Slashdot Mirror
AGP Texture Download Problem Revealed
EconolineCrush writes "The latest high-end graphics cards are capable of rendering games at 1600x1200 in 32-bit color at jaw-dropping frame rates, but that might be all they're good for. For all their gaming prowess, all of these cards have horrific AGP download speeds that realize only 1/100th of their theoretical peak. This article lays it all out, testing video cards from ATI, Matrox, and NVIDIA, and clearly illustrates just how bad the problem is. While these cards have no problems rendering images to your screen, you're out of luck if you want to capture those images with any kind of reasonable frame rate via the AGP bus."
I'd certainly expect the AGP bus to be used asymmetrically, how often do you want to do high speed data capture with a card that's primarily output?
The only situation I can see where you'd want more than PCI bandwidth returning would be for uncompressed HDTV capture, and there are better ways to do that (grab the raw broadcast stream for example)
-Yarn - Rio Karma: Excellent
Maybe you should have read the article? The point is that the slow transfer rate from the card TO the PC's RAM means that capturing video (or recording a gaming session for playback later) is severely hampered.
To be honest though, most people buy a GF4 to play games, not capture video.
Code, Hardware, stuff like that.
In any event, there's another issue he doesn't really touch upon; while he mentions that a single frame at 1600x1200@32bit colour is 7.5MB, he ignores the fact that a 30fps movie would require (30*7.5)=225MB per second uncompressed; you either have to have that much disk bandwidth or have enough CPU grunt to compress that on the fly. I guess a dedicated MPEG encoder card could help, but your average box is going to have trouble keeping up with on-screen gibs, rocket trails and blood splatters and encoding video.
In summary, who the fuck cares?
all of these cards have horrific AGP download speeds that realize only 1/100th of their theoretical peak...you're out of luck if you want to capture those images with any kind of reasonable frame rate via the AGP bus."
As the quoted article clearly indicates, the problem lies with the drivers and not with the cards, the latter which the original poster intimates.
And the underlying reason is immediately understandable: after years of AGP cards and years of noone really complaining raising this issue - (except, now, developers of video-editing software who could benefit) - it seems clear that there isn't much demand for this kind of performance. In the (near ?) future there might be, but why should these companies spend money working on driver performance in areas like this when really customers only care about how well Quake will run ?
When people are willing to pay for these features is when companies will pay to build the requisite drivers. And that is how it should be.
The article presents that once the images are rendered out to the display, they are simply discarded. Sure, for any sort of video capture or whatnot, that sucks. However, the article does not attempt to answer why video card manufacturers do this, or if there are any cards that do take advantage of the AGPx4 bandwidth. My guess is cost. If all AGP video cards provided video feedback into the bus, you're probably looking at a non-consumer level product. And you know what? All I do IS use my GeForce to play video games. If dumping the frames after they are rendered keep the cost of my card down, I'm probably happier for it. Quite simply: Does this matter for the average consumer?
If I'm reading this article right, they're claiming that it also hinders normal screen captures.
That would mean that software like VNC would have much higher performance, if the drivers were updated, the way these guys are demanding. (Wouldn't it?)
That'd be fantastic!
Education is the silver bullet.
This would affect everyone in a different way though. TV stations and production sets, even public access TV, along with low budget movies, would be able to use their PCs with a Radeon 9700 or NV30 card to produce their content. They could not only reproduce many of the effects from movies like Toy Story (notably excluding ray tracing), but do it in real-time for instant feedback, meaning much much faster production cycles. This has the potential to make a big impact.
I know nothing about anything, obviously, but I can see that game designers might think it nice to be able to send stuff to your screen but for you to be unable to send it to storage somewhere.
This *is meant to be* a dumb question. Mod me down if I'm wrong; it's only Karma.
If you're doing multi-pass rendering, it might be extremely convenient to capture the results back to main memory. Especially if the board doesn't have enough texture memory to support all of your temporary buffers.
And boards are starting to ship with 128-bit IEEE floating point buffers.
Essentially, you're right - a human can't tell the difference beyond 24-bit on a given image. But if 100 images were composited together (very likely, to support something like RenderMan-style rendering in hardware), 24 bits is nowhere near enough - you'd get all sorts of accumulation error.
Education is the silver bullet.
Way back when I was working on libfbx, we (the two main libfbx developers) learned of a 48-bit framebuffer developed by SGI. It's used mainly to render special FX for Hollywood. After several composited layers with various effects on an 8-bit per channel system, you can really start to notice the quantization artifacts. Moving to 12- or 16-bits per color channel (depending on whether there's an alpha channel) makes a huge improvement. I've never heard of any 16 byte per pixel (128bit) image format. It'd probably be something like 16-bits per channel RGBA (64), plus 32-bit depth buffer (96), plus 16-bit stencil and select(pick) buffers (128).
A solution to the problem with music today
I wouldn't use one of these cards to capture video though. I can't see why most people would, actually. The Matrox cards might be an exception. Quadro is a CAD/CAM card. These are just consumer grade cards. They buffer and write video directly to the hard disk. Real video editing hardware works differently, but even they often have several gigs of onboard RAM.
