Crossfire benchmarked, X700 and X300 too, apparently

[Rys]

http://www.theinquirer.net/?article=23593

Here's the actual article. Thinking back on Josh's interview with NV, maybe they're after making the pcie connection wide enough to do away with the intra-card connector without impacting performance? Folks might like that. Don't know what it would do to compatibility, if anything.

I had a good read of that linked thread last night and again this morning, since we've had some Crossfire numbers for a while that I'm now thinking about writing up.

It's geo's quote about using unused PCIe lanes for the inter-GPU communications that's been bugging me, as I think about SLI, any possible 'SLI2', and Crossfire as a whole.

The PCI Express specification dicates you negotiate 1x, 2x, 4x, 8x, 16x or 32x lane appropriation when the devices talks to the root port and announces itself on the bus. Following on from those lane bundles, they're the only electrical slots you can have, too.

So to appropriate spare lanes in a PEG16X slot just for inter-board communications, that means a new set of mainboards that put 8X from the slots to the root port, for normal working, and the back 8 lanes from the slots are routed to each other, for the cards to talk over. That's the first lane split you can do with a PEG16X slot, since there's no way to negotiate say the first 14 lanes, then just rebundle the back 2. It's a power of 2 for splits and reassignment.

And if 8 lanes are being pushed that way, each GPU would need another PCI Express controller that also acts as a pseudo root port, as far as I can tell, so the negotiation can be made, unless there's some extra logic on the mainboard that isn't there just now.

In addition, pushing that data over PCI Express, as Xmas has said I think, doesn't guarantee any explicit timing of the transfer of data, so you can't rely on information arriving within a certain time frame. And timing of buffer pushes and whatnot would likely have to be as tightly controlled as possible. Hence the custom Xylinx FPGA on ATI's boards, for controlled composition.

I think it's more likely that the connecting bus that currently exists between the two boards will just get faster and/or wider, and we'll continue to see it for any SLI2 implementation that maybe needs a bit more bandwidth on that connector.

As long as the inter-GPU connectors continue to be bundled with either graphics board (buying two graphics cards and having a spare connector wouldn't be the worst thing in the world, I think) or mainboard as they are now, I don't see the big problem with connecting the boards up that way for the forseeable future and a further generation of SLI.

 

[Hanners]

As long as the inter-GPU connectors continue to be bundled with either graphics board (buying two graphics cards and having a spare connector wouldn't be the worst thing in the world, I think) or mainboard as they are now, I don't see the big problem with connecting the boards up that way for the forseeable future and a further generation of SLI.

Speaking of such things (and a little off-topic), I don't suppose you (or anyone else) has had the opportunity to pair up two 6600 non-GT's without the inter-GPU connector with NVIDIA's latest driver set yet?

I'm kind of curious as to whether they match up against my own numbers with a couple of 6600GTs.

 

[dizietsma]

As long as the inter-GPU connectors continue to be bundled with either graphics board (buying two graphics cards and having a spare connector wouldn't be the worst thing in the world, I think) or mainboard as they are now, I don't see the big problem with connecting the boards up that way for the forseeable future and a further generation of SLI.

Speaking of such things (and a little off-topic), I don't suppose you (or anyone else) has had the opportunity to pair up two 6600 non-GT's without the inter-GPU connector with NVIDIA's latest driver set yet?

I'm kind of curious as to whether they match up against my own numbers with a couple of 6600GTs.

None GT 6600's, when they get to 500-600-700Mhz, are terribly bandwidth limited, so I can imagine that doing tests with SLi on those would also have to take into account this also.

Did you try varying PCie speed to see if that had any influence ?

Regards

Andy

 

[Hanners]

Did you try varying PCie speed to see if that had any influence ?

Nope, literally just ran it 'as is' both with and without the bridge connector.

 

[KimB]

Those benchmark results seem somewhat questionable, considering that in so many of the benchmarks Crossfire is so close to exactly double the performance of one board, and in some cases better.

 

[Geo]

Interesting analysis, Rys, thanks for taking the time. I take it that "SLI2" seems a reasonable possibility from your pov. Not being all that technical, I was approaching it from a marketing/competitive position direction now that ATI is about to be in the game, and assuming that NV would like to see three "goods" (from their pov) happen.

I saw those as:

1). Increased performance
2). Moving proprietary stuff into the mobo chipset and off the graphics card, simplifying the cards and setup/install by getting rid of the connector.
3). Given their much stronger mobo position, having a legitimate technical reason that increases performance for SLI that also makes it difficult/impossible for ATI XFire boards to work in SLI2 mobos. I don't know that moving stuff into the mobo chipset would do that, I just think NV would like it to happen, and happen in a way that would reduce criticism of them --or at least give them a decent-looking reason for it when the criticism does happen. I suppose it works the other way too, that SLI2 cards wouldn't work, or as well, in ATI mobos --and I think NV would think that just fine.

 

[_xxx_]

The connector is here to stay. There's no way you'll ever achieve the same throughput through PCI-E only in todays mobos. Maybe they'll have a few lanes dedicated solely to SLI in the next iteration, that might be the solution but could also introduce a batch of other problems.

