5850 just as fast as a 5870?

[Daedalus685]

"Based on my arbitrary definition of architecture and total lack of understanding how drivers and hardware work, I declare that I am right and no amount of logic or so called experts in any given field can change my mind."

Let me ask then, explain how crossfire works if it increases the bandwidth by a factor of two?

Certainly, you have access to two copies of the memory, thus on reads you have two paths.. so in theory you could 'double' that aspect in a perfect situation, but I was under the impression that almost all of the access to VRAM was to write the frame. Since the write must be done twice, as the memory is akin to raid1 (in that it is mirrored at least) you have the same bandwidth as you have with only a single card. Even if the read/write is 50/50 you are still only going to increase the bandwidth by a theoretical perfect max of 50%.

Also, how exactly do minimum frame rates tell the entire story? Sure, I would rather play a game that is exactly 60fps as compared to half 30 half 60, but are you implying that you think the 5770 has much higher max and thus the average is the same but playability is not? The nadir of a curve is important, but for minimum to not have any bearing on the average they woudl have to be ridiculously narrow peaks, something far more conducive of a driver fault than anything else. For average to be the same the max would have to be much higher than other cards, except in the sections where it is bottle necked by the memory, this does not seem to be the case. It is not very scientific to ignore all of the data with minimum frames only to lambast those that don't have it for being totally wrong for not including it.

Any graphs I have seen show very much the same fps curves. If minimum was more than just an instantaneous spike I'd agree, but it is not. There are no long flat parts to indicate memory hogging sections of the game that perform well on another card, it is lower almost across the board.

What makes you think the isntantaneous dips are not a driver issue with the card? Other than your assumption that it is the same architecture just because you said so... Do you not think that BFG, having actually done tests and written a report on this topic, is not in a better position to draw conclusions? It is one thing to disagree with a report, quite another to disagree without being able to provide one bit of evidence to support your seemingly illogical conclusions.

EDIT: I'm getting this and the 5770 thread mixed up as the same one.. assume this a reply to both..

 

[AzN]

"Based on my arbitrary definition of architecture and total lack of understanding how drivers and hardware work, I declare that I am right and no amount of logic or so called experts in any given field can change my mind."

Who are you referring this to? You should read forum rules before posting. That is considered personal. if you want to talk about tech that's fine but don't you make this into a personal piss match.

Let me ask then, explain how crossfire works if it increases the bandwidth by a factor of two?

Certainly, you have access to two copies of the memory, thus on reads you have two paths.. so in theory you could 'double' that aspect in a perfect situation, but I was under the impression that almost all of the access to VRAM was to write the frame. Since the write must be done twice, as the memory is akin to raid1 (in that it is mirrored at least) you have the same bandwidth as you have with only a single card. Even if the read/write is 50/50 you are still only going to increase the bandwidth by a theoretical perfect max of 50%.

Certainly in perfect situation the bandwidth doubles which is why we refer this as theoretical rates. if you want to believe that crossfireX only pertain 50% of the bandwidth that's you but in crossfireX performance shows this isn't the case as it's showing mostly 80-100% scaling.

Also, how exactly do minimum frame rates tell the entire story? Sure, I would rather play a game that is exactly 60fps as compared to half 30 half 60, but are you implying that you think the 5770 has much higher max and thus the average is the same but playability is not? The nadir of a curve is important, but for minimum to not have any bearing on the average they woudl have to be ridiculously narrow peaks, something far more conducive of a driver fault than anything else. For average to be the same the max would have to be much higher than other cards, except in the sections where it is bottle necked by the memory, this does not seem to be the case. It is not very scientific to ignore all of the data with minimum frames only to lambast those that don't have it for being totally wrong for not including it.

I've never said 5770 only peak higher maximum frames than to certain cards. What I said is that Core peak maximum frame rates and bandwidth plays the equalizer to peak theoretical fillrate.

