Tim's thoughts

[zidane1strife]

I don't think you're talking about the PC space - which is a commodity market and therefore loss leading is not particularly viable. Tim is undoubtedly talking about the PC space.

Yes I know... it was a bit off topic... but itt was just a response to the gpu v.s. gpu+cpu combo comment I got... clearly as you've said the prior solution can't be sold at a loss... thus even in a gpu v.s. gpu situation the latter has a higher trans. budget... when you take into account the fact that the latter is a two chip solution... you realize.... that...

If designed well, such a combo could very well outclass many things... such h/w would be shielded by the fact that the next process(45nm) won't be available to reg foundries until the end of 2007 or beginning of 2008 most likely...

 

[Crazyace]

Intel couldnt design a CPU which did as much work per clock cycle as a modern GPU at the kind of frequencies of a modern CPU ... not with air cooling. They can run their CPUs at that speed because most of the transistors on it are sitting idle anyway.

I tend to think that the calculation units run quite well on the intel units ( especially with the newer hyperthreading versions ) - as there is a lot of cache on a P4 I wont disagree about idle transistors in general.

At a time of spiralling fab costs, Intel reducing gross margin to build market share would not seem to be a good plan. Basic economics says that once you have the biggest slice of the pie, it can be worth reducing market share somewhat to make more margin (once you get over 50% in any market in particular).

My original point was that it might be difficult for Intel to convince people to continually upgrade their cpu's due to general usage not really requiring more power. If they put 'gpu' processing into the mix, then PC manufacturers may just place one expensive intel processor in with minimal video support, knowing that the price points will drop... Dual processor machines would sell as 'traditional' high end games rigs, and power numbercrunchers.

To be honest the only card placed in most machines these days is the graphics card... so if replacing the cpu improved both graphics and general computing it would be worth more incremental business in the long run.

Hmm - we know that the R300 has 8 pixel pipes, each capable of executing 5 instructions per clock (1 texture, 2 scalar, 2 vector), and that it has a long enough pipeline to absorb a texture cache miss without stalling - I would guess that you need about ~50 (?) pipeline steps to do that. Which would give 8*5*50 = 2000 instructions in flight. Which might not be the most meaningful comparison you can do.

I suppose in theory the P4 can dispatch 4 integer ops, one mmx vector integer , one sse2 vector float percycle at nearly 10x the clock rate.. Absorbing a texture cache miss every cycle is pretty impressive though ( which makes me believe it isn't actually true :? )

As a final point I expect any cpu that aims to compete with a graphics chip to contain optimisations that improve it's performance in graphics applications... MPEG encoding has 'special' support instructions, why not texture loading and filtering.[/code]

 

[Dio]

When texturing using mipmaps, R300 produces very nearly 8 pixels per clock. This is roughly equivalent to nearly 8 texel misses per clock cycle (as with mipmapping you have a roughly 1:1 pixel to texel ratio).

 

[Tahir2]

My argument is that if Crazyace is correct then why are we not hearing about this in the Electronic Engineering circles and why hasnt someone attempted this CPU doing GPU work?

Any idea worth a damn has been patented, twice )to borrow a phrase from MFA

 

[Tokelil]

They can run their CPUs at that speed because most of the transistors on it are sitting idle anyway.

I thought 90nm chips leaked around 50% current when idle which would make it quite hard making very big chips at insane clockspeeds air cooled even when the PU is idle. So adding logic to control graphic processing to the CPU would hamper CPU clockspeed growth.

 

[nobie]

We are very very quickly approaching the point of diminishing returns for real-time computer graphics, as CG did a long time ago, where the people actually buying the games won't see enough of a difference to care anymore (i.e. very few people that I talk to outside of computer graphics seem to notice any real difference between Toy Story 1's graphics, and Finding Nemo's).

I wonder if this will mean an increased market for licensed graphics engines like Jupiter/Lithtech. I can imagine a situation 10 years into the future where a competitive graphics engine would take years and millions of dollars to build from scratch.

edit: another thing, has anyone tried the DX9 software reference driver? <1 fps. CPUs have a long way to go before they can take over for GPUs

 

[Ailuros]

Never heard that myth. Where did you hear that? I mean, it's quite obvious that the price goes down with time. Part of it is because the low cost hardware becomes good enough for most people, so more low end hardware is produced and bought. I believe that's true for most hardware (not just computers). There's a point above which fewer people feel the need to go.

People are constantly complaining that hardware gets more expensive as time goes by, whereby ironically it's the exact opposite.

Those reactions usually pop up everytime they see a new high end product with a higher than 400-500$ pricetag released.

