Rust for HPC + GPU Bundle

If you're a Python pro looking to get the most out of your code with GPUs, then Practical GPU Programming is the right book for you. This book will walk you through the basics of GPU architectures, show you hands-on parallel programming techniques, and give you the know-how to confidently speed up real workloads in data processing, analytics, and engineering.

The first thing you'll do is set up the environment, install CUDA, and get a handle on using Python libraries like PyCUDA and CuPy. You'll then dive into memory management, kernel execution, and parallel patterns like reductions and histogram computations. Then, we'll dive into sorting and search techniques, but with a focus on how GPU acceleration transforms business data processing. We'll also put a strong emphasis on linear algebra to show you how to supercharge classic vector and matrix operations with cuBLAS and CuPy. Plus, with batched computations, efficient broadcasting, custom kernels, and mixed-library workflows, you can tackle both standard and advanced problems with ease.

Throughout, we evaluate numerical accuracy and performance side by side, so you can understand both the strengths and limitations of GPU-based solutions. The book covers nearly every essential skill and modern toolkit for practical GPU programming, but it's not going to turn you into a master overnight.

Key Learnings

• Boost processing speed and efficiency for data-intensive tasks.
• Use CuPy and PyCUDA to write and execute custom CUDA kernels.
• Maximize GPU occupancy and throughput efficiency by using optimal thread block and grid configuration.
• Reduce global memory bottlenecks in kernels by using shared memory and coalesced access patterns.
• Perform dynamic kernel compilation to ensure tailored performance.
• Use CuPy to carry out custom, high-speed elementwise GPU operations and expressions.
• Implement bitonic and radix sort algorithms for large or batch integer datasets.
• Execute parallel linear search kernels to detect patterns rapidly.
• Scale matrix operations using Batched GEMM and high-level cuBLAS routines.

Table of Content

1. Introduction to GPU Fundamentals
2. Setting up GPU Programming Environment
3. Basic Data Transfers and Memory Types
4. Simple Parallel Patterns
5. Introduction to Kernel Optimization
6. Working with PyCUDA and CuPy Features
7. Practical Sorting and Search
8. Linear Algebra Essentials on GPU

Whether you're just starting out or a fully-fledged Rust pro, this third Edition is a goldmine of knowledge about Rust. This book has been updated with Rust 1.85 and Rust 2024 Edition, and it takes you on a really structured learning path to write production-grade systems, async services, GPU compute kernels, Linux kernel modules, and Go-integrated libraries.

The book is built around a single, progressive application, and every chapter adds a functional layer to a real Linux system-metrics tool. The knowledge you gain from this book, is working and evolving codes that you can study, extend, and own. You'll get the hang of ownership and borrowing, handle the borrow checker, build concurrent and async programs with Tokio, manage memory without a garbage collector, and use Rust in domains that were once reserved exclusively for C and C++.

Whatever stage you're at, whether you're just starting out and writing your first function or you're a seasoned engineer migrating a critical service to Rust, this strong practically focussed book has got you covered.

• Dive in-depth into ownership model and eliminate memory bugs at compile time.
• Build async, concurrent programs using Tokio without data races or deadlocks.
• Design production-grade error types with ‘thiserror’ and ‘anyhow’ crates.
• Write GPU compute kernels and monitor GPU metrics.
• Integrate Rust libraries into existing Go applications via application binary interfaces.
• Write a loadable Linux kernel module entirely.
• Manage heap memory using Box, Rc, Arc, and smart pointer types.
• Leverage Rust 2024 Edition's new safety rules and migration workflow.
• Make use of Cargo workspaces, features, and optional dependencies to structure projects.

