Crossfire Without Mastercards Explained

[perkam]

Disclaimer: This is from Tech Report's article here: http://techreport.com/onearticle.x/8390 ...also I'm far from an expert on this, I just report what i see

Exhibit A: Different Configuration without Mastercards:

Curiously, the first thing DFI showed us was what they described as a pair of slave cards running together cooperatively. The DVI output to the LCD display was only hooked to one of the two video cards, and the system was running 3DMark05. I asked how the cards were sharing data, and the gent pointed to the north bridge chip between the two cards.

So apparently, the connector to the monitor is only connected to ONE of the cards, the other card is merely placed without ANY connections in the PCI-E slot....so they're sharing data through a mysterious NB which may or may not show up in the CF-DR...but most likely will.

Exhibit B: AFR with only half the frames

I suppose a CrossFire master card with an image compositing engine isn't strictly necessary for multi-card acceleration. He said that you could hook up the monitor to the DVI output on the secondary slave card and see only half of the frames when the cards are rendering in alternate-frame mode

The image compositing engine is the invisible sli bridge if you will, that transfer the data between the two cards when using the dongle. As it suggests here, its not necessary for crossfire, but be warned using AFR or alternate frame rendering (the only mode of crossfire compatible with non-mastercard crossfire) that the frames will skip. In my opinion, the frame skipping shouldn't be too noticeable when one scene is being rendering in excess of 60fps...but i could be completely wrong and it could look like jittery as hell. Point is, you can still benchmark using two slave cards, and magnificiently too...apparently, 2x the score in 3dmark05 is not a problem with two non-mastercard x850xts.



**Requirement for this as stated in the article is that for "dual slave configuration" to work, you need two of the exact SAME cards, down to the exact same bios, along with a crossfire ready mobo for this to work.

Exhibit C: The right way and the wrong way

The dude from DFI then swapped in a true CrossFire master card from ATI alongside one of the the X850 XT PE cards and proceeded to flash the slave card to what he said was a Radeon X800 BIOS, so it would match the CrossFire card. Both had dual-slot coolers that looked for all the world like X850 XT cards, but whatever. He then ran 3DMark for us again, this time with the video cables daisy chained as one would expect for a CrossFire config.

Apparently, this must have been pretty early in the game for crossfire to require flashing the bios EVEN WITH a mastercard. THe config for normal crossfire is shown here. The configuration of the connectors is too ugly for me to post here also because they've refined it a bit on crossfire's second launch as shown in this instructional flash video: http://www.ati.com/technology/crossfire/images/demo.swf (flash movie looks better when size of window its viewed in is ~640x480)

That's all from Tech Report...off to T-break for some:

Exhibit D: Crossfire Nex Gen NEW !!

The following is from an interview an an ati official and thus deemed very reliable.

T-Break: Is there any way to have two cards work together without a compositing engine? D. Nalasco, ATI - Yes- in fact the x1300 does just that.

Yup. Value owners rejoice. Explanation:

D. Nalasco, ATI - Basically after you draw each frame you have to transfer the results- either half of the frame or the whole frame using AFR to the other card. But if your total frame rate is not that high then that’s less frames to transfer. Even on the x1300 we have to do some special things in our driver to minimize the amount of data that goes over the bus

Apparently, the data that is shared between two X1300 is miniscule and thus does not require an interlink for crossfire to work in full AFR mode. This makes it very interesting as initial results show an X1300Pro outperforming x700s in benchmarks and real world performance.

Exhibit E: Slave Crossfire and the difference between High, Mid and Low End NEW !!

D. Nalasco, ATI - On an x1800, if you plug them in and not have a link and we were to enable a software crossfire mode over PCI Express, we could still get games to work but they’d be 20-30% faster with the second card which is not really that interesting. You’d want 2x or 80% (gains) and to get those we obviously need an interlink. With the x1300 we could get that 80% in games with PCI Express using some optimizations.

Interesting. With an X1800, we're not talking about NO performance increase WITHOUT an interlink (which means dual slaves) we're still going to see a 20-30% increase in performance ! Sure, that may not give enough of a boost to counter GTX SLI right now, but for those with multiple X1k cards lying around and a crossfire mobo, we're talking ~11k stock (based on % estimate) in 05 without the currently non-existant x1800 "xfire edition" card.

