The New Face of CrossFire

There haven't been any changes to the way CrossFire works from an internal technical standpoint, but a handful of changes have totally revolutionized the way end users see CrossFire. NVIDIA's SLI approach has always been fundamentally better from an end user standpoint. Internal connectors are cleaner and easier to use than ATI's external dongle, and the ability to use any X1950 Pro in combination with any other X1950 Pro is absolutely more desirable than the dedicated master card approach. ATI has finally done it right and followed in NVIDIA's footsteps.

At the heart of the changes to CrossFire is the movement of ATI's compositing engine from the card onto the GPU itself. This does add cost to every GPU and thus every graphics card, but the added benefits far out weigh any negatives. In early versions of CrossFire, digital pixel information was sent between cards using TMDS transmitters (the same transmitters used to send display information over DVI and HDMI). While this format is fine for displays, it isn't as well suited for chip to chip communication.

With the compositing engine built into every GPU, ATI is now able to send pixel data through an over-the-top NVIDIA style bridge directly to another GPU. This also eliminates the necessity of a TMDS link for use in transmitting pixel data. ATI hasn't talked about what type of communication protocol is used between the compositing engines on each chip, but we suspect that it is a little lower speed than NVIDIA's 1GHz connection. ATI is using a higher bit-width connection split into 2 12-bit parallel channels. At full capacity, ATI states that these connections can support resolutions of up to 2560x2048, but that communication doesn't happen any faster than the old style TMDS method.

ATI did make it clear that even though this incarnation of CrossFire supports a higher resolution than we are currently able to test, it won't necessarily run well. Of course, we'd much rather see a situation where we aren't limited by some technical aspect of the hardware. The first incarnation of CrossFire was quite disappointing due to its low maximum resolution of 1600x1200.

One of the oddities of this multi-GPU implementation is the splitting up of the connector that links the GPUs. Both are required for the driver to enable CrossFire, but only one is technically necessary. As bridges will be bundled with graphics cards, everyone who purchases 2 X1950 Pro cards will have two bridges. This eliminates the need for end users to buy bridges separately or rely on them shipping with their multi-GPU motherboard. When pressed further about why two connectors were used, ATI asked us to envision a system with 3 or 4 graphics cards installed. With 2 channels, cards can be easily chained together. This does offer ATI a little more flexibility than NVIDIA in scaling multi-GPU configurations, but it is also a little more cumbersome and offers more small parts to lose. Overall, though, the 2 channel configuration is a good thing.

Now that we have a chip built specifically for the $200 price point with a robust, full featured, CrossFire implementation, we are very interested in seeing what type of performance ATI is offering.

RV570 and the Demise of the X1900 GT The Test


View All Comments

  • Spoelie - Tuesday, October 17, 2006 - link

    They're total system power, we're not gonna see 180W gpu's till november :) Reply
  • Aikouka - Tuesday, October 17, 2006 - link

    Those cards are going to be so hot.

    Sorry, that was my one and only bad pun of the day :P.
  • Aikouka - Tuesday, October 17, 2006 - link

    First off, it was an intersting article and it's nice to see that ATi is making changes to "fix" the poor dual-card implementation :).

    But now akin to my subject line, I was wondering if you [Anandtech] would be keen on adding in a "Performance / Price" sort of chart at the end. The idea would be to keep different quality modes separate (i.e. No-AA/No-AF and #x-AA/#x-AF would need separate charts or only one chart but using one set of data) while ordering cards by their average FPS or "average ranking". Such as, if Video Card A is seen in spot #1 the most, it stays there in the chart regardless of the figure listed. I'd say this is more of a user-friendly idea than anything required to be practical. But here's what I think a typical bar would represent:

    [ Video_Card_Name - (Average_FPS/Price) - Average_FPS ]

    The idea is that some video cards may be return the same value (i.e. 120/$400 and 60/$200), but listing cards in their typical performance standing allows someone to say "well, hey... I like the performance ratio of this card, and I don't need 120 FPS!"

    I think it may provide a way for people who read the benchmarks to get a real world idea of these cards rather than an "in box" idea, because as nice as it is to see a card produce 120 FPS... how much will we have to produce to purchase it ;).

    Just something I thought up while looking at pages and pages of charts and not knowing how "worthwhile" the cards really were for the performance.