Any graphs I have seen show very much the same fps curves. If minimum was more than just an instantaneous spike I'd agree, but it is not. There are no long flat parts to indicate memory hogging sections of the game that perform well on another card, it is lower almost across the board.

Why do you need multiple accounts to argue Schmide?

Anyways... you get different scenes in games that fluctuate it's the same reason why your frames fluctuate and not "instantaneous" by bandwidth.. As there are too many variables you have to consider in game.

What makes you think the isntantaneous dips are not a driver issue with the card? Other than your assumption that it is the same architecture just because you said so... Do you not think that BFG, having actually done tests and written a report on this topic, is not in a better position to draw conclusions? It is one thing to disagree with a report, quite another to disagree without being able to provide one bit of evidence to support your seemingly illogical conclusions.

http://techreport.com/articles.x/17986/2

These days, I feel like we should attach as many caveats to that table as one might to a federal budget resolution. These are simply theoretical peak values, and they are in some cases quite academic. You're rarely going to hit the peak pixel fill rate on one of these cards, for instance, since memory bandwidth will likely limit you first.

That's the thing BFG's finding was only showing avg. frame rates to come up with his conclusions. From my extensive testings on all the cards I've owned memory makes dramatic difference in minimum fps and not so much on avg. fps.

 

[ugaboga232]

So just trying to recap what you are saying, BFG and Daedalus's findings are not as good as yours even when you have not tested the 5870 and they have.

Also Daedalus brings up a good point with min fps and avg fps. If the mins go up but the average stays the same or increases slightly, assuming the maxes do not go down, that means the gpu spends very little time at its minimum fps. While higher bandwidth might help this, it means very little and more efficient drivers can help make sure the card doesn't have quick spikes and drops.

 

[Meaker10]

" - 1) In computer networks, bandwidth is often used as a synonym for data transfer rate - the amount of data that can be carried from one point to another in a given time period (usually a second). This kind of bandwidth is usually expressed in bits (of data) per second (bps). Occasionally, it's expressed as bytes per second (Bps). A modem that works at 57,600 bps hastwice the bandwidth of a modem that works at 28,800 bps. In general, a link with a high bandwidth is one that may be able to carry enough information to sustain the succession of images in a video presentation. "

http://searchenterprisewan.techtarget.com/sDefinition/0,,sid200_gci211634,00.html

We have already explained that all write requests have to be doubled in order for both cards to have an identical set of data.

Just like a raid 1 array of hard drives you get NO write performance increase.

Writes and reads make up bandwidth and therefore it has not doubled.

How many ways do I have to explain this to you?

 

[AzN]

Jesus how many accounts do you have?

 

[AzN]

" - 1) In computer networks, bandwidth is often used as a synonym for data transfer rate - the amount of data that can be carried from one point to another in a given time period (usually a second). This kind of bandwidth is usually expressed in bits (of data) per second (bps). Occasionally, it's expressed as bytes per second (Bps). A modem that works at 57,600 bps hastwice the bandwidth of a modem that works at 28,800 bps. In general, a link with a high bandwidth is one that may be able to carry enough information to sustain the succession of images in a video presentation. "

http://searchenterprisewan.techtarget.com/sDefinition/0,,sid200_gci211634,00.html

We have already explained that all write requests have to be doubled in order for both cards to have an identical set of data.

Just like a raid 1 array of hard drives you get NO write performance increase.

Writes and reads make up bandwidth and therefore it has not doubled.

How many ways do I have to explain this to you?

how is raid read and write have anything to do with crossfirex?

http://techreport.com/articles.x/8826/3

SuperTiling — This method is the default for Direct3D applications, and it's also the only mode unique to CrossFire. The screen is subdivided into a checkerboard-like pattern of 32x32-pixel squares, with one card rendering what would be the red squares on the checkerboard, and the other rendering what would be the black squares on the board. ATI says this method distributes the load between the cards neatly and efficiently, but SuperTiling offers benefits only in terms of pixel-pushing power, not geometry computation. Both cards must compute the underlying geometry for each frame individually. SuperTiling is not supported in OpenGL.