So really, I guess that I meant to say that I fail to see the relevance of the article. It is kinda of silly, actually, to even want to record real-time game footage with this hardware. Just pipe the video output to a real capture card on another machine. Problem solved.
A couple of salient points come to mind when reading this article:
1) Recording games/presentations/etc. The reason why we don't do it is because if the system was capable of generating it real time in the first place, it's far less space intensive to record the parameters of the animation than the output. i.e. It's cheaper to say "Daemia fires rocket at these coordinates" than record an MPEG of said rocket shot. AND, as hardware gets better, your recording does too.
Which leads me to point 2:
2) Since it's cheaper to capture realtime animation by capturing parameters, the only use of the capture function would be NON-realtime applications - i.e. getting your Geforce5TiUltraPro to render an extremely complex scene with incredible realism at 1 fps. That's not a typo. If we have 10MB/s back-into-the-PC bandwidth and each super high resolution shot takes 10MB on average, we have a wonderful solution working at 1 fps. Spend the fill rates on 600 passes for each pixel or something like that. Imagine the quality of the scenes! Capture the damn things and be glad you're not rendering at 1 frame per hour like they were 5 years ago.
Repeat after me - if you're rendering for posterity you don't need real time... That'll come eventually.
-JackAsh
...that I have ever read. Either that, or I am missing something here... The idea that graphics subsytems have 'bandwidth to burn' is kind of ironic, given that every graphics chip is ultimately held back in performance by the amount of bandwidth available to it - especially when using high quality options like anti-aliasing. The main focus of the article is actually a very niche segment... the idea of transeferring back rendered images over the AGP bus for TV / film / etc. is a joke... Rendering at high quality takes a huge amount of bandwidth (ie. textures and geometry)... as someone else pointed out, transferring back high-res images would take up over 200MB - that's a quarter of your AGP bandwidth! And without taking into account contention and timing issues in uploading/downloading that would mean that you simple couldn't realise the full potential of the bandwidth without a lot of other (expensive?) hardware... The simple fact is that for production uses, you would be *far* better off taking a stream of data from the DVI connector, and storing that for later use... Screen capture for business use is a reasonable point - however when does that require 3d rendering to be taking place? There should be no contention and no reason why the AGP bus couldn't be utilised fully - although would the graphics companies make enough out of this to justify the effort? As for internet streaming - how many people have access to bandwidth fast enough for high quality, full screen video streaming? Enough said...
Once, definitely. Twice, probably. Thrice, perhaps.
You typically composite and re-composite layer after layer to create decent effects, it's not a one-shot thing. Certainly professional video runs at ~48bit for film work.
Simon
Physicists get Hadrons!
Hmmm. Close but still not quite right. Think of the colour space as a cube with RGB as the three axis of the cube. In 32bit colour you have 8 bits per colour plane, giving you a cube that is 256 x 256 x 256. Any gradient from any point on the cube to any other point on the cube is going to be a maximum of 443 (if my maths is freaked out - distance from two opposite corners of the cube). Plus some messing about with the various quantisation that this line will pass through gives you definite banding on all but the lowest resolution displays...
The only Good System is a Sound System
Our ray intersection algorithm implemented on the GPU (an "old" Radeon 8500) was able to intersect 114M rays per second. This was loads faster than the best CPU implementation, which could handle between 20 and 40 intersections.
But when we tried to implement a ray tracer based on this, and an efficient one that didn't intersect every ray with every triangle, the readback rate killed us. Our execution times slowed down to the low end of the fastest CPU implementations.
And the readback delay seems to be completely due to the drivers, which apparently still use the old PCI-bus code. If the drivers could use the full potential of the AGP bus, our ray tracer could approach twice the speed of the best CPU ray tracers.
I can think of several reasons:
- The company hasn't released the game yet, but wants to release a video of gameplay to the public. Current methods would require implementing a "save game as it goes" and then a "replay, in offline rendering mode at a steady frame rate, and record results" pass. Or, you could save it at reduced quality if you had video out on your computer and video in on another computer.. but that's just ridiculous, imo.
- Likewise, you have the game, and your friend hasn't purchased it yet, and lives too far away to just take a glance at it..
- You're having a graphical glitch in a game with your particular card that can't be easily illustrated with screenshots. Think how much easier it would be to just send a video clip than having to send a half-dozen screenshots and a wordy explanation, where they still might not believe you.
- You have a Radeon9700, he has a Geforce2. You want to show him how different Doom III looks on your card, as opposed to his card, in real time.
Etc..
That's what render-to-texture is for, you don't need to read data back to the CPU.
b) split world/image-space occlusion culling.
This wouldn't be too useful for realtime graphics anyways, because of the way the 3D graphics pipeline works. The CPU can already be processing data a few frames ahead of what the GPU is currently working on. If you read back data from the card every frame, you have to wait for the GPU to finish rendering the current frame before you can start work on the next one.
Very few people use their typical desktop video cards for actual video production or anything related to it because the hardware up until now was simply unable to handle that sort of load. Now we have these cards that are the beginning of a new era of computer-generated visuals. The article is saying that they can do quite a bit more than they can do now if someone would just write some better drivers for them.