1. Why Rust in 2025?
2. Setting up Rust 1.85 Development Environment
3. Core Language Essentials
4. Structs and Data Modeling
5. Enums, Pattern Matching, and Error Handling
6. Ownership, Borrowing, Lifetimes, and Raw Pointers
7. Packages, Crates, and Full Command Suite
8. Rust Standard Library
9. Smart Pointers and Memory Management
10. Concurrency, Parallelism, and Async Programming
11. Rust 2024 Edition
12. Rust for GPU and CUDA Programming
13. Rust for Go Developers
14. Rust for Linux Kernel

This machine learning book is a new edition out now for Rust developers to build practical machine learning systems without getting bogged down in the complexities of Rust 1.85. We put together a single workspace over eleven chapters, building it from a Polars-based data pipeline through classical models, deep neural networks, and natural language processing, right up to a deployed REST API.

We teach you to make use of linfa and smartcore crates for regression, classification, ensemble methods, and support vector machines. We build and train neural networks using the Burn framework, work with convolutional architectures on image data, and load pre-trained transformer models through Candle. We also use a lightweight NLP pipeline with TF-IDF and Gaussian Naive Bayes from first principles. At all times, Rust's ownership model and type system work together to make sure things are correct, rather than getting in the way.

This particular book is a perfect knowledge source for developers who already know Rust at a beginner’s level and want to use that knowledge for machine learning tasks. You don't need to have worked with any ML frameworks before.

• Structure a multi-crate Rust workspace for end-to-end machine learning.
• Build type-safe data pipelines using Polars and Apache Arrow.
• Train, evaluate, and improve classical models with linfa and smartcore.
• Implement backpropagation and mini-batch gradient descent.
• Train multilayer perceptrons and convolutional networks using Burn.
• Load and fine-tune pre-trained BERT models using Candle.
• Apply LoRA weight adaptation to transformer layers.
• Construct production TF-IDF and Naive Bayes NLP pipeline.
• Serve trained models using axum and tokio.
• Implement live model hot-reloading and request batching without server downtime.

1. Why Rust for Machine Learning?
2. Data Engineering with Polars and Arrow
3. Regression and Classification
4. Decision Trees, Random Forests, and Gradient Boosting
5. SVMs, Naive Bayes, and k-NN
6. Neural Networks from First Principles
7. Deep Learning with Burn
8. Computer Vision Pipelines with Burn
9. Transformer Inference and Fine-Tuning with Candle
10. Natural Language Processing in Rust
11. Model Serving and REST APIs

This book is all about getting hands-on with different toolchains that connect Rust to NVIDIA hardware. There's RustaCUDA for safe host-side control, the Rust-CUDA project for writing kernels in pure Rust, and NVIDIA's experimental cuda-oxide compiler with its typed launches and async execution graphs.

We're going to build one Cargo workspace that keeps on growing. It'll include device queries, launch planning, Rust-written kernels, memory optimization, parallel reductions and scans, multi-stream pipelines, matrix multiplication benchmarked against cuBLAS, a Monte Carlo option pricer validated against a closed formula, and a complete batched inference application measured against a Python baseline. We'll check every result against a CPU reference, and the reports will give accurate numbers, including where libraries outperform hand-written kernels and where experimental toolchains are still a work in progress.

• Launch, synchronize, and verify GPU kernels with ownership-managed device memory.
• Write real CUDA kernels using Rust-CUDA and cuda-oxide.
• Plan grids, blocks, and warps for 2D workloads.
• Accelerate transfer speeds with pinned memory and coalesced access patterns.
• Build race-free thread cooperation using shared memory, barriers, and atomics.
• Overlap transfers with computation using streams, events, and async Rust pipelines.
• Optimize matrix multiplication and benchmark against cuBLAS ceiling.
• Wrap CUDA C library safely with handles, error enums, and Drop.
• Ship complete batched GPU inference application against Python baselines.
• Diagnose performance with Nsight Systems, Nsight Compute, and compute-sanitizer.

1. New Beneficiary of GPU Computing
2. Thinking in Threads
3. Commanding GPU
4. Writing GPU Kernels
5. Cleaner Kernels with cuda-oxide
6. Mastering GPU Memory
7. Making Threads Cooperate
8. Keeping GPU Busy
9. Delivering Real Math
10. Borrowing NVIDIA's Muscle
11. Shipping Complete GPU Application
12. Proving Performance