Free the video card!

[dakels]

You know with video cards getting so insane, I dont understand why they dont they just strap a real daughter card on there with a full size processor setup? Why can't you have dual opterons processing the video? What is so different about a GPU over a CPU?
IIRC in the old days there were some macs with some insane daughter card add ons that used the PCI bridge and take up 2-3 slots. Even stuff that was so big it was housed externally yet connected to the PCI via a connection card or SCSI.

Wouldnt producing some large scale video processing set up be cheaper but provide alot more power due to lack of size restrictions/miniaturization ? Like an external box which was basically just a huge video processing system/card, with a line that went to plug into the AGP slot. Doesnt have to be huge, but say the size of a external HD, even a little bigger.

I understand that this may not be viewed as a profitable marketing idea but I was just wondering. Honestly though, if by having more freedom of size would allow you to get 2x the processing power of a high end video card for the same price, I think there would be a market for it anyways. I know I and many other video power nuts would go for it. I doubt I am alone in saying if I could buy an external box that would have twice the processing power of the newest Gefore 6800 Ultra, and cost the same, maybe even a little more, I would buy it.

There are already external video importing/processing boxes, I dont see why an external high powered video card would be that far off. Forgive me if this is already out there but even if it is, its not common knowledge, and it's probably for an extremely specified purpose.

Forgive my lack of super technical knowledge...

 

[Badut78]

I guess you're looking forward to SLI then?

Check out this link for more info:
http://www.nvidia.com/page/sli.html

 

[dakels]

yes but its not really what I mean by my post.
Besides I think SLI has been out for a while with companies like the old 3dfx and high end workstation systems.

I dont want linked cards, I want an inexpensive super powerhouse video rendering system for cheap... cheap because they dont care how big the video processing set up is. It could be an external box the size of a stereo reciever for all I care. I think if it produced superior performance numbers at a relatively affordable cost, it would sell. Even like a small box which had 2-4 linked R9700's efficiently multiprocesing would smoke the competition as far as performance, and the cost really couldnt scare someone who spent $600-700 for a GFX 6800 Ultra. But what about 4 athlon or pentium processors on one board dedicated to video rendering? What is so different about GPU's over CPU's? I assume more specified of course for math.

 

[AndyHui]

Presumably, the bus that connects the external box will be slow. You pretty much have to have very high speed devices very close to the processor and RAM in order to get that high performance.

 

[Cattlegod]

Originally posted by: AndyHui
Presumably, the bus that connects the external box will be slow. You pretty much have to have very high speed devices very close to the processor and RAM in order to get that high performance.

Yep, the longer wires are the more capacitance they have, turns square signals into a flat sine wave. It is likely at a high frequency, the *box* wouldn't even see a readable clock signal.

 

[AnnoyedGrunt]

Also, from a cost perspective, the more parts you have the more expensive your product will be. Therefore, making a separate box (even out simple components) will almost always be more expensive than an internal solution. The cost of video cards isn't due to any size restrictions, so an external version won't really save any money by allowing more room (unfortunately).

-D'oh!

 

[Shalmanese]

Difference between GPU's and CPU's: A GPU's task is to move huge amounts of data in and out and to do fairly basic processing on it. You shovel data in one end and then you spit it out and forget about it. Thus, you need large memory busses, relatively low clock speeds and very little cache. A CPU is tasked for various things but is often called on to do fairly intensive processing on a relatively small dataset. Thus, you care more about caching and branch prediction and higher clock speeds. Also, it's only been the last 2 generations of GPU's that can be called truely programmable. DX7 era cards were completely fixed function and DX8 only had fixed shader models. Even now, very little branching or conditionals are done by a GPU.

 

[kautilya]

well its not really free is it?

 

[harrkev]

You need to keep in mind the difference between a CPU and a GPU.

The typical CPU is very inefficient. A large percentage of its resources are spent doing instruction fetch and decode. A CPU has to be general purpose. This means that if you give a GPU some vertex data, it knows what to do with it. If you give vertex datat to a CPU, it has to as "What next?" after each and every step.

In calculating the position of one thing in relation to another thing, you have to do a 4x4 matrix multiplication. Here is how a CPU and GPU might handle this.

--GPU--
1) Load matrix A from memory.
2) Load matrix B from memory.
3) Multiply
4) Use the result.