Now, streaming real-time rendering images over the internet? Maybe not fullscreen stuff right now because of a multitude of hampering factors on affordable internet bandwidth which I won't name for clarity's sake, but for the limiting factor to be the internet itself and not the graphics card is still a significant step.
This would definately be very beneficial to low-budget game developers and movie directors. We could very well see the return of the shareware boom (remember the early-mid 90's?) because of this.
sure, only a small portion of the people who'd buy the cards would use these features that the article talks about, but they'd be people that didn't have that capability before. Whenever this happens in any medium/artform/what-have-you, there is the tendency for a lot of experimental stuff to appear. I think we have some very interesting times ahead of us if someone gets these drivers written.
Just another freak in the freak kingdom.
"What kind of idiot puts their most powerful processor at the end of a one way street?"
Maybe they're the kind of idiots who know most people just want the best possible OUTPUT for gaming possible, and so don't want to add any overhead in card performance - or even additional design time - that isn't related to gaming performance. You know, the idiots who make cards that get award after award from gaming companies, then write near-perfect drivers, port those drivers to linux, and let you overclock the card to your heart's content. Those sort of idiots. My, they're idiotic.
Nobody says, "buy a geforce 4 ti, make the next toy story." No, it's advertised as a gaming card, and that's what its designed to do. If you want to do high-end video rendering things, perhaps a gaming card isn't the best choice.
I'm the stranger...posting to
First, Matrox and 3dLabs are both shipping products that do 10r-10g-10b-2alpha color.
Second, the poster wants to do more than "make due". You can also make due with 16 colors. And no, 256 levels is not enough, if you're compositing many images together, or if your data has a high dynamic range (which would require more gamma range than 256 levels are capable of providing, without serious banding.)
Third, pot. Kettle. Black.
Education is the silver bullet.
I've been doing real-time 3D graphics for 10 years and read-back speeds have been the biggest problem for doing many advanced algorithms. We have asked the companies to improve this many times. The problem as I see it: Quake and other benchmark apps don't rely on readback. ./ may even have run a link to one of these techniques a while back.
Here are a few other important but non-Quake techniques that are driven by readback speeds. I'll go into more detail on the first for illustration purposes.
High-quality real-time occlusion culling -- many techniques render the scene quickly by using a unique color tag per object or polygon and then read back the framebuffer to figure out everything that was visible (and how many pixels for each) for a final high-quality pass. If HW drivers would even just implement the standard glHistogram functions (which essentially compress the framebuffer before readback), this would become practical. NVidia adds their NVOcclusion extension, but it's limited in how many objects at a time you can test, it's very asynchronous, and it requires depth sorting on the CPU to make it most useful. The render-color technique does not. Yet HW makers are spending lots of money adding custom HW to do z-occlusion when a simple driver-based software technique may be easier.
Dynamic Reflection Maps -- for simple, reflective surfaces -- Requires background rendering from multiple POVs (generally six 90 degree views) and caching these. Even if you can cache a small set of maps in AGP memory, you want fast async readback if you have a large fairly static scene and you're roaming around.
Real-time radiosity -- similar to above, but needs more CPU processing of the returned images and possibly depth maps (reading back the depth buffer is often even more expensive than the color).
Real-time ray tracing -- the better quality approaches need fast readback to store intermediate results (due to recursion, etc..). With floating point framebuffers and good vertex/pixel shaders, ray-tracing becomes possible, but not yet practical. I believe
So there's a lot more to this issue than just making movies of your games. Faster, better graphics would be possible. So why isn't this a priority?
------------ cyranose@realityprime.com
OpenGL supports reading back the screen buffer mostly so that the OpenGL validation suite can check the rendering accuracy. For that, it doesn't have to be efficient. And if you read back in some format other than the actual structure of the framebuffer, every pixel gets converted in software and performance will be awful.
This article reads like it was written by an overclocker, not a graphics developer.
Any gradient from any point on the cube to any other point on the cube is going to be a maximum of 443 (if my maths is freaked out - distance from two opposite corners of the cube)
The distance between opposite corners is about 443, but the diagonal distance between color points is 1.732, so you still have 256 points in the gradient.
Think about it this way, the gradient from (0,0,0) to (255,255,255) passes through (1,1,1), (2,2,2), etc. Exactly 256 points.
-
- - You can't take something off the Internet! That's like trying to take pee out of a swimming pool.
If you read the AGP spec, which was written by Intel, you will note that it is based on the PCI 2.0 spec. The PCI 2.0 spec is for a 32 bit, 33 MHz symmetric bus which gives you a max transfer of rate of 132 MB per second. The AGP spec is for an asymmetric bus, 33 MHz read and 66+ MHz write. But writes were optimized at the expense of reads, since Intel was pushing video with NO onboard texture memory, and who would want to read back the image in real-time anyway, right?!?
Yes, I am sure that drivers do have some affect, but the AGP spec is the first bottleneck. On an OpenGL news group it was reported last year that a person tested two identical video cards, the only difference being one was AGP and the other was PCI. The read performance for the PCI version was several times faster than the AGP version.
Of course, some video cards are also to blame because of the frame buffer format they use, but that is another story...