Alternate-frame rendering — The king of all multi-GPU load balancing modes is alternate-frame rendering, or AFR for short. AFR interleaves full frames rendered by the two cards, so that, say, the master card renders odd frames and the slave card renders even ones. This is the preferred load-balancing mode whenever possible, because AFR shows markedly better performance scaling than other modes. Part of the reason for AFR's good performance is the fact that it splits the geometry processing load between the two cards evenly, something no other mode does.

2 cards doing independent things.

 

[ugaboga232]

But if VRAM is copied, that means that the same writes will have to be done for both cards. The reads can be different but since you have less than double memory bandwidth and doubled shaders, you would expect it to be just as memory bottlenecked which is not the case.

 

[Daedalus685]

Who are you referring this to? You should read forum rules before posting. That is considered personal. if you want to talk about tech that's fine but don't you make this into a personal piss match.

It was a random quote, not intended to be personal or derogatory.. more to highlight the frustration one might encounter when reasonable argument gets one nowhere. Though for someone talking about the forum rules you talk to me awful personally..

Certainly in perfect situation the bandwidth doubles which is why we refer this as theoretical rates. if you want to believe that crossfireX only pertain 50% of the bandwidth that's you but in crossfireX performance shows this isn't the case as it's showing mostly 80-100% scaling.

You don't understand how crossfire works. The memory is not doubled, each card can only access its own (thus the frame buffer is not twice as big). I am aware you realize that the memory amount is not doubled, but how then do you figure the bandwidth is doubled? You can ignore my point all you want, but please do explain it to me if I am wrong. If I have two frame buffers, and i have to write the same thing to both then I am writing twice as much as I was to one of them. Sure If I am talking about how much data is written per second, the bandwidth is obviously doubled. But, we are taking something, doubling it, then writing it. The useful bandwidth is the same as it was without crossfire. I'm not sure how much of a cycle is spend reading the frame buffer (not much I'd recon) but even the reads would only have double bandwidth if the data were already there. It is nowhere near double the bandwidth effective, though I do not know exactly how different it is. I'd appreciate if anyone who know for sure could chime in.

We see 80-100% scaling for the reason that the cards are NOT bandwidth starved. It is not proof that the cards are doubling up the bandwidth..

I've never said 5770 only peak higher maximum frames than to certain cards. What I said is that Core peak maximum frame rates and bandwidth plays the equalizer to peak theoretical fillrate.

I was stating that the frame plots look very similar from card to card. If bandwidth was a problem you should see a section where the fps is muted. Also, lower minimums pull the average down. I know average does not tell the whole story but for the min to be much lower than another card but the average stay the same the max must either be higher, or the minimum was over a very short time frame.

Why do you need multiple accounts to argue Schmide?

I don't know who that is, If you wish proof that I am my own person I have been active on toms hardware and the ati forums for some time with this same name.

Anyways... you get different scenes in games that fluctuate it's the same reason why your frames fluctuate and not "instantaneous" by bandwidth.. As there are too many variables you have to consider in game.

A benchmark run is a time demo recording. They are by definition exactly the same from run to run. That is the whole point, as they would be useless otherwise.

That's the thing BFG's finding was only showing avg. frame rates to come up with his conclusions. From my extensive testings on all the cards I've owned memory makes dramatic difference in minimum fps and not so much on avg. fps.

Yes, he only showed average. But who cares, if you are so set against it then run your own tests with the minimum included. There are many sites that list minimum now. A few sites will even show the full fps log (like Hardocp).

Your experience has little to do with this discussion unless you can produce numbers like bfg has. I assume you mean memory bus, as the total amount means nothing to this conversation. Minimum FPS will be higher if bandwidth is not an issue, as has been explained when you run into a scene that is bandwidth starved the performance becomes muted. This is obvious. The point is that in this situation that behavior is not seen. The minimum is very erratic.

 

[ugaboga232]

Also you have yet to give data proving your point while we have a lot of data showing that it is not memory bottlenecked.

