unified shader model 4.0

dh19440113

n00b
Joined
Mar 3, 2004
Messages
52
What do you guys think about the new unified shader (instruction) of shader model 4.0?

What about unified shader hardware?
Do you think it will be better than the separated pixel/vertex shader nvidia 32/32 and ati 16/48 have right now?

Personally I think a unified cross vector and dot scalar units will be a balance of speed and stability. For example, a unified 64 pipeline can allocate pipeline resource according to the need of the 3d software so frame rate will remain steady and playable.

However, a unified pipeline is slower than a dedicated pipeline design to handle specific pixel/vertex.
The result is slower maximum fill rate in exchange for steady fill rate that never drop below threshhold.
 
Personally, I'd rather see a boost to minimum framerates than see even higher maximums. Your minimum always happens at the peak of the action when you really need the performance most :mad:

Getting 100 fps+ doesn't do you much good if that only occurs when you're staring at a wall or straight up at the sky. Of course, I could be way off here and not really have a clue.
 
I'm pro whatever gets me better playability with less heat/noise/power consumption. GPU performance cannot continue to be raised by simply throwing more transistors/clock speed at the problem. The dies get too big, they draw too much power, and keeping them cool within reasonable cost gets more and more difficult each generation.

The CPU industry is moving toward performance per watt. The GPU industry isn't the same as the CPU industry, but more raw firepower is not going to be a complete answer going forward.

Does this mean we need unified shaders? Possibly. There are engineers and the liek that make a lot more money than me that get paid to make decisions like that. They even went to school to figure out stuff like this :eek:
 
Since window vista uses window graphics foundation, and WGF is SM4, vole will force the adaptation of unified shader hardware whether we like it or not. :confused:
 
dh19440113 said:
Since window vista uses window graphics foundation, and WGF is SM4, vole will force the adaptation of unified shader hardware whether we like it or not. :confused:


exactly.....but i think it'll be good
 
Um, nothing about the SM4 spec requires unified shader hardware. It's just that the VS/GS/PS instruction sets are, aside from a few exceptions, (e.g. the stream out functions for GS) exactly the same. Does that mean it may be better to have ALUs that can be used by any part of the rendering process? Possibly. I do not know, because I'm not a HW engineer at ATi or NV. More importantly, should I, or any of us care? No! If you buy a video card, it'll be based on the performance it gets in your games, not whether or not it has a hardware feature that is totally transparent to the running of your games.

Since window vista uses window graphics foundation, and WGF is SM4, vole will force the adaptation of unified shader hardware whether we like it or not.

WGF does not exist any more in any form. What was WGF1 is now known as D3D9Ex (which is simply D3D9 running through D3D10 function calls, and picking up the API benefits of D3D10 along the way) and what was WGF2 is now known as D3D10.
 
DougLite said:
I'm pro whatever gets me better playability with less heat/noise/power consumption. GPU performance cannot continue to be raised by simply throwing more transistors/clock speed at the problem. The dies get too big, they draw too much power, and keeping them cool within reasonable cost gets more and more difficult each generation.

The CPU industry is moving toward performance per watt. The GPU industry isn't the same as the CPU industry, but more raw firepower is not going to be a complete answer going forward.

Does this mean we need unified shaders? Possibly. There are engineers and the liek that make a lot more money than me that get paid to make decisions like that. They even went to school to figure out stuff like this :eek:

I've gotta agree with this. My current videocard (Geforce 6800) probably draws close to 2 times the power compared to what my CPU does while gaming. (Celeron M 350J)

I'd hate to see what the new X1900XTXs and 7800GTX 512's draw.
 
Cypher19 said:
Um, nothing about the SM4 spec requires unified shader hardware and what was WGF2 is now known as D3D10.

DX10 SM4 requires unified code across the shader, so if unified shader hardware isn't inplimented; then its not dx10 sm4 hardware is it?

Just like x800 series isn't dx9c sm3 compatible.
But dx10 spec shows that its more programable than dx9, therefore it might be possible to sparate pure shader instruction to accomadate separate vertex and pixel shaders. But I don't think Nvidia will risk not implimenting unified shader Hardware in G80. :cool:
The truth is I wouldn't even be supprised if dual gpu comes on a single core. I mean its alot more efficient than SLI/crossfire.
 
I have to agree with Doug. And yea it looks like we are going to have to jump on sm4 once dx10 and vista are out, I do not nessesary think it would be a good thing at start becasue I hardly doubght that the first dx10 cards will be able to put out enough power for sm4. I like good graphics, but when you averaging 5fps most of the time, then it is not fun :(
 
DougLite makes a solid point. Regardless of how much raw power there is, I'd MUCH rather see those few instances of low framerates pounded out into a smooth gaming experience, then have more maximum framers per second. Sure, it's nice to have really high framerates when you're walking down a hall, but that doesn't matter so much as the battles and such.

