Slashdot Mirror
ATI Introduces a Parallel Processing Video Card
bilestoad2 writes " ATI has announced the introduction of a new video card, The Rage Fury Maxx. It uses two RAGE 128 Pro Processors and 64 Megs of Ram.
Follow this link for the complete story.
I don't know about you, but I've got to have one of these..."
Splitting full frame rendering among multiple processors...this is patent pending?
Are you kidding?
It's at least arguably unique to split even and odd lines among two cards(like Voodoo 2 SLI), or to split the image into horizontal strips(Metabyte's PGC), or to evenly split the texel reprocessing load among multiple texel processors(Voodoo 2 core design), but to attempt to patent the process of merely having one complete frame go to one processor while having the next complete frame go to the next processor?
The general reason one doesn't want to use a full frame architecture is simple: Per frame times don't budge. Either you have to build a higher latency into your rendering chain, since the chipcluster has to know the next x frames you intend to render, or you get *no* speed boost.
Don't even get me started on out of order frame rendering on a realtime rendering solution.
Each of the previously mentioned solutions(SLI/PGC/Texel x 2), incidentally, lowers per-frame latency.
Granted, there's probably some degree of multi-frame latency built into most drivers, particularly for games. But the concept of patenting the most basic parallelization solution strikes me as absolutely hilarious. It's very likely most 3D rendered movies use the technique ATI is trying to patent. "I'm done finishing this frame, send me a new one."
It's very likely most WORKPLACES work the same way too. "I'm done with this job, assign me a new one."
That being said, I'm looking forward to trying out ATI's new cards. Ever since I noticed their 128's were supported by Metabyte's excellent Eyescream system, I've been much more interested in them.
Yours Truly,
Dan Kaminsky
DoxPara Research
http://www.doxpara.com
summary:
- nothing new under the sun.
- ATI reinvents pipelining, ignores drawbacks.
Firstly a rant about the press release and its quoted 5% drop between 16 bit frame rates and 32 bit frame rates for this new ATI card: Any manufacturer could do this by artificially limiting their 16 bit fill rate. This number says nothing unless combined with an absolute fill rates at either bit depth.
Now to the deja vu: ATI has effectively shoehorned two cards' worth of acceleration into one graphics subsystem. This has been done twice before in the consumer space: first by 3dfx, with Scan Line Interleave, which allowed two cards to work in parallel on any polygon that spans more than one line on screen; more recently by Metabyte, with their Parallel Graphics Configuration, which partitioned the screen vertically into two independent regions and dedicated a card to handling each of the regions.
Both 3dfx and Metabyte use spatial partitioning to get parallelism. 3dfx could do it finer grained because they had control over the chipset design and could include a mechanism for tight synchronisation of two cards. Metabyte went coarse-grained because they had to do the picture recombination from the two cards in external hardware, and it was hard enough to make this work at all without making it work for alternating scanlines. So why didn't Metabyte save themselves a bunch of hassle and use the "temporal" partitioning (or, in other words, pipelining) approach that ATI is now using? Hmmmm...
One issue here is latency. (For this discussion, let's assume that the video refresh rate is arbitrarily high, so that as soon as a frame becomes ready, we get to see it.) When a 3d card completes the rendering of a frame and swaps the front and back draw buffers, you are seeing the state of the world as it was at the time the game engine _began_ to draw the frame. If the current frame render time is x milliseconds, that's x milliseconds latency between the game state and your eyeballs.
With a spatial partitioning like SLI, both chipsets work in parallel to render a particular frame, and each frame is completed before rendering of the next frame begins; the game state to eyeball latency is simply 1/(frame rate).
With the ATI approach however, each of the two Rage chips plugs away at its frame independent of the other (which is working on a frame either one ahead or one behind.) Frame _render_ time is therefore twice the frame _display_ time, and the latency is twice as high as SLI for a given overall frame rate: 2/(frame rate). For a 60Hz frame rate, SLI gives 16.6 ms game state to eyeball latency, while the ATI approach gives 33.3 ms.
I am not a cognitive psychologist, so I don't know if an extra 16.6ms or so is going to make a noticeable difference to most people, but I wouldn't be suprised if experienced first-person-shooter players noticed a difference. Certainly for modem play the extra latency is probably smaller than the variation in ping time to the server, so I wouldn't expect it to make much difference, but on a LAN it might be noticeable. I have turned off sync to vertical refresh and forgone triple buffering in LAN Q3Test games because the variation in latency between frames was driving me batty, so I think this could actually be an issue. Of course, the higher the frame rate, the smaller the extra latency, and the less this will matter.
There is also the other matter that for this to work, there has to be at all times a large amount of rendering waiting to go so that each chipset stays busy. The drivers will presumably have to do a *lot* of buffering and then spoon feed each chip as its command FIFO is exhausted. I really wonder whether this will fit in well with what currently written applications are expecting from 3d acceleration hardware; if an application wants to have any synchronous interaction at all with the hardware, such as reading back values from a stencil buffer each frame after drawing is complete, it will totally screw this kind of pipelining. Somehow I'm just not convinced.
-Snorbert, somewhere in the antipodes
Tom's Hardware has a full preview, although he's not allowed to print the performance results. This isn't SLI or PGC -- ATI is actually having the CPU's draw full alternate frames, so the image quality will be high yet the speed will be doubled. It's even buffered so if CPU0 is taking a long time drawing frame 0, CPU1 can keep drawing 1,2,3,... until CPU0 is done. What I like best about this is that the MAXX architecture will allow them to drop their latest chips in as they are developed, so even if their CPU architectures remain a little behind, they'll be able to keep competitive. I like ATI cards because of all of the MPEG and TV toys they build into them; the only other company that even comes close to offering those kind of toys is Matrox, but they're just too damn expensive for the full-featured cards.
SharkyExtreme has a more lengthy writeup, including some initial performance comparisons from a prerelease version (chips clocked to 125Mhz instead of 143Mhz, beta drivers).
Does anyone here remember when www.ati.com was run by some company calling itself "Artificial Turd Industries"? The home page featured a very large, very detailed image of rubber doggie doo.
/. and I'm sure the fall of such a rebel would have been noticed...? Does anyone know if the guy finally sold out or if some how the courts decided that ATI should get the site (even though there are many other companies that have the same initials trademarked!)
= -=-=-=-=-=-=-=-=-=-
The thing that was neat was that this page stayed that way for as long as I could remember. The owner took great delight in posting letters from lawyers demanding he turn over the domain name. Companies like ATI Technologies (the graphics card maker most people are trying to find when they type in www.ati.com), American Tractor Incorporated, Arand Typeset and Ink...and about a dozen others.
ATI ended up getting www.atitech.com which they still own. But now I just found that they also have acquired www.ati.com!
How did this happen? I don't remember reading about it on
- JoeCurious
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
-- I wonder which will go down in history as the bigger failure: the War on Drugs or the War on Filesharing