 

[Geo]

I think it's more likely that the connecting bus that currently exists between the two boards will just get faster and/or wider, and we'll continue to see it for any SLI2 implementation that maybe needs a bit more bandwidth on that connector.

As long as the inter-GPU connectors continue to be bundled with either graphics board (buying two graphics cards and having a spare connector wouldn't be the worst thing in the world, I think) or mainboard as they are now, I don't see the big problem with connecting the boards up that way for the forseeable future and a further generation of SLI.

So you're leaning "SLI2" is a wider intracard connector? Any other content in the "SLI2" bucket, ya think? The Inq article seems much more focused on mobo changes.

 

[Rys]

So you're leaning "SLI2" is a wider intracard connector? Any other content in the "SLI2" bucket, ya think? The Inq article seems much more focused on mobo changes.

Yeah. The clock frequency of that little bus is fairly high (hundreds of MHz) from what I can tell and I don't think it's that wide (pin count at least hints at that). So I think there's still room for development there, and NVIDIA will use an inter-GPU connector again for any SLI2, as they use it now.

 

[Geo]

I had a good read of that linked thread last night and again this morning, since we've had some Crossfire numbers for a while that I'm now thinking about writing up.

http://www.hexus.net/content/reviews/review.php?dXJsX3Jldmlld19JRD0xMzUx

That was quick! I'm not going to have to sic Digi on you am I, for how it really is okay to pimp your own stuff here when it is very much on subject? Personally, one of the things I love about B3D is that everyone points at the interesting articles elsewhere. . .saves me a goodly bit of tramping about.

 

[digitalwanderer]

I'm not going to have to sic Digi on you am I, for how it really is okay to pimp your own stuff here when it is very much on subject?

Pfffft! Rys is free to pimp HEXUS all he wants, it's a great site and I pimp it regularly myself.

There isn't anything wrong with a little pimping, as long as it's pertinant to the subject at hand. Then it sort of transcends "pimping" and becomes "informative".

Rys is almost always the latter, he's good peeps.

 

[Geo]

I know that. You know that. Rys apparently needs some encouragement. 8)

 

[Rys]

That was quick!

You linked to it, not me I probably wouldn't have, especially since I personally wasn't going to publish for a day or so, but it looks like we did anyway. Anyway. the interesting stuff is really in the data I chose to keep back

As for the pimping, no need, you should all read my crap anyway by default

 

[Geo]

Assuming that:

1. ATI AFR doesn't introduce anamolies (obviously you can't see it without the second output).
2. hkepc's chart really does prove what they assert --that the benchies don't change for AFR in slave/slave vs master/slave even if the compositing isn't there.

Then:

Am I right in reading that 12x10 4xaa/16xaf. . .is a 50% can of whup-ass that X850 XF/AFR is opening on 6800u SLI/AFR?

Is that right? And based on your comment of 10k for 7800gtx sli in that setting, that's about 250% faster for 7800 vs 6800?

 

[pat777]

Am I right in reading that 12x10 4xaa/16xaf. . .is a 50% can of whup-ass that X850 XF/AFR is opening on 6800u SLI/AFR?

I see 12*10 but not 12*10 4xaa/16af.
Edit: I thought you were talking about the link from the original poster.

 

[Rys]

Am I right in reading that 12x10 4xaa/16xaf. . .is a 50% can of whup-ass that X850 XF/AFR is opening on 6800u SLI/AFR?

After you wrote that something didn't seem right, so I went back to check the numbers. I'd not only reported a single-card score for the 6800 Ultra SLI, but I'd also neglected to mention the Crossfire box was running with 6xAA, not 4xAA.

There's a new graph that shows the corrections. Changes things a bit, sorry for the mixup.

 

[Geo]

Ah. Okay. I wanted to be excited. . but the more I thot of the 7800sli vs 6800sli relationship in that graph (well, the 7800sli is stated rather than "in that graph", but you know what I mean), the more suspicious I was getting.

 

[Geo]

And I'm mentioned on the front page of Hexus, with no finger pointing and/or laughing involved.

My career in the graphics community can only go downhill from here.

 

[CJ]

I'd not only reported a single-card score for the 6800 Ultra SLI, but I'd also neglected to mention the Crossfire box was running with 6xAA, not 4xAA.

The statement "The AFR-driven slave-slave is nearly twice as fast as Crossfire Supertiling" doesn't seem right to me. In that chart the X850XT CrossFire scores 3016, while in Tom's Hardwares latest VGA Charts, a single X850XT scores 3367 points at 1600x1200 with 4xAA/8xAF.

Are you sure that the 3016 score is a CrossFire (supertiling) setup? I've seen people reach around 4000 on a single X850XT PE on those settings (1280/1024 6xAA/16xAF). Or am I missing something here or completely misinterpreting the chart?

 

[Rys]

I've double checked and the data I have is the data that shows up in the graph. I'd say driver issue or testing quirk, since it does seem out of place. We'll see in due course I guess. Wish I'd done the testing myself, but I was elsewhere at the time DFI had a slot for us.