    P.S. I'd love to hear comments!
  • Spoelie - Tuesday, October 17, 2006 - link

    I don't think it'll be very practical for most people. When a graphic card purchase is being considered, most people just have a hard cash target in mind (for example, 200$) and just look for the fastest card at that price point. The only exception to this is when those people notice that a small rise in their budget allows a much more powerful graphic card purchase, e.g. 6600gt for 100$ but 7600gt for 130$

    So as long as anandtech keeps comparing on pricepoints and of course mentioning any caveats/featuredifferences/possible better deals higher up, there's no real need for such a chart.

    but that's just my opinion
  • Aikouka - Tuesday, October 17, 2006 - link

    That's really a view of someone who's looking to spend x amount of dollars trying to find the best item for that price, which is one thing that's kept in focus. My proposed chat method helps to show two things really, what's the best "bang for your buck" and allows you to see if that performance pricing is the level of performance that you want.

    To go into detail a bit about that, if the best performance ratio was .3 (120fps/$400 or 60fps/$200, etc), someone might go, "Well, hey... I like the performance rating on these two cards, but I don't need something that fast." So, the user would shoot for the 60fps card.

    But in your case, you most likely benefit from what's currently done. For me, I could really care less how much a new video card costs me. I got tired of spending smaller amounts on mediocre performance years ago and I've tried to keep my dollars spent in the high-end sector (the range of high-end cards, not necessarily referring to the ultra-high-end-uber cards only) as much as possible to avoid having to perform constant upgrades to keep that level of detail that I like (I can't stand "jaggies" ... I'm just too anal about things like that).

    For example, my current 6800GT was the 2nd top performing graphics card when I bought it, but it simply doesn't suit my needs anymore as below 30FPS (in more render-heavy areas) in WoW at 1280x1024 with high graphics settings is just subpar. Now, my card is old and about to be replaced by another generation (the G80). Or maybe I just need one of them Killer NICs (just kidding :P).

    I think a good thing to mention is that the bar system that I mentioned may be better for a bulk review more than a single card review. Especially since a lot of people may come into the review of a single card knowing all the information about the other cards, so they're only performing minor mental comparisons. But in a bulk review, there's a lot more information to keep in check, so an overall comparison that takes a couple of the specifics about a card ( average performance and price mainly ) could be beneficial.

    Thanks for the comments!
  • Aikouka - Tuesday, October 17, 2006 - link

    "The idea is that some video cards may be return the same value"

    should be...

    "The idea is that some video cards may return the same value"

    Stupid me deleting and rewriting so much that I'll leave in a word from a prior revision XD. Oh yeah, and I had to make sure to post this fix before someone sought it upon themselves to ignore the entire post, but to make a comment on my little accident :P.
  • bupkus - Tuesday, October 17, 2006 - link

    With other CrossFire configurations as a guide, we can easily expect X1950 Pro to nearly double its single card performance and put it on par with the 7950 GX2 and 7900 GT SLI configurations. As for single card performance, we see the trend of X1950 Pro domination continuing. Performance greater than that of the 7900 GS and GT for $200 is quite a plus. Reply
  • takumsawsherman - Tuesday, October 17, 2006 - link

    Now all ATI has to do is get rid of that Catalyst Control Center. I have been using a Radeon 9700 Pro for two years now, and I have been almost perfectly happy with it except for two problems: It sometimes loses TV mode with no explanation, and I have to remove and re-add the TV in CCC. The other is that CCC takes anywhere from 10-20 seconds to load up on an Athlon XP 2500+. The old control panel set opened within 2 or three seconds (the advanced screens). I've set up a quick profile, but even right-clicking the taskbar icon has a huge delay (often ten seconds or more) before the menu appears, and then sometimes activating the profile doesn't work, and I have to set it manually (sometimes it does work). After initial launch, it does open more quickly (cached, I assume). I assume that this is because of the use of .NET.

    Also, ZoneAlarm reports that CLI.EXE (listed as various ATI apps) is listening to TCP ports 1052, 1057, and 1058. There are currently three CLI.exe processes running on my system, taking 6,132K, 5,800K, and 3,572K of memory respectively. That is about 15MB of memory, and I don't have CCC open. My old Matrox G400Max had tons of options in the advanced screens, had a better TV output quality, and didn't require all of this garbage. The old ATI control panel even seemed easier to use.

    I would love to get a nice new x1950, as Oblivion gets pretty choppy even at fairly low settings. But I really want to get away from these bulky system tray apps. And look at the update process here:">;task=...

    That is ridiculous. Can they not make an installer that removes their own cruft before installing the new version? It's a chore to put a new driver in, and it shouldn't be.
  • Zoomer - Thursday, October 19, 2006 - link

    I think this is partly microsoft's fault. .Net framework is a POS.

    ATi should shoulder some of the blame too, for choosing a crappy base to program on. I would think even java based apps are not this bad.
  • mesyn191 - Friday, October 20, 2006 - link

    I dunno...

    I personally don't like the whole CCC/CP approach to config. the graphics card either, but nV's implementation aint' half bad.

Log in

Don't have an account? Sign up now