He then mentions that the X1300 will be able to get that elusive 80% performance increase in games !! without an interlink ! Seeing as its official confirmation by ATI and that the x1300 will be at a ~$99 pricepoint means ATI has certainly got something in store for value enthusiasts.

Confirmation that this configuration isn't affected much by dual 16x lanes for 1600XT and up:

T-Break: What happens when we use 16X for both cards. Do we have enough bandwidth at that time to power higher-end cards? D. Nalasco, ATI - I would say, its unlikely but we have to do more experimentation and see how that works out.

T-Break: Until your next generation comes up and within that time we’re likely to have two x16 lanes on a motherboard. D. Nalasco, ATI - The x1300 will definitely run better with that configuration (but) the x1600 is quite a bit faster. Its got three times the shader processors of x1300. Assuming games are limited by shader processing, its gonna be a fair bit faster so its unlikely it would work well on the x1600 without the interlink.

There you have it...not much interest in dual x16 lanes and recent launch of dual 16 lane sli mobos confirmed that we dont really need them, even for the high-end.

That's all for now. Hope ppl are as excited as I am at the possibilities of crossfire with or without the mastercard. Thx to Tech Report and T-Break for their journalistic inspiration. This analysis is far from perfect and will likely undergo a few changes, but good stuff nonetheless.

Perkam

[n00b 0f l337]

Perkam lately all I have to do is search for your posts to really get a good read. Thanks for the info!

[P_1]

Tehnically you wouldnt even have to have another slave card as the alternate frames get thrown away, all you would need is some card to feed information that the driver for crossfire is looking for so that it will initiate the afr mode anyway. Also this can probably be done with a driver hack where they just make it think there are 2 cards anyway. I do not believe this is a good solution other then for benching, as its not a true dual solution... ur missing half of the frames, its sort of like cheating ur self.

[n00b 0f l337]

Thats the problem I posted in the thread about crossfire and dual slaves, arent we going to see a lot of cheated benches now?

[perkam]

Thx for the complement guys. New info though. It seems when you have the will to explore something, the information just comes to you

I just read a boat load of official new info on crossfire with and without mastercard including such possibilities FOR THE X1000 series from T-Break's interview with David Nalasco here: http://www.tbreak.com/reviews/articl...8&pagenumber=1 .... and will be adding to the post in a moment

Perkam

[Mr. Tinker]

I'm not trying to argue with you perkam, but it was revealed a couple days after that trade show that DFI's crossfire demo was "bunk," and that it was running a demo crossfire driver for motherboard makers to simulate crossfire performance. It wasn't true crossfire and we won't be able to get our hands on that driver. Sorry, I really wanted this to be true but it just isn't. When I find the articel that specifically talked about the tech report article and how it was wrong I will post it.

Edit: here we go:

http://www.ocworkbench.com/ocwbcgi/n...21995844,44289,

Lately, there is a website that posted results of CrossFire claiming that you can just flash update a normal X800 to enable CrossFire.

We spoke to a friend from ATi and he told us that there was indeed a driver that was provided to the distributors as a "proof of concept" driver to showcase the anticipated functionality and performance of our product without the compositing chip - by rendering odd frames only. It did not require primary or secondary crossfire boards. The current X800/X850 cards used do not contain the composition chip.

Again, this driver was to give motherboard partners ability to test dual graphics while bringing up their Radeon Xpress Crossfire Edition motherboards.

This driver will not allow full functionality of crossfire.

The only products that will function correctly are a crossfire edition X8XX based board (with compositor chip built in) w/ regular X8XX, so it can communicate with additional board via external dongle.

X700 are not currently crossfire supported. Leaking this driver will only cause confusion as it is not endorsed by ATI nor is it Crossfire ready.

[perkam]

Mr. Tinker, np. I stated several times that i cant be completely right...but the nex gen section of my review wont be as unreliable

Perkam

[Mr. Tinker]

updated my post

[nn_step]

But is there a way to do that on an Sli board for the sole purpose of extreme folding?

[perkam]

First post updated with X1k info

updated my post

We already know that it wont enable full cf functionality, thats what the thread is all about And HKEPC has already tested dual slave xfire with cards ranging from x300, x700, x800pro all the way up to x850xt, all without any mastercard, so we defenitly know it works , even if not as well.

But is there a way to do that on an Sli board for the sole purpose of extreme folding?

