Slashdot Mirror
Yet Another Serial Graphics Bus From Intel
ottotto writes: "Techweb has a story about Intel's High Speed Graphics Initiative. After discussing another doubling of the AGP, VP Pat Gelsinger said "The next part of that road map is AGP8x, an evolutionary step from AGP4x, to be followed by a future serial graphics bus." ANOTHER serial graphics bus? Is not the upgrade path to IEEE 1394B (800 Mbps Fire Wire) and beyond sufficent? Is this, along with the USB 2.0 spec another
way around giving any credit or royalties to Apple?" I suppose companies have to make their plans somehow, and new products are better than living in the 1960s forever. But sometimes these "roadmaps" (which often turn out to be more like directions scribbled on the backs of napkins) seem to smack of planned obsolesence. Do you ever skip the current latest/greatest because you know what's around the corner?
The hardware you buy will in fact become obsolete before you can "give" it to your kids. Sorry to dissapoint you after that peptalk you from the local big box salesman about "that top end computer will last you for 5 years!"
1) NVIDIA GeForce2 drivers are rock solid
2) They support Linux.
3) Yea, but todays mediocre is tomarrows Mac hardware.
4) If you can live with a $400 system, then more power to you. Some people actually NEED the extra power, and will pay for it. Question: Would you buy a Porsche?
5) It usually does. Either way, you get bragging rights.
6) I'm a consumer swayed by 40fps in 16K/12K Quake III.
7) To fill my needs, my system needs a lot of power. Live with it.
PS> Before y'all get all hot and bothered about that Mac comment, consider it. Right now, Macs are up to AGP2x and Radeon cards while PCs are at 4x GeForce2 Ultra cards.
A deep unwavering belief is a sure sign you're missing something...
Flipping through a computer catalog today, I saw that there were 4 pin and 6 pin FireWire cables. How on earth, after twenty years of personal computer incompatibilities, have we developed a ultrafast serial transport, and yet still have incompatible cables?
Just about everyone would be a "Microsoft" if they could: Intel intentionally crippling Celeron 2's so not to offer a reasonable price point vis a vis the Pentium 3; Rambus; Pentium Pro's could go 4 or 8 way, with P2 or P3, you can only go 2 way, need to pay the Xeon tax to go more the 2 way SMP. IBM used to be the "Microsoft". Sun's silly Java tricks (standardize, or not.... let us get back to you), etc.
So, this is mildly on topic, but it just shows how we shouldn't be surprised by stupid market leader tricks.
matt
Actually broadcast quality fullscreen fullmotion video is pretty low-res by PC standards. A 1600x1200 32bpp 60 fps video chews up around 460MB of bandwidth, significantly more than Firewire's paltry 50MB/sec. Although, I have now idea why the hell you'd pump a game through firewire, but higher quality video IS beyond firewire. (In fact, it would seem that Firewire can only barely keep up with a 30fps HDTV 1080i clip.)
A deep unwavering belief is a sure sign you're missing something...
The new "Serial Graphics Bus" referred to in this article may actually be a follow-on to the serial digital standards being pushed today for flat panels. Digial Video Interface, Plug n Display, etc, are all in various states of acceptance, and then there is always the fact that Silicon Image owns the patent for PanelLink, which is the link layer protocol that runs on the DVI/PnD connection (Intel is intimately involved in this effort, but does not own anything outright)
Also interesting is the fact that the MPAA is emphatic about an encrypted link from the source (DVD, for example), right to the display... they want to disable any possibility of copying pristine digital content-- you may have heard about this elsewhere... When I worked on flat panel stuff (at Philips, go figure), intel was definitely getting behind such an encrypted link (which would be serial).
Supporting larger, higher-density digital displays with a digital input stream will require better and better connections as well, and if a P4 is needed for better peer-to-peer networking, than certainly intel will find themselves getting involved in some "critical" way here as well.
But all of this is just a guess
"This is the reason that Rambus (16-bit 400 MHz) is so hard to manufacture compared to SDRAM. Moving to a serial bus allows the clock speed to be cranked much, much, MUCH higher without worrying as much about data errors."
