NVIDIA Working on GPGPUs for Macs?

[MacRumors]

Appleinsider claims that NVIDIA is working on bringing "general-purpose computing on graphics processing units" (GPGPUs) to the Mac.

GPGPUs are described as a new type of graphics processors that can perform complex computations typically reserved for the system's primary CPU.

The technology -- in Nvidia's case -- leverages a proprietary architecture called CUDA, which is short for Compute Unified Device Architecture. It's currently compatible with the company's new GeForce 8 Series of graphics cards, allowing developers to use the C programming language to write algorithms for execution on the GPU.

According to Appleinsider, the GPGPUs can be beneficial in a number of applications with complex mathematical requirements, such as raytracing, scientific applications, cryptography, and audio and image processing.

NVIDIA's $1500 Telsa card is the first example of this class of graphics card. When launched for Mac, these GPGPUs will likely be a high-end build-to-order option for Mac Pros.

Article Link

[ricosuave]

I knew this day would come! I have no idea of what you guys are talking about.

[haunebu]

In the olden days, we greybeards called this AltiVec.

[Abstract]

This GP GPU sounds like another CPU. I mean, mathematical tasks such as image and sound processing? If they're really interested in another CPU, just add another CPU. It doesn't have to be from NVIDIA. It can be from Intel, no?

I don't really get it. Or is this just like what AltiVec used to be?

[exabytes18]

GPGPU is more-or-less for highly parallel operations. It can achieve an order of magnitude higher performance than a normal CPU.

I suppose it's like altivec in the sense that it's not x86... GPGPU is more like a single processor with hundreds of more specialized cores.

[MrCrowbar]

Well GPU's are good in dumb parrallel processing, which is what rendering 3d images is all about. Your CPU can't do it that fast because it has to be able to do all kinds of differnet stuff.

[bigwig]

Specialized coprocessors were all the rage in the 80s and into the very early 90s. Rapidly advancing speeds of general purpose CPUs rendered them moot and that line of development was dropped. Now we're back to repeating history. No different than high-speed interconnects, really. They started out serial, went parallel for the next round of speed bumps, and now we're back to serial.

[iSee]

Originally Posted by Abstract

This GP GPU sounds like another CPU. I mean, mathematical tasks such as image and sound processing? If they're really interested in another CPU, just add another CPU. It doesn't have to be from NVIDIA. It can be from Intel, no?

I don't really get it. Or is this just like what AltiVec used to be?

This is really an alternative to things like Altivec. It is a really good development to my mind because it might be better. If the GPU guys apply all they know about highly-paralllelized computation, maximizing throughput, etc. to more general computation, they might make some incredible things possible.

It's also interesting because it decouples the coprocessing unit from the CPU, which could allow for more flexible, customized computing environments.

There are some interesting problems just dying for coprocessors and this might be a way to address that. For example, there was talk at one point about "Physics" coprocessors. There's been some work in that area, but nothing mainstream really caught on. Some other possibilities: AI processors. Even sound processing seems to be moving back to the CPU. Individually, these things don't seem to create a viable market for a coprocessor. But a most general purpose coprocessor might help some of this stuff to take off.

On the other hand, it seems likely that computers will be getting more and more CPU cores, which can be utilized for the same purposes.

So we've got a competition between the CPU guys and the GPU guys to provide us with the fastest, most useful, cheapest processing power. And that's got to be good for us users.

[SPUY767]

Originally Posted by Abstract

This GP GPU sounds like another CPU. I mean, mathematical tasks such as image and sound processing? If they're really interested in another CPU, just add another CPU. It doesn't have to be from NVIDIA. It can be from Intel, no?

I don't really get it. Or is this just like what AltiVec used to be?

GPU's are already capable of vastly more processing power than today's CPU's. The problem is that the GPU's are very specific in what they can do. an 8800GT has 112 Pixel pipelines, you can view that as 112 seperate thread execuition paths. It can run highly specialized processes like a 112 core processor. GPU's only run in the 6-700 MHz range, but with parallel processing capability like they have, who needs clock speed. With AMD snatching up ATi, I see intel potentially forming a quasi-partnership with nVidia, with the combination of a processor like the Yorkfield, and a GPU that is GP capable, bringing the next wave of computer graphics. On-Demand ray-tracing.

[skunk]

Originally Posted by SPUY767

but with parallel processing capability like they have, who needs cock speed.

Are we in the right thread?

[Cromulent]

Originally Posted by haunebu

In the olden days, we greybeards called this AltiVec.

This is nothing like AltiVec.

Edit : AltiVec was a way to pass 128bit instructions through the CPU in 1 cycle rather than having to split them up into separate instructions. This is a method of passing instructions to the graphics card which execute them faster and in parallel.

