NVIDIA's evolution from a graphics hardware company to a leader in general-purpose computing was accelerated by the use of GPUs for AI with models like AlexNet.
As single-threaded CPU performance plateaued, GPUs offered a path forward with their massively parallel architecture designed for simultaneous computation.
GPUs work with CPUs by offloading compute-intensive code and use thousands of threads to hide memory latency, leveraging streaming multiprocessors and high-bandwidth memory.
The CUDA platform is a complete ecosystem with compilers, libraries, and frameworks that enables developers to program GPUs using various languages and abstraction levels.
Specialized hardware like Tensor Cores can be used transparently through high-level libraries like cuDNN or programmed directly with low-level APIs for maximum performance.
Frameworks like Rapids provide GPU-accelerated, drop-in replacements for popular data science libraries such as Pandas (cuDF) and NetworkX (cuGraph) with minimal code changes.
Developers can choose from a spectrum of Python libraries, from simple drop-in replacements like CuNumeric and CuPy to JIT compilers like Numba and direct kernel programming with PyCUDA.
NVIDIA offers a comprehensive suite of developer tools for profiling and debugging, along with learning resources like the NGC repository, DLI courses, and community events.
21:02Andy Terrel
22:18Paul Graham
24:39Paul Graham
33:44Fabian Schindler
57:49Jodie Burchell
46:30Anirudh Koul
39:04Bobur Umurzokov
57:46Adrian Schmitt
Jobs that call for the skills explored in this talk.
CONTIAMO GMBH
Berlin, Germany
MedAustron EBG
CloudStream
London, United Kingdom
ETH Zürich
Zürich, Switzerland