I wasn't just considering those but if you look at the performance of CPU VS GPU solutions for just VS you'll see that, even with the CPU's much higher frequency, the GPU will easily outperform it.

Did that already a page or two before. I switched between HW T&L/VS off and on in UT2k3. Here again for those who missed it:

SW T&L on 2.0GHz Athlon: 27.9 fps
HW T&L/VS on 4 VS@325MHz + 2.0GHz Athlon: 50.4 fps

 

[akira888]

Hi Simon,

I believe that as an example the PPC970 has around 220 instructions in flight at any one time...

Impressive though it is, that figure, no doubt, is the total of all instructions in various stages of decode. There is still going to be a relatively small number of pipelines running in parallel - this is still small compared to, for example, even the very first powerVR system.

Crazyace, the reason as to why PPC970 has actually 212 instys in flight is due to the fact that this larger "window" allows the chip more flexibility to optimally schedule instructions in a way to maximize throughput. PPC970 is a five-way superscalar core however and can therefore only retire five instructions per clock at most. Its dhrystone rating FWIW IIRC was only around 4.5 times that of the GameCube's PowerPC Gekko.

Which also means that a simple texel retrieval would take ten cycles, during which time nothing else would happen (tesselation, coordinate interpolation, subsample blends, transformation, pixel shading...)

 

[passerby]

why are we not hearing about this in the Electronic Engineering circles and why hasnt someone attempted this CPU doing GPU work?

PS2 was created along the lines of such design philosophies. Taken in certain perspective, it fits the description well. But it has its own unique architecture, so such an approach is likely not useful for the PC platform.

 

[Pavlos]

My argument is that if Crazyace is correct then why are we not hearing about this in the Electronic Engineering circles and why hasnt someone attempted this CPU doing GPU work?

Any idea worth a damn has been patented, twice )to borrow a phrase from MFA

Gordon Moore made the same prediction about 30 years ago: More and more functionality will be integrated in a single chip as the time passes. And when someone is right for about 30 years you’d better believe him.

He gave some nice presentations about a year ago about the future of Intel CPUs until the end of the decade. Guess what? They will include integrated graphics hardware functionality. He’s not explicitly saying GPUs will be obsolete, but when Intel already has 35% of the graphics market share, I think the future is obvious.

A very interesting note is that some of the graphics functionality will be hardwired, not for performance reasons but to lower the power consumption.

So, it’s not exactly software rendering, but the GPUs sooner or later will be redundant.

PS. Sorry, I don’t have the time to search for the link to Moore’s presentations right now, but I’ m sure someone else can point to them. They are somewhere at the Intel’s site.

 

[jvd]

He gave some nice presentations about a year ago about the future of Intel CPUs until the end of the decade. Guess what? They will include integrated graphics hardware functionality. He’s not explicitly saying GPUs will be obsolete, but when Intel already has 35% of the graphics market share, I think the future is obvious.

but is that what we want ? I really don't want intel controlling the future of 3d graphics because they may never move past whatever is the cheapest thing for intel to sell. Gpus are very needed for advancement . There will allways be someone willing to spend extra to get a better framerate or higher aniso or fsaa or a higher res for the game. So as long as there are people willing to buy it there should be a company willing to keep pushing the tech and another company to make sure the prices stay in a sane area .

 

[sonix666]

While I don't see it happening very fast, I wouldn't be surprised to see GPUs being replaced again by the CPU. Looking at what many CPU manufacturers think of the future, it seems that we'll be getting multiple simpler CPUs on one die. That could bring back parallellity that GPUs right now have.

 

[Crazyace]

Did that already a page or two before. I switched between HW T&L/VS off and on in UT2k3. Here again for those who missed it:

SW T&L on 2.0GHz Athlon: 27.9 fps
HW T&L/VS on 4 VS@325MHz + 2.0GHz Athlon: 50.4 fps

Hi Ailluros,

The interesting point about your comparision comes when you consider that running UT2k3 'game' code probally takes 50% of your cpu cycles... So effectively the SW T&L will probally be just as quick on its own as the HW T&L.

Hi Akira - a Dhrystone rating isn't really much to compare graphics functionality - I wonder what score you would get from compiling the dhrystone test in Cg... Also you are comparing the 'current' architectures against future machines which will have further optimisations.

 

[Simon F]

(as with mipmapping you have a roughly 1:1 pixel to texel ratio).

4:1 would make Mr Nyquist happier.

 

[see colon]

here's a quote from tim regarding his predicted return to software rendering....

