Nvidia Working on a CPU+GPU Combo

Max Romantschuk writes "Nvidia is apparently working on an x86 CPU with integrated graphics. The target market seems to be OEMs, but what other prospects could a solution like this have? Given recent development with projects like Folding@Home's GPU client you can't help but wonder about the possibilities of a CPU with an integrated GPU. Things like video encoding and decoding, audio processing and other applications could benefit a lot from a low latency CPU+GPU combo. What if you could put multiple chips like these in one machine? With AMD+ATI and Intel's own integrated graphics, will basic GPU functionality be integrated in all CPU's eventually? Will dedicated graphics cards become a niche product for enthusiasts and pros, like audio cards already largely have?" The article is from the Inquirer, so a dash of salt might make this more palatable.

Comments

Heard This One Before

[eldavojohn]

Sounds like Nvidia is just firing back at the ATI-AMD claim from two months ago. Oh, you say that you're integrating GPUs and CPUs? "Well, we can say that too!"

What I don't understand is that I thought GPUs were made to offload a lot of graphics computations from the CPU. So why are we merging them again? Isn't a GPU supposed to be an auxillary CPU only for graphics? I'm so confused.

What I'm not confused about is the sentence from the above article:

DAAMIT engineers will be looking to shift to 65 nanometre if not even to 45 nanometre to make such a complex chip as a CPU/GPU possible.

Oh, I've worked with my fair share of DAAMIT engineers. They're the ones that go, "Yeah, it's pretty good but ... DAAMIT, we just need more power!"

Re:Heard This One Before

[everphilski]

What I don't understand is that I thought GPUs were made to offload a lot of graphics computations from the CPU. So why are we merging them again?

a really, really fast pipe. It is a lot quicker to push stuff from CPU->GPU when they are on the same piece of silicon, versus the PCIe or AGP bus. Speed is what matters, it doesn't look like they are moving the load one way or another (although moving some load from CPU->GPU for vector based stuff would be cool if they had a general purpose toolkit, which I'd imagine one of these three companies will think about).

Re:Heard This One Before

[purpledinoz]

It seems like this type of product would be marketed towards the budget segment, which really doesn't care about graphics performance. However, the huge advantage of having a GPU on the same silicon as the CPU would be a big boost in performance. The low cost advantage has already been attained with the integrated graphics chipsets (like nForce). So that would mean this might be marketed towards the high-performance crowd.

But I highly doubt that nVidia will be able to get a CPU out that out-performs an Intel or AMD, which the high-performance junkies would want. Intel and AMD put a HUGE amount of money into research, development, and fabrication to attain their performance. This is going to be interesting to watch. Hopefully nVidia doesn't dig themselves into a hole with this attempt.

Re:Heard This One Before

[arth1]

At one time floating point was done by software it still is one some cpus.
Then floating point co-processors became available. For some applications you really needed to speed up floating point so it was worth shelling out the big bucks for a chip to speed it up.

Then people started using floats for the convenience, not because the accuracy was needed, and performance suffered greatly as a result. Granted, there are a lot of situations where accuracy is needed in 3D, but many of the calculations that are done could be better done in integer math and table lookups.
Does it often matter whether a pixel has position (542,396) or (542.0518434,395.97862456)?
Using a lookup table of twice the resolution (or two tables where there's non-square pixels) will give you enough precision for pixel-perfect placement, and can quite often speed up things remarkably. Alas, this, and many other techniques have been mostly forgotten, and it's easier to leave it to the MMU or graphics card, even if you compute the same unneccessary calculations and conversions a million times.

Fast MMUs, CPU extensions and 3D graphics routines are good, but I'm not too sure they're always used correctly. Does a new game that's fifty times as graphically advanced as a game from six years ago really need a thousand times the processing power, or is it just easier to throw hardware at a problem?

Re:Heard This One Before

[NerveGas]

I don't think that the CPU->GPU pipe is any limitation. Going from AGP 4x->8X gave very little speed benefit, and on PCI-E connections, you have to go from the normal 16x down to a 4x before you see any slowdown.

Memory size and bandwidth are the usual limitations. Remember that if you want 2x AA, you double your memory usage, and if you want 4x AA, you quadruple it. So, that game that needed 128 megs on the video card, with 4x AA, can suddenly need 512.

steve

Re:Heard This One Before

[LWATCDR]

Exactly.
I was using floating point as an example.
I don't know if Nvidia can pull this off without a partner. Too build a really good X86 core isn't easy. I wonder if they may not do a PPC or Arm core instead. That could make nVidia a big player in the cell phone and mobile video market. At some point there will be portable HD-DVD players.

My crystal ball says.
AMD will produce these products.
1. A low end CPU it integrated GPU for the Vista market. This will be a nice inexpensive option for home and corporate users. It might also end up in some set-top boxes. This will the next generation Geode.
2 A family of medium and high end video products that use Hyperchannel to interface with Opteron and Athlon64 line.

Intel will
Adopt Hyperchannel or reinvent it. Once we hit four cores Intel will hit a front bus wall.
Intel will produce a replacement for the Celeron that is Duo2Core with integrated graphics on one die. This is to compete with AMD new integrated solution.
Intel will not go in to the high end graphics line.

nVidia will fail to produce an X86+GPU to compete with AMD and Intel.
nVidia produces an integrated ARM+GPU and dominates the embedded market. Soon every cellphone and media player has an nVidia chipset at it's heart. ARM and nVidia merge.

Of course I am just making all this up but so what, electrons are cheap.

Thank MicroSoft

[powerlord]

Okay, I admit, I haven't RTFA yet, but if GPUs do get folded back into CPUs, I think we need to thank MS.

No. ... Seriously. Think for a minute.

The major driving force right now in GPU development and purchase are games.

The major factor that they have to contend with is DirectX.

As of DirectXv10. A card either IS, or IS NOT compliant. None of this "We are 67.3% compliant".

This provides a known target that can be reached. I wouldn't be surprised if the DirectX10 (video) featureset becomes synonymous with 'VGA Graphics' given enough time.

Yeah, sure, MS will come out with DX11, and those CPUs won't be compatible, but so what?, If you upgrade your CPU and GPU regularly anyway to maintain the 'killer rig', why not just upgrade them together? :)

It's a logical extension of the NVidia NForce line

[Animats]

I've been expecting this for a while, ever since the transistor count of the GPU passed that of the CPU. Actually, I thought it would happen sooner. It's certainly time. Putting more transistors into a single CPU doesn't help any more, which is why we now have "multicore" machines. So it's time to put more of the computer into a single part.

NVidia already makes the nForce line, the "everything but the CPU" part, with graphics, Ethernet, disk interface, etc. If they stick a CPU in there, they have a whole computer.

Chip designers can license x86 implementations; they don't have to be redesigned from scratch. This isn't going to be a tough job for NVidia.

What we're headed for is the one-chip "value PC", the one that sits on every corporate desk. That's where the best price/performance is.