mdnax.blogg.se - Parallel computing in pix4dmapper

no conditional processing like IF, CASE, GOTO etc +9 more terms.

=> intelligent spars processing needs to work on list of non-zero values but not dense matrixes => no support for sparse/not dense matrixes or vectors

full speed up for multidimensional tensors operations but plain vector operations cannot exploit parallelism.

=> data should stay in the GPU memory as long as possible (faster processing => data locality)

every bit of data has to be transferred from main memory to GPU memory and back.

GPU: Ups and Downs GPUs are fast when processing functions which can be expressed by matrix/tensor operation but are not generally faster: ** tensor = algebraic object that describes a (multilinear) relationship between sets of algebraic objects related to a vector space. * Gradient (Steigung) descent is a first-order iterative optimization algorithm for finding a local minimum of a differentiable function. Note: GPU support parallel computing of mathematical operations on tensors which is not specific to Deep Learning ⇒computations on tensors** can be done order of magnitude faster than on a CPU

Gradient Descent* on Deep Neural Networks can be mapped easily to these APIs (all necessary operations are available).

GPU vendors provided API for computing on the GPU (CUDA, OpenGL.

** PCIe = peripheral component interconnect express is an interface standard for connecting high-speed components. * sigmoid = The term "sigmoid" means S-shaped, and it is also known as a squashing function, as it maps the whole real range of z into in the g(z). dedicated HPC (High Performance Computing) cards might not have monitor ports anymore component original handling the video IO /currently connected via PCIe** bus to the main board-Cards have their own memory up to multiple GPUs (transfer main memory to GPU is necessary) recently more and more native support for ML functions (e.g. less functionality than a CPU but much more parallelism ⇒rendering can be parallelized since applying a linear transformations to points is independent from each otherĭedicated processor providing parallel mathematical operations on tensor data. => rendering the next frame applies update on each point of multiple polygon meshes => a 3D scene is a collection polygon meshes

rendering 3D scenes in real-time exploits parallel processing.

=>fast computation requires quick access to graphics memory

=> translation, scaling, rotation of 3D points can be expressed by matrix multiplications

3D rendering is based on algebraic tensor operations:.

"muscle" behind a CPU if processing power of CPU is not sufficient.

originally designed to support 3D rendering in CAD, graphic design and computer games.

TERMS IN THIS SET (13) Graphical Processing Units (GPU)