--CPU--
1) Load instruction to load first element of A
2) Load first element of A
3) Load instructino to load first element of B
4) Load first element of B
5) Load instruction to do a scalar multiplication.
6) Multiply.
7) Store result
8) Load instruction to load second element of A.
9) Load second element of A
10) etc. etc. etc.

The data flow of a GPU is optimized to do the sort of thing that a GPU has to do. A GPU in certain aspects probably approaces a systolic array.

On the other hand, a GPU would make a terrible general-purpose processor. So you trade off efficiency for flexibility. Think fork vs. spork.

 

[Titan]

not to change the subject, but with current trend of CPU integration (mem controllers, dual cores, etc) Does anyone think we will see a built in GPU as part of the CPU in say 10 years, maybe with it's own mem controller? That would greatly improve graphics processing, would it not? (of course the chip soze goes way up)

 

[dakels]

OK I understand the specialization of the GPU.

I dont see the AGP bus as a limiting factor. Not for transmission rate. Isnt an 8X AGP capable of like 2.04gb/sec? Latency I could see as being a huge issue with an external AGP device.
From what I am told the bus is not the limiting factor in video card speeds, since no current video card can saturate a AGP 8x bus. So what is the limiting factor?

what happened to the 12" cards? I still have an old one around somewhere.

Another really cool use for an external video card would be laptops or other small form factor machines. They have external audio cards like the Extigy... why not a video? What's so different there?
As many laptops are being used desktop replacements, I dont see why a beefy external video card wouldnt or couldnt be viable. Maybe it's not for the common market but hey neither are the 2 Athlon 64 gaming laptops I just got. The laptop does have a cool bottom panel access for swapping out video cards, but you're still limited to the choice of the weaker small laptop video cards on the market.

 

[Shalmanese]

Well, when you invent a low latency, cheap 2Gb/s link, then we can talk about external video cards.

 

[Burbot]

The difference is that audio requires a lot less bandwidth then graphics. It is quite possible to jam sound signal into confines of USB or Firewire. Video card data, such as textures and models require a lot more bandwidth that simply cannot be provided by modern external interfaces. And no laptop manufacturer is going to spend money on a non-standard expensive high speed connector that would be used by 1% of people that are not satisfied with Mobile Radeon 9600 or a similar internal solution.

 

[k1pp3r]

the best idea for cutting down vid card cost wouldn't be by making it bigger, shrinking a circut makes if run faster. The best idea for vid cards would be interchangable parts, much like the MB's. The GPU and memory chips could be removed from the board and upgraded, for instance i could stick a GPU from a 9800 speed in my 9600 Pro along with the memory, this could cut the cost down by just having to buy the GPU instead of the whole board

 

[SonicIce]

Maybe anything is possible but we're limited to ATX

 

[Pudgygiant]

Call me dense, but how different is a GPU from a RISC processor? They both have limited instruction sets and very restricted purposes.

 

[Steg55]

Well you could do vertex processing etc on the CPU, but that idea was ditched long ago when hardware accelerated 3D graphics cards were first designed. This, as already has been said, allows a very specialised, high effiency processing unit designed to go what a GPU has to do.

So what your suggesting is just 'specialise' a duel opteron set-up or similar and use that for GPU data?

The other major hole in that plan is that GPUs are immensly powerful when compaired with CPUs. A 6800 Ultra is supposed to push 40gigaflops, several times that of a high end P4, and hense doing the calcualations on a couple of CPUs would probebly be slower in that respect as well.

Also, the latest GPUs can process up to 16 pixels at once which when compaired to a single threaded Opteron makes them look SLOW.

Steg

 

[kpb]

Originally posted by: k1pp3r
the best idea for cutting down vid card cost wouldn't be by making it bigger, shrinking a circut makes if run faster. The best idea for vid cards would be interchangable parts, much like the MB's. The GPU and memory chips could be removed from the board and upgraded, for instance i could stick a GPU from a 9800 speed in my 9600 Pro along with the memory, this could cut the cost down by just having to buy the GPU instead of the whole board

too many issues to list with that.

1) You'd have to define a standard pin out in order to have any attempt at this.

2) Bus widths. A 9600 has a 64 or 128 bit memory bus depending on which version and a 9800 has 256 bit.

