Welcome to the official website for The CUDA Handbook
The CUDA Handbook
is a comprehensive guide to programming GPUs with CUDA. It covers every detail about CUDA, from system
architecture, address spaces, machine instructions and warp synchrony to
the CUDA runtime and driver API to key algorithms such as reduction, parallel prefix sum (scan) , and N-body.
The CUDA Handbook
will have a couple different aspects:
- Detailed descriptions of every CUDA abstraction and how it maps onto the hardware.
to measure performance of everything from peak global memory bandwidth.
- Microdemos to illustrate in-depth the workings of features such as texture-mapping and the new warp shuffle instruction (new in SM 3.0).
- Detailed descriptions of key parallel algorithms, such as reduction, Scan, and N-body, with source code implementations.
- It covers CUDA in the Cloud, focusing on Amazon's CUDA-capable EC2 instance type, cg1.4xlarge.
The source code accompanying The CUDA Handbook
is open source, available on github
Check out The CUDA Handbook blog
Like The CUDA Handbook
Follow The CUDA Handbook on Twitter (@CUDAHandbook
Nicholas Wilt was on the inception team for CUDA, starting in 2005 and continuing work on it until he left NVIDIA in late 2010. As the primary architect of CUDA's driver API, he is responsible for most of the low-level abstractions that CUDA programs deal with: from contexts and modules to streams and events, portable and mapped pinned memory, CUDA arrays and texture and surface references.
Before joining NVIDIA in 2002, Mr. Wilt worked on multimedia at Microsoft in various roles. Among other things, he was the development lead for Direct3D 5.0 and Direct3D 6.0, worked on the Video Mixing Renderer (VMR) that shipped in Windows XP, and built a prototype page-flipping desktop that eventually became the Windows Desktop Manager.