In other words, I agree that overall smoother play is more important than "higher highs." :)

Me = Pro Unified Shader Model 4.0 (For now, but we'll have to see ;) )
 
dh19440113 said:
However, a unified pipeline is slower than a dedicated pipeline design to handle specific pixel/vertex.
The result is slower maximum fill rate in exchange for steady fill rate that never drop below threshhold.

I'm not too sure about that. Since you have unified pipes, you can have more pipes on a single GPU than pixel pipes. So while a single pipeline may be slower, you may be able to use more of them at the same time, thereby still getting the same fillrate, or more.
I think the main problem might be the allocation of these pipelines. Even with fixed pipelines we are already seeing quite diminished returns when you go over, say 12 pixel pipelines. As polygons get smaller, there will be relatively more edges on the screen, and since pipelines are allocated as quads, you will more often get 1 or more idle pipelines in a quad, because it falls outside the polygon.

I really can't say which is better, it depends on various factors, like how much slower a unified pipeline is, how many you can put on a chip, and how efficient the allocation will be in an average scene. The idea sounds elegant though, and elegant is often good.
 
I meant pipe for pipe its slower, and you definitely can have more shader in unified hardware. Just like pentium 4 alot of MHZ but little performance gain LOL. :D

Hey if the final hardware fail to be faster than current gpu, they can alway consider dual gpu on a single core that render alternate frames. The cpu seems to go that way why not gpu.
 
dh19440113 said:
Hey if the final hardware fail to be faster than current gpu, they can alway consider dual gpu on a single core that render alternate frames. The cpu seems to go that way why not gpu.

Because a GPU is a completely different thing. With CPUs you've always designed around the concept of processing one thread at a time, and trying to do that as quickly as possible. Since we can't seem to get more performance on a single thread (clockspeeds can't be scaled much higher, and adding extra execution units is not an option, since they can't be kept busy anyway), the only way to get more performance is to process multiple threads at the same time. But this is a less efficient option, and also one that doesn't scale as well.

With GPUs we've always had massive parallelism. One of the main assumptions with triangle rasterization is that every triangle is independent of every other triangle, and every pixel is independent of every other pixel. Which means that the order of rendering of triangles and pixels is pretty much unimportant, and they can be rendered independently (give or take a few exceptions).
So if you compare it to a CPU, then every pixel pipeline is a CPU in its own right, and you have a sort of super multicore GPU.
With multiple GPUs however, you have all kinds of synchronization issues, and the parallelism can quickly fall apart in very common situations (eg render-to-texture, where one GPU renders a texture, and the second GPU has to wait for it to complete).

I think it's far more efficient to just build one core with extra pipelines than it is to put two or more GPUs on a single die.
Ofcourse when you can't build larger dies, using multiple dies on a single card, or even a SLI/CrossFire setup, is an acceptable way to increase performance, but as long as you're working with a single die, I think one mega-core will always be more efficient.
 
You guys make a solid point, but I think it wil lbe basically the smae thing as we had with sm3.0. tell me, about 4 months ago when x850xt and 6800ultra were top dogs on the market, were 6800's able to utilize sm3.0 smoothly? Hell no, there were tons of threads about that problem, I have a feeling that sm4.0 will be something simular.
 
Russian said:
You guys make a solid point, but I think it wil lbe basically the smae thing as we had with sm3.0. tell me, about 4 months ago when x850xt and 6800ultra were top dogs on the market, were 6800's able to utilize sm3.0 smoothly? Hell no, there were tons of threads about that problem, I have a feeling that sm4.0 will be something simular.

What problem exactly?
Afaik, SM3.0 was really no slower than SM2.0, so using it was no problem. Unlike ofcourse the problem with the FX series, when SM2.0 was much slower than SM1.x.
But again, on the R9x00 series, there was no difference.
I don't think SM4.0 has any reasons to be much slower either, apart from badly designed hardware ofcourse.
 
DX10 SM4 requires unified code across the shader, so if unified shader hardware isn't inplimented; then its not dx10 sm4 hardware is it?

A unified instruction set is totally different from unified hardware. Unified hardware simply means having the same hardware (e.g. ALUs, memory access) to perform VS/GS/PS execution. That is, instead of having ALUs that are dedicated purely for PS usage, they can be used for processing VS's and GS's. Having a unified instruction set HELPS in that area, but by no means is unified hardware a requirement. Hell, current video cards with SM3 or SM2 could probably do unified hardware right now, s.t. that hardware can perform the union of the VS/PS instruction sets.

Scali said:
What problem exactly?
Afaik, SM3.0 was really no slower than SM2.0, so using it was no problem. Unlike ofcourse the problem with the FX series, when SM2.0 was much slower than SM1.x.
But again, on the R9x00 series, there was no difference.
I don't think SM4.0 has any reasons to be much slower either, apart from badly designed hardware ofcourse.

