Crossfire limitation

[Jawed]

Ooh, another thing, if the CrossFired cards can't get above 60fps at 1600x1200, say, that's going to really hurt in benchmark comparisons, isn't it?

Since benchmark comparisons "rely" on stupidly high spikes of 200fps (etc.) to drag the average upwards, CrossFire is at an inherent disadvantage.

Jawed

 

[Jawed]

So, we have a limitation that was known about in June (hell, I'm on that thread :smile: )...

Blimey, so was I!

In fact I started that thread. And I have no memory, whatsoever, of this issue being raised before

Jawed

 

[Dave Baumann]

This is perhaps a stupid question but... couldn't you use the dual-dvi output of the slave card to get around this issue?

Read my earlier message - exactly how many of the Slave boards actually have Dual-DVI?

In CrossFire there appears to be a 1600x1200/1920x1080-ish @ 60Hz limitation for framebuffer data.

Framebuffer update rate does not equal CRT refresh rate.

But anyone with a 2560x1600 or 2048x1536 display is so out of luck.

The way I see it is the max 3D resolution will be 1600x1200/1920x1080 because this is what the Slave can transmit. The CRT refresh rate will be dictated by the RAMDAC, the LCD (DVI) frquency is ditacted by the TMDS, but then you were already running that over a single link TMDS in non-crossfire mode.

Ooh, another thing, if the CrossFired cards can't get above 60fps at 1600x1200, say, that's going to really hurt in benchmark comparisons, isn't it?

Although I said the composite device is acting as the final frame buffer, each of the boards still have their own frame buffer so they will be rendering to those at the 3D FPS rate, and that's what is reported by the benchmarks. Images displayed != frames rendered (and this was really always the case).

 

[Rys]

WRT the refresh rate, why are CRT's going to be disadvantaged by a TMDS? This is for digital display, is it not? Surely the analogue refresh rate would be dictated by the RAMDAC?

Because there's no analogue out? You convert DVI to VGA if you want with a convertor, but that means you take on the limitations of the DVI host output since there's no explicit DAC.

As for the rest of the chatter in the thread...

And because it's a Sil1162 and follows the PanelLink spec (well, Silicon Image did invent it), it can probably be configured by the video host to drive 1920x1200 @ 60 Hz (154MHz DVI pixel clock) and 1920x1080 (HD, using panel blanking at 131MHz DVI pixel clock) just like nearly every other single-link TMDS!

Why do you think today's popular 24 inch panels with > 1600x1200 are drivable with single-link DVI ports?

The single-link parts on compatible Radeon 'slave' hardware can be configured to drive those entirely single-link compatible modes. Connect a 2405FPW to an X800 sometime

I'm not saying that ATI will program the TMDS transmitters on the slave and the 1162 on the Crossfire master to do that, but it's certainly technically possible and inside of single-link DVI transmission rates.

 

[Dave Baumann]

Because there's no analogue out? You convert DVI to VGA if you want with a convertor, but that means you take on the limitations of the DVI host output since there's no explicit DAC.

There's a DAC sitting next to the composite chip.

 

[Jawed]

Although I said the composite device is acting as the final frame buffer, each of the boards still have their own frame buffer so they will be rendering to those at the 3D FPS rate, and that's what is reported by the benchmarks. Images displayed != frames rendered (and this was really always the case).

But you can't afford to have tearing in a composited frame. That'll prolly generate "holes" of mismatched compositing. So in any compositing mode I'd expect the frame rate to be capped because the two cards have to render their portions of the frame in lock-step - even if they take unequal time to do so.

In CrossFire the framerate cap is set by the resolution, courtesy of the DVI standard.

Though AFR mode, where the images are simply alternated (and which will presumably be the highest performing mode, where available in a game) obviously doesn't do any compositing.

Jawed

 

[Rys]

There's a DAC sitting next to the composite chip.

Makes sense since the FPGA is the final output device and doesn't have its own DAC. In truth, I didn't ponder where the analogue feed would come from, from the FPGA

But it won't help you bypass the digital limitations in Crossfire mode, either, which is the crux of the issue.

 

[Kombatant]

Question (and if it's rubbish, excuse moi, I haven't had any coffee yet): Isn't this 60Hz "limitation" thing like saying that, if you connect your CRT to the DVI-Out of the card (via a DVI-to-VGA converter) you are limited to 60Hz? Which means it's totally irrational?

 

[KimB]

Although I said the composite device is acting as the final frame buffer, each of the boards still have their own frame buffer so they will be rendering to those at the 3D FPS rate, and that's what is reported by the benchmarks. Images displayed != frames rendered (and this was really always the case).

Except the composite device doesn't necessarily have to act as the final frame buffer...it may just have a small buffer that is big enough to correct any slight timing differences between the cards, and just output the incoming data stream as it enters the compositing buffer.

