XFX 6950 2GB Memory Bandwidth Issue?

[Mistwalker]

Running 6950 CF with shader-unlocked BIOS, getting 5GB/sec one way and 500 MB/sec the other. CPU-Z does confirm 16x/16x PCI lane speeds...

 

[biostud]

my 6850 got 5.7Gb/700mb

 

[VirtualLarry]

This is seriously F'ed up that people are only getting ~500MB/sec from GPU to system RAM. PCI-E's big advantage (so it was marketed) was that it's speed was bi-directional, unlike AGP. But here we have clear evidence that it is no better, in many cases.

 

[biostud]

but you have to wonder whether is a driver issue or hardware problem.

 

[Ben1979]

but you have to wonder whether is a driver issue or hardware problem.

ppl with 6950 with the same 11.1a driver got 500 mb/sec & 5,000 gb/s

It's actually all mixed up result.

Also on mine i used the same driver for the 5870 & 2x 6950 ..

 

[Pohemi]

From you're quote you sure didn't read my reference thread.

Actually, I did read your *first* thread, and saw no reason why you felt it necessary to make a new thread about problems with the same video card and system.

 

[Ben1979]

Actually, I did read your *first* thread, and saw no reason why you felt it necessary to make a new thread about problems with the same video card and system.

Cause a new thread was more appropriate to raise a issue with the PCI-E Speed test.

So we could get result from ppl all over. Also those low result related to the PCI-E doing 500~800 mb/sec could be a bottleneck for crossfire. Depending of the games tested and such. So until someone show up this ain't affecting a crossfire setup. I wouldn't consider it a issue only related to my system\cards\spec.

As mentioned in my case one card (xfx 6950) even if it show 500~800 mb/sec act normally in bench & games but 2 of them in crossfire act weird compared to all the bench i saw online.

 

[MTDEW]

Pc specs....
Win7 64bit
i7 920@4.2ghz
12gb ddr3
Gigabyte GA-X58A-UD3R
Intel SSD
2gb Sapphire HD6950 x 2

So, i'm running two sapphire 6950's in X-fire with unlocked shaders and 11.1a drivers and i get 5GB/sec one way and 700 MB/sec or so the other way also.

Yet, my laptop with a mobility hd5870 gets 3GB/sec both ways.

So yeah, this is interesting as something seems odd.

 

[HumblePie]

I have for my system...

2600K @ 4.8 Ghz
16GB DDR3 ripjaws
Asrock P67 Fatality
OCZ 120GB Vertex
2x XFX 6950's flashed to 6970's and @ 920/1400 each

I get 2.8 GB bandwitdth from CPU and back.


I'm assuming since I'm at x8/x8 lanes for my crossfired solution that this is about right.

 

[biostud]

seems to be 6xxx on p55/x58 boards

 

[Bill Brasky]

Here is an interesting thread on this topic. In short, this is a unique problem with a specific combination of intel mainboards (x58/P55) and ATI cards. Games shouldn't be affected, but GPGPU applications will.

http://forums.amd.com/devforum/messageview.cfm?catid=328&threadid=130923&STARTPAGE=1&FTVAR_FORUMVIEWTMP=Linear


I also have the low GPU->CPU score on my P55/ HD5850 rig. Maybe I'm off-base here, but for games it seems to be more important that the gpu can receive data from the cpu and memory. However, sending data to the cpu and ram FROM the gpu wouldn't create any sort of bottleneck. This is a non-issue in my opinion.

 

[Ben1979]

I just bought 2x 6950 Sapphire. I'll report tonight with the same motherboard drivers.

 

[Arkadrel]

a unique problem with a specific combination of intel mainboards (x58/P55)

ohhh.. there we go ^-^ lmao.... motherboard issues, thats gotta be a first for Intel motherboards right?

 

[NoQuarter]

When I run this tool on my Radeon 5770 it kicks the GPU clock up from idle to 3d speeds. My guess is the 6xxx series has a different power management config that keeps the GPU at idle when doing this test since no rendering is happening, and thus can't max out the GPU->CPU transfer.

Try manually clocking your card up or running a 3d app to force it into 3d clocks then run the test.

 

[artvscommerce]

I didn't realize Bill Brasky was on anandtech.. I know bill brasky! He's a ten-foot tall beast man, who showers in vodka, and feeds his baby shrimp scampi

 

[thilanliyan]

When I run this tool on my Radeon 5770 it kicks the GPU clock up from idle to 3d speeds. My guess is the 6xxx series has a different power management config that keeps the GPU at idle when doing this test since no rendering is happening, and thus can't max out the GPU->CPU transfer.

Try manually clocking your card up or running a 3d app to force it into 3d clocks then run the test.

Nope, my 6950 clocked up automatically as well.

