⚡CUDA

NEW

Also known as: Compute Unified Device Architecture

NVIDIA’s parallel computing platform and programming model enabling general-purpose GPU acceleration.

DESCRIPTION

CUDA provides APIs, languages, and libraries to write kernels that execute in parallel on NVIDIA GPUs. It supports fine-grained thread hierarchies, shared memory, and warp-level primitives. A rich ecosystem includes cuDNN, cuBLAS, Thrust, and NVRTC. CUDA underpins modern deep learning training and inference workflows. Tooling spans profiling, debugging, and deployment across desktops, data centers, and embedded devices.

ALTERNATIVE DEFINITIONS

Wikipedia

CUDA is a parallel computing platform and programming model developed by Nvidia for general computing on its own GPUs.

General-purpose computing emphasis on NVIDIA GPUs.

Source

NVIDIA

CUDA is a parallel computing platform and programming model developed by NVIDIA for general computing on GPUs.

Official vendor definition highlighting platform and model aspects.

Source

NVIDIA Documentation

CUDA is a general purpose parallel computing platform and programming model.

Documentation phrasing used in the CUDA C Programming Guide.

Source

RELATED TERMS

Graphics Processing Unit Parallel Computing CUDA C/C++ PTX cuDNN cuBLAS NCCL Tensor Cores

COMPETING TERMS

OpenCL ROCm

RELATIONS

• Broader: Parallel Computing

KNOWLEDGE AREAS

Infrastructure AI Software Engineering Data Science

TARGET AUDIENCE

AI Engineer Software Developer MLOps Engineer Cloud Architect Data Scientist

EXTERNAL SOURCES

• Wikipedia - CUDA
• NVIDIA - Official CUDA documentation hub
• NVIDIA Developer Blog - Best practices and performance guides
• NVIDIA cuDNN - Deep learning primitives on CUDA

TOP QUESTIONS

Why use CUDA for AI workloads?

CUDA exposes massive parallelism and optimized libraries, enabling significant speedups for deep learning training and inference on NVIDIA GPUs.

Category: Relevancy

How does CUDA differ from OpenCL?

CUDA is NVIDIA-specific with mature AI libraries; OpenCL is vendor-neutral but often lags in ecosystem support for deep learning.

Category: Comparison

What is a CUDA kernel?

A kernel is a function executed in parallel by many GPU threads, organized into blocks and grids.

Category: Understanding

What are common performance pitfalls in CUDA?

Uncoalesced memory access, excessive divergence within warps, and poor occupancy often degrade performance.

Category: Mistakes

How do I profile CUDA code?

Use Nsight Systems/Compute and CUDA profiling APIs to analyze kernels, memory throughput, and warp efficiency.

Category: HowTo

Can CUDA run on non-NVIDIA GPUs?

No. CUDA targets NVIDIA GPUs. Alternatives like OpenCL or ROCm support other vendors.

Category: Limitations

What is PTX in CUDA?

PTX is an intermediate assembly-like language for NVIDIA GPUs that CUDA compiles to before device-specific code generation.

Category: Understanding

How does mixed precision training work with CUDA?

CUDA libraries and tensor cores accelerate FP16/BF16 operations with loss scaling to maintain numerical stability.

Category: Technique

What role does NCCL play in multi-GPU CUDA?

NCCL provides high-performance collective communication primitives essential for distributed training across GPUs and nodes.

Category: Tools