3) Memory types and speed- Take a look at history of video cards article. Video cards have gone from 183 mhz memory to over a ghz with varing types of memory up to gddr3 on some of the newer cards.

 

[iwantanewcomputer]

I was thinking the same thing a while back
It seems like to remove performance bottlenecks, everything is slowly being moved on the chip and i've heard that this trend should continue.

first math coprocessors were moved on-chip(floating point units i believe), then cache, then memory controllers. it seems that for everything else to keep up with processor improvement, we need to put stuf on the chip to remove the bottlenecks. soon we may see actual dram or some form thereof, and gpus being integrated.

but that just moves more computing and storage onto the chip which means it's still got to have io connections if all this stuff gets integrated into the chip somehow i can imagine a single chip with connections to video displays, hard drives, lan, etc

 

[Sahakiel]

Originally posted by: iwantanewcomputer
I was thinking the same thing a while back
It seems like to remove performance bottlenecks, everything is slowly being moved on the chip and i've heard that this trend should continue.

first math coprocessors were moved on-chip(floating point units i believe), then cache, then memory controllers. it seems that for everything else to keep up with processor improvement, we need to put stuf on the chip to remove the bottlenecks. soon we may see actual dram or some form thereof, and gpus being integrated.

but that just moves more computing and storage onto the chip which means it's still got to have io connections if all this stuff gets integrated into the chip somehow i can imagine a single chip with connections to video displays, hard drives, lan, etc

SoC

 

[BenSkywalker]

Top of the line graphics chips currently push ~double the transistor count of top tier consumer grade processors and they have less cache(which takes up comparitively little die space per transistor). In all honesty from an engineering point integrating the GPU on to the CPU is bass ackwards.

Adding a GPU to a CPU- transistor count ~+200%

Adding a CPU to a GPU- transistor count ~+50%

Call me dense, but how different is a GPU from a RISC processor? They both have limited instruction sets and very restricted purposes.

RISC is reserved to Mac marketing speak now(in terms of comparing "RISC" to "CISC"). PPC currently has a larger instruction set then x86(seriously). Data dependancy is the big issue comparing the architectures. Harrkev's post above gives a rough idea of how to think about how a processor compares to a GPU in terms of how effective they are at handling graphics data but that is massively compouneded by the fact that GPUs are now packing 16 pipelines and escalating quickly. If you were to take a look at simply filtering a single pixel utilizing 128 sample anisotropic(16x w/tril) a processor would need to utilize several hundred cycles while a GPU can output 16 per cycle- GPUs are several orders of magnitude faster then the fastest processor on the market. The fastest CPU out today for that matter can not handle raster functions as fast a Voodoo1- and everyone has a fairly good general idea of how much faster the NV40/R420 is/are compared to that relic. The situation is in relative terms a lot worse for processors when looking at pixel shaders then it is with base filtering. GPUs aren't going anywhere for a long time yet.

 

[dakels]

wow I had no idea GPU's were that complex. I thought they were dumbed down specialied CPU's or math coprocessors in a way. I just saw numbers like 433mhz and thought, why cant it be like a 3ghz P4?

 

[everman]

Your modern GPU is extremely complex. Even though they are clocked lower they're designed to do more instructions per clock cycle, more "parallel".

Anyways, I've always wondered why we can't have video "motherboards". They would really be the same size as cards are today. You could say for example: Buy an nVidia compatible video mobo which will allow you to pop in a range of compatible nVidia GPUs. You could also upgrade the ram (not the same ram you use as system ram, video cards today use faster stuff). Then as time goes on you'd end up getting a new mobo and cpu, maybe use the same ram for a while and upgrade as you go.


Is there any word of possibly dual core GPUs on the horizon? I guess that would be the next evolutionary step.

 

[dakels]

What keeps motherboards from using the high speed (DDR3 is it?) RAM a video card uses? Price?

 

[sao123]

The largest price adding component of any video card is the amount high speed memory.
Most of the gpus in a line of cards from top (high end) to value (budget) use the same gpu core
(possibly underclocked) but the real difference is the trip from 64MB of GDDR all the way to 256 MB of GDDR3.


What i dont like about this is...(assuming no imperfections in the cpu/gpu core) why they sell underclocked versions. I mean, it costs the same to manufacture them all...they all come off the same assembly line.