 

[Skurge]

OP are you sure you have everything setup? Crossfire bridges, drivers, etc

 

[Jules]

Im having the same issue with CF 6950's in my rig. Sig below

 

[techie81]

P55 and 6950 here with 11.1a

 

[MTDEW]

Nope, my 6950 clocked up automatically as well.

I used GPUZ to check when i ran it on my x-fired 6950s.
Main card clocked up , but the second card did not.

It doesnt seem to hinder gaming performance at all though as mentioned earlier.

 

[thilanliyan]

^Ah maybe that is the problem...maybe the 2nd card is not clocking and is causing those weird results for the OP I mean?

 

[Ben1979]

I used GPUZ to check when i ran it on my x-fired 6950s.
Main card clocked up , but the second card did not.

It doesnt seem to hinder gaming performance at all though as mentioned earlier.

The test run it on the first device (first card..) then second device..

It say's DEVICE: 1 when you start it.

 

[MTDEW]

The test run it on the first device (first card..) then second device..

It say's DEVICE: 1 when you start it.

Ok, i'll do it later tonight when i get a chance, but i doubt it will matter because even a single card should hit more than 700-800mbs.
I'm on my laptop now so i can't.

Anyway, after reading Bill Braskys link above i'm not as concerned about it as i was when i first ran this test, and to be honest i would have never noticed if it wasn't for this thread since my gaming/benchmarks seem right on par with what they should be.

BTW: GPUZ does report PCI-E 2.0 x 16 @ x16 2.0 for my cards in case anyone asks me to check that.

Ben1979,
Just curious , I didn't read all of your other threads, are you getting poor gaming and/or benchmark performance that led you to try this test?
Or is this the only test that seems "off/odd" with your setup?

 

[Ben1979]

Ok, i'll do it later tonight when i get a chance, but i doubt it will matter because even a single card should hit more than 700-800mbs.
I'm on my laptop now so i can't.

Anyway, after reading Bill Braskys link above i'm not as concerned about it as i was when i first ran this test, and to be honest i would have never noticed if it wasn't for this thread since my gaming/benchmarks seem right on par with what they should be.

BTW: GPUZ does report PCI-E 2.0 x 16 @ x16 2.0 for my cards in case anyone asks me to check that.

Ben1979,
Just curious , I didn't read all of your other threads, are you getting poor gaming and/or benchmark performance that led you to try this test?
Or is this the only test that seems "off/odd" with your setup?

To begin with on my last board i had issue with the PSU. I had a 18 AWG cable on instead of a 15 AWG, then i had normal FPS in some games.

Then i was worried about the 4x/8x lanes on my old board so i try with another PSU\MB(8x\8x). I had about the same results.

I bought 2x Sapphire just to make sure the XFX wasn't defective and they act the same with this PCI TEST about 500~ mb\sec.

Only game that I can't get over 50 FPS is Crysis Warhead. (1 card do 50fps alone) at 1920x1080 Gamer settings 4x msaa.

3DMarks11: P7900
Unigene Heaven Benchmark v2.1: 1920x1080 high 2xAA: 2101 - 83.4 FPS
3DMarks Vantage 1.1: P23857
3DMarks06: 1920x1080 Optimal texture 0msaa 0af Score: 19087


1920x1080:

Lost Planet 2 bench level A, high setting D11: middle 0aa\0af
91.9 FPS
Lost Planet 2 bench level A, high setting D11: high 2aa\0af
72.4 FPS

Dirt2: Everything Ultra 8xAA: 103.6 fps (1680x1050 108 FPS)

Call Of Pripyat Bench Preset Ultra D11 4xAA : 99~ avg

Mafia 2 Bench: 16AF Maxed Phsyx off: 76 fps (1680x1050 75.9 fps)

AVP D11 Bench: Very High,16xAF,Tessalation On,1xAA 106.3 FPS


So yah most of the bench seem normal.. except for low 3Dmark, might be because I use i7 860 stock.

For the PCI-E Speed issue i guess it's related to P55\X58 because no AMD users report low speed. P67 seem fine. (All the reviews site tested with X58 mostly so there no way to tell if there a impact or not)

Still doesn't make any sense to me that a 5870 get over 5gb/sec. Also don't understand how it can't impact the performance somehow.

So basically someone who bought a X58 with 16x/16x pci-e lanes wasted his money.. if the GPU > CPU cap at 500~800mb/sec on both platform.

Just weird nobody\review sites didn't report that before.

 

[papatsonis]

i can confirm that 4 more ppl with X58-6950 in a Greek forum, experience the same low scores.. (1 AMD score reported+6970, with normal scores).Seeing that only P55-X58+69x0 combinations are affected, i start to suspect that there might be some bug(?) in pci-e interface in 69x0, or intel took some shortcut..? when designed the pci-e controller in X58 and in lynnfield (Sandybridge pci-e controller, seems to have normal scores with 69x0, so far)

edit: also my setup (Q9650-X48 with 2x6950) has normal scores ~5k cpu->gpu, ~6k gpu->pu