What Are The Functions Of A GPU?

[Hornbein]

What are the operations of a GPU? I tried to look it up online and all I could find was "they speed up graphics" and long lists of brand names. They do "vector operations." I want a more technical source of info. Is that linear algebra? If so, what operations?

 

[Baluncore]

It comes down to the two different meanings of the term "vector", which in computing, can refer to both a 1D array of many elements, or a mathematical vector, arrow, or phasor, represented by a single complex number.

A GPU is used to speed up graphics, where rows, columns or blocks of pixels need to be processed in parallel, to get a high throughput.

A "vector processor" performs the same arithmetic operation on all the elements of a vector (a 1D array) of elements. A GPU can do that on thousands of elements, all at the same time, in parallel.

A GPU can do "vector rotations" using complex numbers, if that is what you require, and they can do it on thousands of elements from an array in parallel.

 

[DaveE]

I'm no expert, but it's basically vector/matrix operations. MAC (multiply-accumulate) and inversion operations are the common benchmarks. These are crucial to solving linear algebra equations. This includes things like coordinate transformations, rotations, scaling etc. The emphasis is on efficient operations on large data sets. If I had to pick one thing to highlight it would be multiplying two large matrices efficiently.

 

[Hornbein]

Multiplying matrices. So for graphics it's rotation and scaling. Translation would be matrix addition. I presume that the matrices are of the same dimension.

For AI I suppose projections are important and maybe change of basis. I'm interested because AI took off when they figured out how to use GPUs to train their models. I read that this happened gradually so no one person gets credit. GPUs are also used for mining bitcoin.

 

[sbrothy]

Historically, the GPU powered the first AI attempts.

https://people.idsia.ch/~juergen/

[...] In 2010, his lab's fast and deep feedforward NNs on GPUs greatly outperformed previous methods, [...]

EDIT:

Accelerating Artificial Intelligence: The Role of GPUs in Deep Learning and Computational Advancements

 

[sbrothy]

I read somewhere that Michael Bay was the first person to fry a GPU on a super computer during a CGI computing. It would probably have burned out anyway though. But it's typical of Michael Bay and the amount of CGI he uses in his films.

 

[Filip Larsen]

What are the operations of a GPU?

It may also be instructive for you (and others) to look into how one would use a GPU to speed up calculations, for example by writing a so-called kernel using CUDA, OpenCL, OpenGL or similar framework. An example on the OpenCL wiki page show how a simple matrix multiplication can be done. Another example, the original purpose of GPU's, is to speed up the rendering of on-screen computer graphics, e.g. rendering a 3D scene in real time.

Typically the GPU are setup up by (application) code to repeat the same set of processing pipelines on data that changes. An example here could be that if a pipeline can process an image in some way (e.g. contrast enhance), then this pipeline can be used to process a sequence of images (i.e. video), possibly in real time.

And just in case you didn't notice, the why all this is done is to get a massive performance boost compared to letting the CPU carry out the calculations, even when that CPU has a lot of cores and there are overhead associated with setting up the GPU and transfer data to it. So usually, as long as you need to process a lot of data in parallel, the GPU wins performance-wise.

 

[sbrothy]

Isn't OpenGL technically obsolete by now?

EDIT: Nah, sorry. Development has slowed and it is being superceded by Vulkan.

EDIT2: As they say in biology: functionally extinct.

 

[sbrothy]

Also for someone who've tried both OpenGL and DirectX I can't say I'm sorry.

 

[sbrothy]

It may also be instructive for you (and others) to look into how one would use a GPU to speed up calculations, for example by writing a so-called kernel using CUDA, OpenCL, OpenGL or similar framework. An example on the OpenCL wiki page show how a simple matrix multiplication can be done. Another example, the original purpose of GPU's, is to speed up the rendering of on-screen computer graphics, e.g. rendering a 3D scene in real time.

Typically the GPU are setup up by (application) code to repeat the same set of processing pipelines on data that changes. An example here could be that if a pipeline can process an image in some way (e.g. contrast enhance), then this pipeline can be used to process a sequence of images (i.e. video), possibly in real time.

And just in case you didn't notice, the why all this is done is to get a massive performance boost compared to letting the CPU carry out the calculations, even when that CPU has a lot of cores and there are overhead associated with setting up the GPU and transfer data to it. So usually, as long as you need to process a lot of data in parallel, the GPU wins performance-wise.

Yes. obviously a supercomputer is basically just a ton of GPUs working together. In parallel or serial I don't know. I'd suspect the former.