Either you just accidently contradicted yourself, you forgot to add a final sentence, or you're confused about RDRAM and SDRAM.
SDRAM is parallel. The Rambus is hard to manufacture because it has to be high clock speeds.. very high clock speeds, to beat out the parallel technology. That's because a normal PC100 DIMM doesn't have to be very fast (thus avoiding the more severe effects of cross talk) because it can push 64bits per clock cycle. It can acheive a peak bandwidth of 800Mbps. This is half of what an 800Mhz RDRAM RIMM can do. But the "simpler" RDRAM RIMM is much, much more expensive. Adding extra datapins at slow speeds isn't too hard compared to cranking even the smallest number of pins to incredibly high speeds.
Based on your assertation that crosstalk is bad thing, one would think that RDRAM would be easier to manufacture as they use the simple serial process. This isn't completely true, as you've oversimplified the case and sound like you contracticted yourself ("RDRAM is hard to mfr.. Serial RDRAM is easy to mfr").
---
--
Internet Explorer (n): Another bug -- that is, a feature that can't be turned off -- in Windows.
There are really two questions people seemed to be confused over:
There is only one reason for more bandwidth: 3D graphics. AGP was originally designed so that graphics cards could use main memory for textures. Alas, it was too slow, and graphics card vendors began to just pile up RAM on the video card. However, most people don't realize that textures aren't the only thing sent down bus. Polygon data needs to be sent down every frame as well. Enter hardware transform and lighting. Suddenly games are starting to be designed with a LOT more polygons. Due to things like animation and dynamic level of detail, this polygon data can't just be stored on the graphics card like textures are, they have to be downloaded every frame. If you extrapolate the curve of supported polygons / second by the history of Nvidia cards, you'll see that AGP4X and AGP8X will be saturated relatively soon. All graphics card vendors (except maybe 3dfx, we'll see what they say after the Rampage ships) are clamoring for more bandwidth.
The reason for going to serial is physics oriented. Parallel wires switching at high speeds can generate electrical fields that cause signals in nearby wires to change. This is known as crosstalk (or noise). This is the reason that Rambus (16-bit 400 MHz) is so hard to manufacture compared to SDRAM. Moving to a serial bus allows the clock speed to be cranked much, much, MUCH higher without worrying as much about data errors.
Dan
A lot of people misunderstand and underestimate the benefits of AGP. It is not particularly great at making video faster; this is because very few graphics cards actually bother to use the primary function of AGP, texture caching in main memory.
What AGP is very good at is improving total system performance by taking graphics data off the main PCI bus.
The PCI bus is a non-switched fabric bus. All PCI devices have to share the bandwith of the bus and a graphics card can consume almost the entire bus bandwidth. By moving graphics data to the PCI bus, you can significantly improve total system performance. In that sense, it roughly is the equivalent of segmenting a network.
A better fix for this problem would be a switched-fabric main bus, like the kind that high-end workstations use. Unfortunately, that solution is very expensive and AGP is relatively cheap.
Actually, I think the opposite might happen - eventually, all graphics interfaces will use shared memory with the CPU. Neither serial or parallel interfaces will be fast enough in the long run.
There's a big problem with using a purely serial interface to the video card. For really good 3D graphics, the serial interface would need to be horrifically fast, or you would need an awful lot of RAM on the video card.
Most games will run in 16 MB or less of video RAM for textures, but this is using fairly low-res / small textures. I play a lot of 3D games, and they all have the same visual problems: lighting and shading anomalies, low resolution textures tiled over surfaces, and low polygon counts. These are some of the main problems preventing current 3D games from looking "real". John Carmack and the other gods of 3D programming do a lot of work to get around these problems and disguise them, but it would be better to just fix them.
The use of really high resolution textures would make games look dramatically better, but is difficult or impossible with current video cards. A single high resolution, 24-bit texture, covering a wall or floor for instance, might easily be over 1 MB. And a typical 3D game needs dozens of textures all at once.
For example, assume that you are running a game at 60 frames per second. Suppose your screen resolution is 1280 x 1024. Then for really realistic graphics in a big outdoor area you might want to have about 32 high-res (1024x1024), 24-bit static textures. That's 96 MB right there. Add in some mip-mapped downsampled versions and it's at least 128 MB.