[takao]

this is nothing like altivec .... looks like the greybeards don't know what they are talking about

in fact similiar stuff has already been able before using custom C libraries etc. but it was mostly too difficult to do and too tied to plattforms (thus nobody invested serious money into it)
speedup for scientific apps can be huge if the amount of parallelism can be really used

thou actually i would think the great breakthrough will come once the GPu can be directly intigrated into the operating as an another "pseudo - cpu" which would make it much more easier to get the performance out of it... or at least try to get MPI/OpenMP to work with it

edit: to those saying "why not add a second CPU"
well the GPU has to be there anyway on a desktop machine and most of the time will be sitting around idling when no 3D apps are running so why not use it for something ...
also GPUs are more parallel than another CPU and much, much better at floating point operations

[avkills]

Originally Posted by Cromulent

This is nothing like AltiVec.

Edit : AltiVec was a way to pass 128bit instructions through the CPU in 1 cycle rather than having to split them up into separate instructions. This is a method of passing instructions to the graphics card which execute them faster and in parallel.

This guy needs more grey!

AltiVec was 128bit wide, but it was usually used to process 4 32bit vector calls per cycle. So yes this is like AltiVec but with a lot more parallelism.

-mark

[headfuzz]

So am I right in thinking this is along the lines of the x87 maths coprocesssor architecture from the 80s?

[hugodrax]

Originally Posted by Cromulent

This is nothing like AltiVec.

Edit : AltiVec was a way to pass 128bit instructions through the CPU in 1 cycle rather than having to split them up into separate instructions. This is a method of passing instructions to the graphics card which execute them faster and in parallel.

Wow, being clueless there. Altivec and SSE is the same crap, Single Instruction Multiple Data. Each implementation have their own strengths and weakness but overall they provide the capability to do high performance Vector calculation. Stuff that required multimillion dollar Crays.

[Evangelion]

Originally Posted by haunebu

In the olden days, we greybeards called this AltiVec.

Except Altivec was few orders of magnitude slower... Altivec is for all intents and purposes identical to SSE. Sure, it was better at somethings while being worse at some others. But the basic idea was/is identical. So if you can compare Altivec to this, you can compare SSE as well. And I don't think you can do that, since there are significant differences between this and SSE/Altivec.

[gkarris]

Originally Posted by ricosuave

I knew this day would come! I have no idea of what you guys are talking about.

Whatever it is they're talking about, I hope we can get it in the Mini...

[exabytes18]

Tesla is expensive, but CUDA's not limited to just Teslas. You can execute it on any 8-series graphics cards. I don't think it was ever intended to be a mainstream technology, at least that's the sense I get from the marketing of it. Your average Joe is never going to use 300 GFLOPS.

It's there for those who need such compute power. Just because it's not supported by major software vendors doesn't mean that there's not any developers using it.

[Cloudane]

Here was me hoping it was a Game Player's GPU

(Actually for my uses I don't mind the one in the iMac too much but still)

[EagerDragon]

Sorry Intel, sounds like the days of x86 instruction set CPU(s) are coming soon to an end. In a few years these babies maybe be providing most of the horse power for general purpose computing.

[bête noire]

Sounds like just the thing for BOINC tasks.

Originally Posted by MacRumors

When launched for Mac, these GPGPUs will likely be a high-end build-to-order option for Mac Pros.

Yeah, that'd be right. If it's any good, Apple will overprice it.

[Nichod]

This is one advantage of the Apple marketshare growing. Increased support from the hardware community.

[SirOmega]

The only issue I have is that I think the current GF8 series of cards are only FP16 - 16 bit floating point aka single precision. This may not be enough for most apps. It would be great for running test runs, but I know for scientific calcs FP16 isn't sufficient, they need FP32.

[AidenShaw]

Originally Posted by SirOmega

The only issue I have is that I think the current GF8 series of cards are only FP16 - 16 bit floating point aka single precision. This may not be enough for most apps. It would be great for running test runs, but I know for scientific calcs FP16 isn't sufficient, they need FP32.

I don't understand your post at all.

Standard "single" floating precision is 32-bit floating, "double" precision is 64-bit floating. 16-bit floating is an oddity that has very little traction (it is not supported by the x86 architecture, for example).

See http://en.wikipedia.org/wiki/Floating_point

The "half-floats" in CUDA probably aren't interesting at all. http://forums.nvidia.com/index.php?showtopic=36286

In fact, since modern CPUs run 64-bit floating at almost the same performance as 32-bit single precision - the use of 32-bit float is dropping. Using "half-precision" float would not be interesting for very many applications.

[qtx43]

Originally Posted by AidenShaw

I don't understand your post at all.

Standard "single" floating precision is 32-bit floating, "double" precision is 64-bit floating. 16-bit floating is an oddity that has very little traction (it is not supported by the x86 architecture, for example).

See http://en.wikipedia.org/wiki/Floating_point

The "half-floats" in CUDA probably aren't interesting at all. http://forums.nvidia.com/index.php?showtopic=36286

In fact, since modern CPUs run 64-bit floating at almost the same performance as 32-bit single precision - the use of 32-bit float is dropping. Using "half-precision" float would not be interesting for very many applications.

I disagree. Granted, as time goes on, these video cards will get more memory, but like graphics, sometimes you don't care a whole lot about the precision. For example, imagine doing real time image processing on a high resolution video, using fft or other transforms. On a bargain basement PC. It's coming.