 

[Daedalus685]

how is raid read and write have anything to do with crossfirex?

http://techreport.com/articles.x/8826/3

SuperTiling — This method is the default for Direct3D applications, and it's also the only mode unique to CrossFire. The screen is subdivided into a checkerboard-like pattern of 32x32-pixel squares, with one card rendering what would be the red squares on the checkerboard, and the other rendering what would be the black squares on the board. ATI says this method distributes the load between the cards neatly and efficiently, but SuperTiling offers benefits only in terms of pixel-pushing power, not geometry computation. Both cards must compute the underlying geometry for each frame individually. SuperTiling is not supported in OpenGL.

Alternate-frame rendering — The king of all multi-GPU load balancing modes is alternate-frame rendering, or AFR for short. AFR interleaves full frames rendered by the two cards, so that, say, the master card renders odd frames and the slave card renders even ones. This is the preferred load-balancing mode whenever possible, because AFR shows markedly better performance scaling than other modes. Part of the reason for AFR's good performance is the fact that it splits the geometry processing load between the two cards evenly, something no other mode does.

2 cards doing independent things.

Uhhh..

They are not doing independent things, unless you have a crazy set up with both cards plugged into half of your display.

If this were the case the frame buffer would double with crossfire, it does not. The video memory (which has read and write last i checked.. thus the point to bring it up) has to be mirrored, as in the same on both. The cards do their thing, load their tasks (alternate frame or whatever) into the frame buffer (which is written to both cards memory). From there it is then put on the display through the ramdaq (or whatever we happen to be displaying with). Nowhere is the bandwidth doubled. In theory we could increase bandwidth based on the reads, but even so, the bandwidth of corssfire is far less than double the single card, yet we still see great scaling, if the single card was starved, the two cards should be even more starved.

 

[AzN]

But if VRAM is copied, that means that the same writes will have to be done for both cards. The reads can be different but since you have less than double memory bandwidth and doubled shaders, you would expect it to be just as memory bottlenecked which is not the case.

Not when cards independent from 2 different things the cards are rendering. They don't share the data from techreport's explanation.

If whatever Meaker10 says is true why don't we have an explanation of these things from crossfireX reviews? All reviewers double the bandwidth in crossfireX.

 

[AzN]

Uhhh..

They are not doing independent things, unless you have a crazy set up with both cards plugged into half of your display.

If this were the case the frame buffer would double with crossfire, it does not. The video memory (which has read and write last i checked.. thus the point to bring it up) has to be mirrored, as in the same on both. The cards do their thing, load their tasks (alternate frame or whatever) into the frame buffer (which is written to both cards memory). From there it is then put on the display through the ramdaq (or whatever we happen to be displaying with). Nowhere is the bandwidth doubled. In theory we could increase bandwidth based on the reads, but even so, the bandwidth of corssfire is far less than double the single card, yet we still see great scaling, if the single card was starved, the two cards should be even more starved.

That's why crossfireX needs a gaming profile to work. When it's mirrored like you propose you wouldn't even need a profile and should work without needing a profile much like a setup RAID device. Has to be? is something I have a problem with.

 

[Daedalus685]

Not when cards independent from 2 different things the cards are rendering. They don't share the data from techreport's explanation.

If whatever Meaker10 says is true why don't we have an explanation of these things from crossfireX reviews? All reviewers double the bandwidth in crossfireX.

Are you implying that all reviewers know what they are talking about?

Things are not doubled, sure, coloqually many reviewers will say that "crossfire is bringing 2*256bits of bus to teh fight" but that I ahve ever read woudl go so far as to state something like "The 5970 has a 512bit bus." It is technically correct to state that the 5970 has two 256 bit buses, but this does not mean that the effective bandwidth is doubled.

