Does x8 bandwidth become less capable to scale... lets say two GTX 480 or 5970?

How does this change when we talk about CF or SLI?

with dual gpu based cards yes at x8 it would be more limiting for the card. if talking about single gpu threads, check out this thread (http://www.overclockers.com/forums/showthread.php?t=619246).

youra6

04-17-10, 02:23 PM

Evilsizer

04-17-10, 03:03 PM

thanks for the reply.

How would x8 be more limiting? on CF and SLI?

This is just to expand my knowledge on the topic.

Most people say that "oh performance will only go down a few notches with SLI x8" but most dont know why

with a card using two gpus on a 8x bandwidth it is more limiting. less bandwidth to each gpu since they both are using one slot. the video card with dual gpus has a PCIE controller onboard to talke to both GPUs. you wind up with slot x8->gpu pcie controller x16-> GPU1/GPU2.

Not only that but current high end GPUs still hardly use over PCIE 1.0 x8 bandwidth. since they have PCIE 2.0 which is double the bandwidth, that means a PCIE 2.0 x8 slot is really a PCIE 1.0 x16 slot. as far as bandwidth goes, there are other things like how much ram the gpu has.

it really has to do with how much bandwidth GPU's really need vs what they are getting. do you really need 500GB/s bandwidth if the gpu only needs 200GB/s? those are not the real numbers for bandwidth im just using them for a example.

it is also the same thing with cpus as well, why do you need DDR3-2000 ram. when the cpu might only use 50-75% of the bandwidth the ram can provide. lots of trade offs in the bigger picture...