Slashdot Mirror
ATi's Multi-GPU CrossFire Graphics Card Unveiled
MojoDog writes "ATi has unveiled their new Multi-GPU technology dubbed "CrossFire" today out at the
Computex show in Taiwan.
HotHardware has a full preview of the technology, which requires both a
Radeon Xpress 200 CrossFire based motherboard and a CrossFire graphics card, in
addition to another Radeon X800 series PCI Express card, for dual 3D Graphics
processing with three available types of load balancing.
CrossFire supports Split-Screen, Alternate Frame Rendering and SuperTiling
mode load balancing between the GPUs."
http://www.anandtech.com/video/showdoc.aspx?i=2432
Just thought would be good to add variety.
...32 graphic chips!!!8 43.html
:-)
From TomsHardware http://www.tomshardware.com/hardnews/20050526_155
I will live on bread and water from now on to afford a system with this... in the far future!
HardOCP (http://www.hardocp.com/article.html?art=Nzc4) also has a decent preview. If you look down the list of the various news items for today, the [H] has included links to other previews. Also, they have some photographs from CompuTex (???) in Taipei from this week.
I skimmed both the Anandtech and HardOCP articles, and the basic gist about ATI's "SLI" is:
- needs an ATI chipset (the 200 -- for both Intel and AMD right now)
- "SLI" connector is external via some sort of weird DVI dongle
- uses one (1) existing X800 or X850 flavor card + a special CrossFire edition of same card models = no real need to get TWO CrossFire cards at one time if you already have the above models
Looks like I'm gonna need a monster case to ever be able to do this setup (ATI's demos at CompuTex take up 4 friggin' slots on the back of a case).
IronChefMorimoto
Wrong. Instead they stated that the 'optimum' platform is the Xpress 200 CrossFire.
However, between the marketing bullshit, you can clearly see that the motherboard is just a dupe of NForce4 SLI (and of similar Intel chipset coming up). Exact same PCIE setup. So it's almost certain that CrossFire will run just fine on nVidia chipset SLI motherboards.
I doubt they'd do a commercial suicide to prevent it on driver side. Today ATI has 0 SLI boards out. Nvidia has a gazillion - many of which are currently running X800/X850 cards. Nforce4 was first working PCIE AMD chipset, so many bought it - even the more expensive A8N-SLI or similar from other manufacturers, because nothing else was available at the time. Then they noticed how sucky the 6800GT/Ultra drivers currently are (stuutttteeerr bug in EQ2 comes to mind) and decided to fill the board with top of the line ATI card.
Such people are the PRIME candidates for forking out extra 500$+ for a CrossFire card, and I'm quite sure that they'll want the money from these people WITHOUT forcing upon them a crappy unproven ATI chipset based motherboard.
Now I do admit that ATI has been very elusive about this in their marketing material (ahem, I mean 'exclusive previews'), but if you go over them all, nowhere it says the thing *requires* ATI chipset, and I'm quite sure that detail is missing for a very good reason - they are late to the party on the motherboard side, and their system is exactly same (two x16 slots, running at x8 mode), that doing it any other way would be just silly.
By the way, when 2 cards are installed, the PCI express bus speed gets divided between the two.
BTW, PCI devices all share the same bus.
Same with ISA and VLB.
I guess that's why it's called a bus.
PCIe at least has the advantage of offering switching, meaning if you have multiple communications going on simultaneously, they're not necessarily all waiting on each other (though I don't think that's going to be very useful in a real world environment)
The only graphics interface format that ever dedicated bandwidth is AGP... but fun multi-GPU tricks aren't possible there.
Honestly, I don't think the bus speed is the bottleneck on PCIe... I think that a top of the line comp today has more graphics bus than it can use (which of course is exactly the condition you want)... I think the fundamental limiter is the CPU in a lot of titles... and the GPU in everything else.
I am disrespectful to dirt! Can you see that I am serious?!
ATI has been doing multi-GPU since the R300. Evans and Sutherland has been using multiple GPUs to provide rock-solid frame rates and 24x anti-aliasing for military simulators. ATI have been considering multi-GPU for longer than nVidia and I suspect their consumer solutions will be much more robust and thought out than SLI.
http://www.gamespot.com/news/2005/05/26/news_61265 52.html
GameSpot's quarterly report said PC sales were down, and that they only account for 4% of sales. You could argue that PC sales remained the same and console related sales skyrocketed, but this is the tail end of a console generation when people are usually saving up money for the next console.
Since GameSpot doesn't sell PCs that I know of and console hardware sales are around 20% from that same report, you can venture that roughly 70% of sales are from console games. That's a staggering number from the number one games only retailer.
But really, just look at your Walmart and Best Buy. Console games store space take up at least 2 times as much space as PC game titles. Look at the sheer number of console games too compared to PC games. Ever wonder why companies like Epic are moving to console games and supporting those platforms? It's not a big mystery why console games outsell PC games.
Not really. Making GPUs faster is relatively trivial - just add more pipelines. There is very little in the rendering process that can't be excessively parallelised. Vertex shaders will only get into diminishing returns once there is one vertex shader pipeline per vertex (well, per primitive). Pixel shader pipelines will only get into diminishing returns once there is one per pixel. Other components can easily be parallelised more (e.g. compositing) by splitting up the screen into smaller fragments (not quite a linear speed-up, because of overlaps, but a significant one).
The problem is fitting all of these pipelines into an IC that doesn't spontaneously ignite when you remove it from the liquid nitrogen tank. Multi-GPU solutions step around this problem by putting some of the pipelines in a different package, so the cooling required is spread between two physical packages.
CPUs are different. They are designed for performing inherently serial calculations (while GPUs are inherently parallel). This means that doubling the speed of a CPU becomes increasingly difficult every time it is done, until eventually it will be impossible[1]. Dual Cores are a way to side step this, by making CPUs more parallel.
[1] This has been Real Soon Now(tm) for about 20 year, but isn't here quite yet, although we are seeing the first serious hurdles.
I am TheRaven on Soylent News
The difference between the two is that the consol is locked down. This is the goal of the DMCA driven dreams of the software industry. You'll take what they give you and that's it.
Because they're not as hand-designed as processors are. For the most part GPUs are written in higher level ASIC tools [e.g. verilog/vhdl] which then filter through the tool chains [automatic...] with some human intervention along the way.
The benefit is that they can produce a graphics card in half the time [hint: the last significant AMD change in terms of logic was from the K6 to the K7...] as a typical CPU like an x86 but the downside is they're less efficient.
Processors are written with high level tools but there is much more human intervention before it goes to tapeout.
The software analogy would be comparing someone who goes from C to binary to someone who goes from C to assembly, tweaks it and then to binary.
Tom
Someday, I'll have a real sig.