Nope As i said, at the very least, a commitment to a crossfire motherboard is required for this to work.

Perkam

[P_1]

First post updated with X1k info



We already know that it wont enable full cf functionality, thats what the thread is all about And HKEPC has already tested dual slave xfire with cards ranging from x300, x700, x800pro all the way up to x850xt, all without any mastercard, so we defenitly know it works , even if not as well.


Nope As i said, at the very least, a commitment to a crossfire motherboard is required for this to work.

Perkam

Also you should tell people not to buy 2 slave cards right now as its a waste of money, you can acheive the same results by hacking the drivers to go into crossfire afr mode so it will just drop the alternate frames.

[Grayskull]

The driver that DFI showed is indeed a true Crossfire driver just without the compositing done at the last stage (obviously). All frames were rendered using the 2 cards just half of them are displayed.

I'm not trying to argue with you perkam, but it was revealed a couple days after that trade show that DFI's crossfire demo was "bunk," and that it was running a demo crossfire driver for motherboard makers to simulate crossfire performance. It wasn't true crossfire and we won't be able to get our hands on that driver. Sorry, I really wanted this to be true but it just isn't. When I find the articel that specifically talked about the tech report article and how it was wrong I will post it.

Edit: here we go:

http://www.ocworkbench.com/ocwbcgi/n...21995844,44289,

[perkam]

Thanks Grayskull, Tech Report's findings were accurate then. It's intriguing though. I was under the impression that the other frames were just "skipped" entirely and not even rendered with the lack of a compositing engine. SO its just a display issue then...nice.

Though would you be able to tell us if the frame-skip is really that noticeable at high framerates ?? And hopefully David Nalasco's information is accurately portrayed here too

Also you should tell people not to buy 2 slave cards right now as its a waste of money, you can acheive the same results by hacking the drivers to go into crossfire afr mode so it will just drop the alternate frames.

Look again Buying two X1300s can allow you to use crossfire without a hitch

Perkam

[Revv23]

so whats the point, sounds like cheating to me, and why would you game with it, why have two card redering 60 but your only seeing 30 when you could just have one card rendering a smooth 30?

[Mr. Tinker]

Thanks Grayskull, Tech Report's findings were accurate then. It's intriguing though. I was under the impression that the other frames were just "skipped" entirely and not even rendered with the lack of a compositing engine. SO its just a display issue then...nice.

Though would you be able to tell us if the frame-skip is really that noticeable at high framerates ?? And hopefully David Nalasco's information is accurately portrayed here too


Look again Buying two X1300s can allow you to use crossfire without a hitch

Perkam

What's the point if you can only see half the frames?

[perkam]

Uh... With the X1300, you dont skip frames even without a compositing engine or an interlink!! That's the whole point. Secondly, you'll still see a 20-30% HIGHER frame rate even with frame skipping !!, on dual x1800s.

For those asking whats the point, this analysis was meant to explore the dual slave configuration for x8xx and x1xxx line of graphics card, not meant to say that its a definite alternative to crossfire with a master card

Perkam

[cadaveca]

Yes, they are transferring frames via the pci-e bus, Perkam, but the main thing you missed in the interview with David was that this ONLY WORKS WHEN SHADER DATA EXCHANGED IS AT A MINUMUM. This means that superAA and such will not properly work, as the slave card cannot push enough data through the bus.

This however, is only for the 8xPCI-Ex2, as as far as i understand it, the x1300 only needs 4 lanes, and 4 are for cross-communication. Once 16xPCI-Ex2 chipsets come out, for example, the newer cards will not be as limited, as is the current situation. You fail to see that the move IS GOING TO 16x PCI-E, but it's not gonna happen mainstream just yet, so is not CURRENTLY viable. that can change by the end of the month tho.

Its(it=x1600,]got three times the shader processors of x1300. Assuming games are limited by shader processing, its gonna be a fair bit faster so its unlikely it would work well on the x1600 without the interlink.

[perkam]

Well, the ati rep certainly didn't care much for dual 16x pci-e lanes...but perhaps your right in that it will move more data between the cards. Also, yes shader data has to be at a minimum, but an 80% increase in performance means it can keep data being sent across the lanes at the minimum required level while not foregoing performance.