I could set up a switch that toggles a million LEDs on and off. I could then say that each of those is a bit. When i flip the switch would you say that I have a million bit bus? or only one bit? Technically I have the hardware to change the value of a million bits at once. Certainly I could argue that if a million different people were to look at each LED separately they would count a million bits every time I flipped teh switch. But every rational person woudl argue that since I am only conveying a single bit of meaningful information I only have a 1 bit bus. It is much the same in Crossfire.

Sure, the tech is there for two totally separate buses, but that is not how it works. Because the data is written twice the useful bandwidth on a write is only that of a single unit.

 

[Daedalus685]

That's why crossfireX needs a gaming profile to work. When it's mirrored like you propose you wouldn't even need a profile and should work without needing a profile much like a setup RAID device. Has to be? is something I have a problem with.

It needs a profile to split up the data stream to the card, so each card knows what it is supposed to do. It is still requirement on all of the frame buffer being stored in both sets of VRAM.

The memory sub system is one of the simplest parts of crossfire, not what the profiles are really aimed to solve. The profiles are so teh drivers know how to split up the work load and still ahve the game look presentable, or work at all.

 

[ugaboga232]

One of the bad parts of AFR and thusly Xfire and SLI is that both cards need ALL the data thus all the writes must be duplicated. Reads don't have to but still it isn't a doubling of bandwidth. Lucid Hydra is trying to stop this but as of now yes the VRAM must be copied on both cards and thus while the bandwidth does double, it has double the shaders to feed and there is lots of overhead as shown.

 

[AzN]

Are you implying that all reviewers know what they are talking about?

Of course not but when we've got every single reviewer including anandtech doubling memory bandwidth on crossfireX I kind of question who are you to tell them they aren't right when they get inside information to the tech they are reviewing.

Things are not doubled, sure, coloqually many reviewers will say that "crossfire is bringing 2*256bits of bus to teh fight" but that I ahve ever read woudl go so far as to state something like "The 5970 has a 512bit bus." It is technically correct to state that the 5970 has two 256 bit buses, but this does not mean that the effective bandwidth is doubled.

So who said 5970 has 512 bit bus? Which comes to the question above. Who are you to say the bandwidth hasn't doubled when all reviewers says it does?

I could set up a switch that toggles a million LEDs on and off. I could then say that each of those is a bit. When i flip the switch would you say that I have a million bit bus? or only one bit? Technically I have the hardware to change the value of a million bits at once. Certainly I could argue that if a million different people were to look at each LED separately they would count a million bits every time I flipped teh switch. But every rational person woudl argue that since I am only conveying a single bit of meaningful information I only have a 1 bit bus. It is much the same in Crossfire.

But you have 1 million 1bits much like when you have 2 256bit bus.

Sure, the tech is there for two totally separate buses, but that is not how it works. Because the data is written twice the useful bandwidth on a write is only that of a single unit.

It's not written twice when they are rendering 2 different things. If it was rendering the same thing that I would have to agree with you.

 

[AzN]

It needs a profile to split up the data stream to the card, so each card knows what it is supposed to do. It is still requirement on all of the frame buffer being stored in both sets of VRAM.

The memory sub system is one of the simplest parts of crossfire, not what the profiles are really aimed to solve. The profiles are so teh drivers know how to split up the work load and still ahve the game look presentable, or work at all.

But in a RAiD mirror device you do not need a profile to read and write data.

Split up the data stream. That's the key word here. Rendering is being split and not shared between loads by each card.

 

[ugaboga232]

Azn clearly has no clue what AFR and XFIRE and SLI mean. They have to have all the same stuff in framebuffer meaning the extra bandwidth does not matter.

 

[Daedalus685]

You don't understand how crossfire works.

Think of the video memory as a different system than the GPU. The GPUs are rendering different things, but the two GPUs are sharing the SAME frame buffer.

Any rendering that GPU #1 does must be written to the frame buffer. Since the frame buffer is stored on both cards this has to be written twice.

who said 5970 has 512 bit bus? Which comes to the question above. Who are you to say the bandwidth hasn't doubled when all reviewers says it does?