For performance, it would seem that the above would be the thing to do. No reason to add extra latency if it's not needed. And besides, running such a display at 85Hz when the output from the cards is at 60Hz would be really bad anyway: you'd be forced to 42.5 fps! (Not enabling vsync, as mentioned before, would probably be really bad for either supertiling or split frame rendering).

 

[Jawed]

Question (and if it's rubbish, excuse moi, I haven't had any coffee yet): Isn't this 60Hz "limitation" thing like saying that, if you connect your CRT to the DVI-Out of the card (via a DVI-to-VGA converter) you are limited to 60Hz? Which means it's totally irrational?

No, because the DVI-I connector on the back of a video card carries the analogue VGA signal in pass-through to the VGA connector. That's what the "+" is at the side of the grid of pins.

Jawed

 

[Xmas]

One more thing I'd like to quickly reiterate:
Alternate frame rendering shouldn't be nearly as limited as the other modes, provided ATI can fudge it so that the cards will output at strange refresh rates like 37.5Hz or 42.5Hz (which, when combined, will result in 75Hz and 85Hz, respectively). But the resolution will still be limited to 1600x1200, which a single R520 again should do just fine.

I don't think so. This would require buffering of whole frames, and I don't think the composition chip has access to any significant amount of memory.

 

[Kombatant]

No, because the DVI-I connector on the back of a video card carries the analogue VGA signal in pass-through to the VGA connector. That's what the "+" is at the side of the grid of pins.

Jawed

Thx for the clarification man

 

[Geo]

Read my earlier message - exactly how many of the Slave boards actually have Dual-DVI?

All of the X1800 ones? [he asked, hopefully & suggestively]

 

[Karma Police]

Well, the Inq has some "good" info:

Crossfire does 1600x1200@60 or 1920x1200@52 with pictures to prove it

 

[KimB]

I don't think so. This would require buffering of whole frames, and I don't think the composition chip has access to any significant amount of memory.

Oh, you're right. Didn't think about the time it would take to display an image at the lower refresh rate. Looks like even alternate frame rendering will be affected: in this mode, each card's every other frame will just be blank.

 

[JoshMST]

Haha, do I dare post this?

http://www.penstarsys.com/#xfire_sil

I really think that it is essentially a non-issue. We are dealing with bandwidth here, and there is no reason why the master card can't output higher resolutions and refresh rates because the receiver is only the Sil 1161.

Please let me know if my analysis is way off base or that the tech information is wrong.

 

[Dave Baumann]

Josh, you are thinking about what is happenng on the master card, but the reality of the story lies in what can be transmitted by the slave, thats where the limitation comes in. Also, refresh rates / frames displayed are actually fairly separate thing to the actual image being displayed - the a refresh of 60Hz but rendering at 200FPS can still give a better gaming experience because the apparent latencies are reduced.

Also note, that none of internal display devices will produce the final displayed frame under 3D in a Crossfire system, this is the domain of the composite device, the TMDS or external RAMDAC. In this case note that displayed refresh rates / resolutions will not be the same as a standard Radeon as the external RAMDAC is 240MHz.

Finally, in all modes the slave cards are transmitting a full frame - in Scissor or Tiling mode the sections that weren't rendered are effectively just black.

 

[Xmas]

Josh,

That data is sent to the compositing engine, which then takes the final frame/s and sends it to either its built in TMDS transmitter or the RAMDAC, both of which can handle 2048x1536 @ 85 Hz.

The TMDS transmitter can't handle this resolution.

What we are basically talking about here is bandwidth between the slave card and the master card.

Unfortunately, not quite. The master card has no way of telling the slave card when transmission would be appropriate. The slave card just sends pixel for pixel at a fixed rate. Now, if the output refresh rate is supposed to be different from the input refresh rate, you have to buffer the incoming data. However, AFAICS the composition chip has no access to any significant amount of memory. So the output refresh rate simply has to be equal to the input refresh rate.

Now, in saying that, I think that ATI should have used the Sil 1171 receiver that can handle 2048x1536 @ 60 Hz, as it would have given them a LOT more headroom and would have made this controversy a non-issue.

A better receiver buys you nothing if you can't put better transmitters on the slave cards.

 

[Karma Police]

But you can put it on the R520's right?

Could that be why the Crossfire R520's don't work with the slave X800XT's?

 

[JoshMST]

Yes, fps and Hz are most definitely different, and I did my best to differentiate those two things. Don't get me wrong, they have limited their setup, but my point is that it is not the showstopper many are thinking.

I think one of the big problems here is that we don't know exactly how ATI is actually making CrossFire work. When transmitting, are those areas in Scissor or SuperTiling really just black pixels? I really don't see a reason why the driver does not make the slave card run at 1920x600 for Scissor mode (but I of course have no idea what SuperTiling would do).

And yes, the more I look into this, the more it becomes apparent that the TMDS transmitter on the X8x0 series cards is also holding it back, but then how do current X800 cards successfully power these 23" and 24" widescreen LCD's @ 1920x1200 @ 60 Hz? Reduced blanking is obviously part of it.