Perkam

[cadaveca]

I'm just awaiting shared framebuffers over the bus. AFR works because a rendered frame equates to very little...6-12mb...but @ 60FPS, that's 300-750mb/second!! @ 200mb/second per lane, i think, that's cutting the budget pretty close. it might be only 100mb/sec per lane.

Although the timing for Crossfire/SLi without external connections is close..it's a bit far off. I'm sure they have working samples already, and are weighing in the cost/benefit ratios already, and what gets decided we will see in 18-24 months.


BTW, as far as i understand it, this is how the 6200 and 6600's from the other side are communicating without a bridge, so i don't quite understand why all the enthusiasm. As David said, and you pointed out earlier yourself, Perkam, it's not really a marketable benefit for this generation, but maybe for the next. It really depends on what Intel and AMD bring with thier newest chipsets for up-and-coming sockets.

[Mr. Tinker]

Uh... With the X1300, you dont skip frames even without a compositing engine or an interlink!! That's the whole point. Secondly, you'll still see a 20-30% HIGHER frame rate even with frame skipping !!, on dual x1800s.

For those asking whats the point, this analysis was meant to explore the dual slave configuration for x8xx and x1xxx line of graphics card, not meant to say that its a definite alternative to crossfire with a master card

Perkam

but without the compositing card you only SEE the frames rendered on the card hooked up to the monitor, and not any frames rendered by the other card. So in other words, if both cards collectively manage 60fps in AFR, you will only get to see 30fps. See why this is stupid?

[perkam]

if both cards collectively manage 60fps in AFR, you will only get to see 30fps.

Not quite as I've said before: There is no performance drop relative to one card as a result of using this method. So if one X1300 manages 26 fps in a game at 1024x768, then with crossfire, you're looking at approximately ~40 fps in the same game. With the X1800 this difference is somewhere along the line sof 60 fps to 65 fps perhaps, in real world performance.

What apparently Mr. Nalasco didnt comment on is the capabilities of this configuration for 3dmark05, which benefits a lot more irrespective of the lack of the interlink...as you see in the x850 section of the analysis - 10.6k stock in 05 with dual slaves.

Perkam

[Mr. Tinker]

Not quite as I've said before: There is no performance drop relative to one card as a result of using this method. So if one X1300 manages 26 fps in a game at 1024x768, then with crossfire, you're looking at approximately ~40 fps in the same game.

You're not getting what I'm saying. In the scenario you described (assuming rendering method is AFR) 40fps would be rendered collectively by both cards, but 20fps would be displayed by the monitor.

So if the single card does 26fps, and mastercard-less crossfire renders 40fps, you will get to view 6fps more with the single card.

What's the point of this? Higher synthetic scores? You're assuming that the public drivers will allow you to run Crossfire without a MC, which they probably won't. The drivers HKEPC and DFI ran were not public.

T-report was the unfortunate recipient of misinformation in this case.

[cadaveca]

YOU are missing the point...drivers aside, the hardware does not need the composting chip. The problem at hand that makes the use of the chip nessecary is the lack of bandwidth available on the motherboard. This is why the new revision of the chipsets was apparantly making some OEM's thinking of holding off on release of boards until late this year...rumour had it that they would have 2x 16lanes pci-e to each card, which should provide enough bandwidth for decent scores without use of the composting chip. This would make both the motherboards and the videocards cheaper to produce, but fact of the matter is, old stock must go first.

[Vapor]

How do you suppose a card will composite the two data paths if there's no compositing chip? What it MAY be able to go without is the dongle, but certainly not the compositing chip! Unless there is CORE logic to composite the two streams in R5xx (like there is in NV4x and G70), then you'll never see more than one card's stream without the compositing chip and dongle.

It's all well and good to not have the compositing chip for benchmarks where you don't need a playable framerate, just as high as it thinks it's going, but for games you need to composite the damn images--there's currently no way to do that on ATi hardware without the compositing chips.

[cadaveca]

meh..you're kinda right...it's not the lack of the need for the composting chip, but of the dongle.

D. Nalasco, ATI - On an x1800, if you plug them in and not have a link and we were to enable a software crossfire mode over PCI Express, we could still get games to work but they’d be 20-30% faster with the second card which is not really that interesting. You’d want 2x or 80% (gains) and to get those we obviously need an interlink. With the x1300 we could get that 80% in games with PCI Express using some optimizations.