Reviewers do not say it has double the useful bus, they say it has two of them. These are entirely different things. The 5970 is reported as 2*256bit, not as 512bit, there is a huge difference.

I am not saying the bus has not doubled as it clearly has. I am saying the useable bandwidth has not doubled, no good review has ever claimed that it does. As to who I am to claim such a thing.. I don't know.. I'm Dan, and I'm an applied physicist.. while not a computer engineer I'd like to think I know something of the topic.

 

[AzN]

Your experience has little to do with this discussion unless you can produce numbers like bfg has. I assume you mean memory bus, as the total amount means nothing to this conversation. Minimum FPS will be higher if bandwidth is not an issue, as has been explained when you run into a scene that is bandwidth starved the performance becomes muted. This is obvious. The point is that in this situation that behavior is not seen. The minimum is very erratic.

LOL schmide. You use same words and mention hardocp. 1+1=2 not 1.5

That is why I mentioned other reviews from your favorite xbit and stop being so lop sided with BFG's average frame results to conclude bandwidth isn't playing the role here.

Minimum is not erratic when you have multiple games showing the exact same thing.

 

[ugaboga232]

Azn show us one bit of solid data that the 5800 is memory bandwidth bottlenecked. Until then do not change the discussion into one about fake accounts.

 

[Daedalus685]

But in a RAiD mirror device you do not need a profile to read and write data.

Split up the data stream. That's the key word here. Rendering is being split and not shared between loads by each card.

I'm sorry AZN, I don't know how else to explain it. You just don't seem to understand how crossfire/sli work.

Don't think of crossfire like raid. It is not like raid at all. Raid was only mentioned because the frame buffer system in crossfire works akin to raid1. Before anyone goes claiming how different they are I know.. but the point is the data is written twice for every requested write, and may provide you with twice the speed with which to read.

The profile is telling the driver how to instruct the gpu's to render. Some games may not render the same as others, they may not like to be split up frame by frame, and there might be other calculations that have to be made that do not logically split up into "frame 1 and frame 2." This has nothing to do with the technology of the memory. It is just how it works that both cards need to duplicate memory, as they need to write and read from the same place in order to work.

 

[Daedalus685]

LOL schmide. You use same words and mention hardocp. 1+1=2 not 1.5

That is why I mentioned other reviews from your favorite xbit and stop being so lop sided with BFG's average frame results to conclude bandwidth isn't playing the role here.

Minimum is not erratic when you have multiple games showing the exact same thing.

What?

What does schmide mean? Now who is being personal.. I'll assume it is slang for something non offensive as I have never heard it before, and give you the benefit of the doubt.

I have no idea what you are on about. You will have to explain this again.

 

[ugaboga232]

Daedalus it sounds like schmide is another poster and he believes you are an alternate account of his.

 

[AzN]

You don't understand how crossfire works.

Actually I do have basic idea how crossfireX works but you seem to put shove everything in my throat because you said so.

Think of the video memory as a different system than the GPU. The GPUs are rendering different things, but the two GPUs are sharing the SAME frame buffer.

Any rendering that GPU #1 does must be written to the frame buffer. Since the frame buffer is stored on both cards this has to be written twice.

That's the thing crossfire isn't sharing the same frame buffer. Either alternate frame buffer or super tiling where 1 card would be rendering red checkers and the other rendering black checkers.

Reviewers do not say it has double the useful bus, they say it has two of them. These are entirely different things. The 5970 is reported as 2*256bit, not as 512bit, there is a huge difference.

Did I say 5970 has 512bit bus? I don't know why you keep saying the same thing over and over again when I specifically in the other thread said 2 128bit bus.

I am not saying the bus has not doubled as it clearly has. I am saying the useable bandwidth has not doubled, no good review has ever claimed that it does. As to who I am to claim such a thing.. I don't know.. I'm Dan, and I'm an applied physicist.. while not a computer engineer I'd like to think I know something of the topic.

They put up double gb/s numbers though including techreport, anandtech, and so forth...