Suppose you've "only" got 64MB of RAM on your video card. For a really rough approximation, you will have to transfer about 64MB of textures through your bus, 60 times a second. Do the math - it's 30 gigabits per second. I'm no hardware engineer, but I don't think it's that easy to even make silicon that switches at 30 gigahertz, and that's what you would need for a serial bus.
Even AGP 4x only manages 8 gigabits per second, and it's a parallel bus.
Even if you have enough texture RAM on the video card, many games use procedurally generated textures, usually for things like water, fire, and other effects. These change every frame, and look dramatically better than a simple repeating loop. Unfortunately, procedurally generated textures must be uploaded to the video card every frame. Even one high-res procedural texture could suck up 500 Mbps of bandwidth.
Hardware supported texture compression helps a lot, but can't completely solve the problem.
Really, I think the best thing for high speed graphics would be for the video card and the CPU to just share a big whack of high-speed DDR-DRAM. Interestingly enough, this is the approach that Microsoft's X-Box is taking.
Using shared memory between the CPU and the video card would also make it much easier to experiment with more esoteric forms of 3D graphics generation, like hardware support for voxels.
That could lead to some gaming breakthroughs. I'm getting tired of 3D games with worlds built out of perfectly flat triangles and rectangles with blurry textures plastered across them.
Torrey Hoffman (Azog)
Torrey Hoffman (Azog)
"HTML needs a rant tag" - Alan Cox
Well, you know what?
Most games don't even take full advantage of AGP 2x! Because video RAM is so cheap nowadays, with 32 meg cards being common, the AGP bus is going to waste most of the time now. There are benchmarks showing very insignifcant gains going from AGP 2x to AGP 4x... and they're talking about AGP 8x now? Big waste of money.
And yes, serial lines *ARE* much more able to made faster than parallel. Parallel links suffer from one problem - the need for every dataline to be locked 'solid' before sending the next logic level. A slight skew in arrival times can corrupt data. The limit is just how powerful your parallel line drivers are to quickly propagate all the lines. (This also explains all the zig-zags you see on motherboards and RAM blocks - they act as delay lines so the signal doesn't arrive too early before the other signals).
I'm becoming digusted with slashdot's journalistic ability. For all practical purposes, the link contains no information what so ever about a "Serial Graphics Bus" except that there might possibly be another one (and couldn't you assume as much?), and yet it's important enough for the title of the post?
We can't even know what they mean a "Serial Graphics Bus", but I would bet its not a replacement for Firewire or USB. Please save the mindless speculation for the comment area
If slashdot wants to be a rumor site, how about you post some real rumors?
Actually Apple shares all the royalties from IEEE1394 licensees with a consortium of other manufacturers including Sony (who brand it as iLink), Canon, etc. read all about it: http://www.1394la.com/lic_agreement.html
lisa bonet ate no basil
The problem with these crazy hardware plans is that they're adding additional cost and waste for people who don't need it. AGP was nice...for gamers. Didn't make a bit of difference for anyone else. And in all honesty it didn't do anything for gamers either. All that talk about using main RAM for textures went away, and video card makers just starting putting more and more VRAM on their cards. Everyone lost. Kinda like MMX.
Intel needs to stop taking niche products and pushing them to everyone as a necessity. How about just focusing on decent processors, okay?
Well the only reason graphics cards originally mapped thier frame buffer onto the computers memory space was because as you know the graphics hardware relied on the CPU to do the grunt work.
Your 2D graphics card (lets say a bog standard Matrox Millenium) takes a list of commands (i.e, draw a few lines, do a few bitblts) and processes them indepenentantly of the CPU. Infact you can pretty much bet that your video driver hardly reads/write directly to the frame buffer. This is pretty much extended to 3D except they take it a step further with geometry engines and so forth.
As things have got to this stage now it pretty much makes sense to just have a really fast serial link to your graphics card which would use up less tracks on the mother board.
I personally like the way things have turned full circle. X is designed from a serial point of view and all the nay-sayers and critiques should take a step back and see why it is such a good design (after you take away some of the bloated "extensions").