SM3 is not slower than SM2, but some of the features that SM3 introduced (i.e. vertex texturing) are, for the most part, fairly half-assed both on NV and ATi's part. For ATi, I think everyone here has heard about the fiasco with it not officially support vertex texturing and I'll leave it at that. For NV, the vertex texturing support is there, yes, but it simply stinks. As it is, it's currently too slow to be used on a practical scale.

SM4 is already going to be similar. For example, ATi and NV have been saying that even though the GS can emit 1024 primitives a pass, they're telling devs to keep it down to a couple dozen at the most. Also, there's still the issue about how fast the other features of SM4 like stream out, bool/int support, depth-stencil buffer lookups, or bitlevel control of variables are. Even beyond that, going into the rest of D3D10's Everest-esque feature set, there's still the issue of things like how fast memory virtualization will be, support of atypical texture formats, and so on. It'll hardly be a case of badly designed hardware, but more likely not enough time/space to pack all of D3D10's features in. Chances are that a lot of (more obscure parts of...) the feature set will be emulated on software or something simply because they could not put it into the hardware. In fact, because ATi and NV can no longer bicker about picky things like "we've got HDR" due to the 100% guaranteed feature set, the only option left IS performance of various features of D3D10, including SM4.
 
Cypher19 said:
A unified instruction set is totally different from unified hardware. Hell, current video cards with SM3 or SM2 could probably do unified hardware right now, s.t. that hardware can perform the union of the VS/PS instruction sets.

I think you misunderstood what i was originally trying to say.
sm4 =unified shader instruction The hardware from orig post was a typo. :p

current hardware can do sm4 with some innovative programming but it will be so slow LOL :D .
Nothing takes advantage of unified instruction like unified hardware.
 
DougLite said:
The CPU industry is moving toward performance per watt. The GPU industry isn't the same as the CPU industry, but more raw firepower is not going to be a complete answer going forward.

They can do this performance per watt by getting smaller and smaller to create, Intel is at 45 nanometer now which is in testing I believe and do out next year I think. AMD is just now getting to 65 nanometer.

Also, CPU industry is going towards dual cores/quad cores
GPU industry already has dual GPU solution in SLI and Crossfire and even 2 gpus on one card.

I'm taking a guess that if they seem to hit a wall, just throw more GPU dies together and Dual GPU here we come.

Isnt there a rumor going around that G80 is just that, a dual GPU offering?
 
Cypher19 said:
For NV, the vertex texturing support is there, yes, but it simply stinks. As it is, it's currently too slow to be used on a practical scale.

I think that depends on your expectations.

Cypher19 said:
It'll hardly be a case of badly designed hardware, but more likely not enough time/space to pack all of D3D10's features in.

Well yes, I suppose you can say the risk of 'badly designed hardware' is a lot bigger with SM4.0. But as long as the 'main features' are fast enough, it should be quite useable, although not every single detail may work as expected. Same thing with SM3.0. I think it's worked just fine since the first 6800.
 
rmed64 said:
They can do this performance per watt by getting smaller and smaller to create, Intel is at 45 nanometer now which is in testing I believe and do out next year I think. AMD is just now getting to 65 nanometer.

Also, CPU industry is going towards dual cores/quad cores
GPU industry already has dual GPU solution in SLI and Crossfire and even 2 gpus on one card.

I'm taking a guess that if they seem to hit a wall, just throw more GPU dies together and Dual GPU here we come.

Isnt there a rumor going around that G80 is just that, a dual GPU offering?
GPUs have been "multi core" for close to a decade now. CPU architecure is entirely different from GPU architecture. While splitting GPUs into multiple dies may make them easier to manufacture (smaller dies get higher yields and are easier to bin for high clock speeds), paralleling multiple GPUs doesn't scale nearly as well as multiple CPUs. Benchies show that CPU SMP gains 95%+ more performance, while SLI/Crossfire gains are 70-75% under the best of conditions.
 
SMP just got to that point tho, it wasn't long ago an extra CPU only showed about a 45-50% improvement over a single.

I think the multicore GPU comment was more directed at the current trend of SLI/Crossfire, I personally don't see why there aren't more of them sold as a single dual chip card. I guess the price is just too high to really sell, but thats where I'd go rather than the added trouble of 2 cards, 2 slots taken up, etc...

edit:
GPUs have been "multi core" for close to a decade now.
multicore GPU is much more comparable to massively multithreaded CPU, imho.
 
computational power wise yes its more comparable, but each GPU needs its own set of ram, which is why its not cost effective.
 
I guess there's really no way around that. If there was a dual core 7800, it would still need a 512bit memory bus or 2x 256bit.

SLI on one card still seems a natural step vs the 2 card method.
 
Back
Top