Most CUDA programmers never see the real program their GPU actually runs.
They write CUDA C++. They launch kernels. They profile. They tune block sizes, adjust memory access, stare at Nsight reports, and hope the compiler has done what they think it has done.
But the truth is lower down.
The truth is in PTX. The truth is in SASS.
Mastering PTX and SASS is for CUDA programmers, GPU performance engineers, ML systems developers, compiler-minded programmers, and high-performance computing specialists who want to understand NVIDIA GPU execution below the source level.
This is not a beginner CUDA book.
It is for readers who already understand the basic GPU programming model and now want to inspect the layer where performance is really exposed: instruction streams, register allocation, memory transactions, predicate logic, compiler output, synchronization, tensor pipelines, and architecture-specific machine code.
Inside, you will learn how to:
- Understand how CUDA source becomes PTX, and how PTX becomes SASS
- Use PTX as a portable low-level view of compiler intent
- Use SASS as evidence of what the GPU actually executes
- Interpret register pressure, spills, occupancy, instruction scheduling, and memory traffic
- Diagnose coalescing problems, bank conflicts, branch divergence, dependency stalls, and synchronization costs
- Work with tensor cores, MMA, WGMMA, TMA, low-precision formats, and modern NVIDIA architecture features
- Use nvcc, ptxas, cuobjdump, nvdisasm, Nsight Compute, Nsight Systems, and Compute Sanitizer as practical engineering tools
- Decide when CUDA C++, compiler flags, intrinsics, inline PTX, full PTX, or SASS inspection is the right level to touch
The goal is not assembly for its own sake.
The goal is diagnosis.
The goal is control.
The goal is to look at a slow kernel and know whether the limiting factor is memory movement, instruction throughput, register pressure, tensor-pipeline starvation, branch behavior, synchronization overhead, launch cost, or compiler transformation.
Modern GPU performance is not won by writing code that merely runs.
It is won by understanding how the machine schedules work, moves data, allocates registers, forms instructions, hides latency, feeds tensor units, and exposes bottlenecks through measurable evidence.
This book is for you if you already write CUDA code and want to understand what happens after compilation.
If you are still learning what a thread block is or how kernels launch, start with an introductory CUDA book first.
But if you want to read below the source level, connect profiler symptoms to architectural causes, and move closer to the real hardware ceiling, Mastering PTX and SASS was written for you.
Correctness is only the beginning.
The real question is: how close can you get to the hardware ceiling?
