Accelerating Python on GPUs

Accelerate your existing NumPy and pandas code with massive speedups on GPUs. Learn about drop-in replacement libraries that require minimal code changes.

#1about 1 minute

The evolution of GPU programming with Python

Python has become a first-class citizen in the CUDA ecosystem, making it easier to accelerate software on GPUs.

#2about 2 minutes

How GPUs evolved from graphics to AI powerhouses

The development of CUDA unlocked general-purpose GPU computing, which was supercharged by the AlexNet breakthrough in AI.

#3about 2 minutes

Understanding modern GPU architecture for parallelism

A look inside a modern data center GPU reveals thousands of cores and specialized hardware like Tensor Cores designed for massive parallelism.

#4about 2 minutes

Navigating the CUDA Python software ecosystem

The CUDA platform provides a layered stack of libraries, frameworks, and tools to access GPU power at your preferred level of abstraction.

#5about 3 minutes

Using high-level frameworks like Rapids for acceleration

Frameworks like Rapids provide GPU-accelerated versions of tools like pandas and scikit-learn, often requiring zero code changes for massive speedups.

#6about 1 minute

Using CuPy as a drop-in replacement for NumPy

CuPy offers a familiar NumPy-like API that allows you to move array computations to the GPU by simply changing the import statement.

#7about 5 minutes

Optimizing code with nvmath-python and a case study

The nvmath-python library enables kernel fusion for significant speedups, as demonstrated by a supernova detection project that went from 45 minutes to one minute.

#8about 2 minutes

A look at upcoming Python GPU programming tools

New tools like CuTe for array-based programming and Python bindings for CUDA Core Compute Libraries are making GPU development even more accessible.

#9about 2 minutes

Strategies for scaling your code to multiple GPUs

Explore various approaches for multi-GPU programming, from high-level libraries like Dask and JAX to lower-level communication libraries like NCCL and NVSHMEM.

#10about 2 minutes