-------------------------------------------------------------------------------
http://www.gamespy.com/legacy/interviews/sweeney.shtm

"Gamespy - Do you ever think you'll tinker with a voxel engine, or combining a voxel and a polygon engine?

Tim - I don't think voxels are going to be applicable for a while. My thinking on the evolution of realtime computer graphics is as follows:

1999: Large triangles as rendering primitives, software T&L.

2000: Large triangles, with widespread use software-tesselated curved surfaces, limited hardware T&L.

2001: Small triangles, with hardware continuous tesselation of displacement-mapped surfaces, massive hardware T&L.

2002-3: Tiny triangles, full hardware tesselation of curved and displacement-mapped surfaces, limited hardware pixel shaders a la RenderMan.

2004-5: Hardware tesselation of everything down to anti-aliased sub-pixel triangles, fully general hardware pixel shaders. Though the performance will be staggering, the pipeline is still fairly traditional at this point, with straightforward extensions for displacement map tesselation and pixel shading, which fit into the OpenGL/Direct3D schema in a clean and modular way.

2006-7: CPU's become so fast and powerful that 3D hardware will be only marginally benfical for rendering relative to the limits of the human visual system, therefore 3D chips will likely be deemed a waste of silicon (and more expensive bus plumbing), so the world will transition back to software-driven rendering. And, at this point, there will be a new renaissance in non-traditional architectures such as voxel rendering and REYES-style microfacets, enabled by the generality of CPU's driving the rendering process. If this is a case, then the 3D hardware revolution sparked by 3dfx in 1997 will prove to only be a 10-year hiatus from the natural evolution of CPU-driven rendering."
------------------------------------------------------------------------------------

so i guess we only have a few more years before we find out if he's right or not. what do i think? i'm quite certain that software rendering will not make a comback as tim suggests. it might rear it's ugly pixelated head up a few times here and there, but to say it will be the primary rendering platform is pretty far fetched, from what i see.

gpu's are getting more programable every generation, and i believe rendering meathods currently looked at as "non-traditional" (those available only in software engines now) will begin to be supported in hardware. doesn't dx10 have support for hardware voxel rendering? how long do you thing before we have hardware support for splines or other, non-polygon primatives in hardware? there won't be a need to shift back to software rendering because the hardware will have adapted and will be capable of doing everything you could achieve via software, only faster, with faster memory, more bandwidth, and higher accuracy.

if anything i see gpu's becoming even more general purpose where non-graphical tasks can be performed. maybe the video card market being replaced by a "gaming card" market, where the add in board has more than just a gpu, but also dedicated hardware for physics calculations, ai, ect; or possibly those feature being added to the gpu core altogether. budget solutions will only have programable shaders, and will have to rely on "software physics". and the term "software physics" will carrry the dirty (low end) connotation software t&l currently does.

c:

 

[ET]

I think the way of the future is to have more generalised computing resources. So if I want more vector processing for graphics you simply plugin a CPU which does that job. Of course this decentralised architecture already exists to some degree it's just taking time.

It's not a matter of time. It's a paradigm that has always existed. It's just that whenever a function proves useful enough it's integrated.

 

[Dio]

When texturing using mipmaps, R300 produces very nearly 8 pixels per clock. This is roughly equivalent to nearly 8 texel misses per clock cycle (as with mipmapping you have a roughly 1:1 pixel to texel ratio).

I need to correct myself. R300 can't quite achieve 8 pixels/clock in single texturing mode. However, it can achieve 8 texels/clock in 2-texture multitexture with no problems.

 

[Dio]

(as with mipmapping you have a roughly 1:1 pixel to texel ratio).

4:1 would make Mr Nyquist happier.

I was assuming that bilinear overlap would imply that 3 of the 4 are already in the cache. I think.

 

[Dio]

In 1996 I reckoned that I only had a job in this industry for about five more years, because after that things would have got so powerful we wouldn't need new chips any more.

In 2001 I looked at it again, and still reckoned I only had a job for about five more years.

In 2003... well, it looks like about five more years to me...

 

[Dio]

2006-7: CPU's become so fast and powerful that 3D hardware will be only marginally benfical for rendering relative to the limits of the human visual system, therefore 3D chips will likely be deemed a waste of silicon (and more expensive bus plumbing), so the world will transition back to software-driven rendering.

I'm not convinced. This implies that in less than 4 years CPU's will be a lot more capable in 3D rendering than a current high-end videocard (I wouldn't say we were at 'the limits of the human visual system' yet).

3 1/2 ago was the launch of the GeForce 2 GTS. Could a current CPU software renderer match that for performance? That's an interesting question... is Nick around, or any of the other guys who are working on software pipelines?