Modern GPU Architecture Third Edition

Modern GPU Architecture — Volume 2: Acceleration, Integration, and Physical Design

A modern GPU is no longer just a graphics processor.

It is a compute engine, AI accelerator, ray-tracing machine, video processor, interconnect fabric, synchronization system, thermal design problem, and silicon product — all in one.

Volume 1 showed how the graphics pipeline becomes hardware.

Volume 2 shows how that hardware becomes a complete modern GPU.

This volume moves beyond the geometry-to-pixels pipeline and into the accelerators, system engines, performance limits, and physical-design realities that define modern GPU architecture.

If you want to understand compute dispatch, tensor cores, ray-tracing units, synchronization, memory ordering, advanced rendering, video engines, interconnects, performance counters, timing closure, packaging, and emerging accelerator design, this book connects them into one coherent system.

Inside Volume 2, you will learn:

• How GPU compute architecture launches and schedules massively parallel kernels

• How occupancy, registers, shared memory, warp slots, and barriers limit resident work

• How tensor and matrix accelerators use tiling, systolic datapaths, mixed precision, and sparsity

• How ray-tracing hardware performs BVH traversal, primitive testing, ray scheduling, and shader integration

• How barriers, fences, atomics, cache coherence, and memory ordering preserve correctness

• How mesh shaders, VRS, deferred rendering, display engines, video engines, and interconnects fit into the larger GPU

• How counters, roofline limits, workload signatures, power, and thermals shape performance

• How floorplanning, synthesis, timing closure, place-and-route, DFT, packaging, and chiplet integration turn architecture into silicon

• Where GPU architecture is heading as AI accelerators, new memory systems, and post-Moore scaling reshape the field

This book is for readers who want the layer beneath the marketing terms.

Not just “tensor cores are fast,” but how matrix acceleration changes datapaths, precision, memory layout, and scheduling.

Not just “ray tracing uses RT cores,” but how traversal, primitive intersections, coherence, sorting, and shader handoff fit into hardware.

Not just “GPUs are massively parallel,” but how synchronization, resource allocation, interconnect congestion, and thermal limits decide whether that parallelism survives.

This book is for you if you are:

• A hardware or computer-engineering student moving beyond CPU architecture

• An FPGA, RTL, or ASIC learner studying GPU-like design

• A graphics programmer who wants to understand the silicon behind modern rendering

• A CUDA, HPC, or AI performance reader who wants the hardware layer below kernels and libraries

• An early-career engineer who wants a map of how GPU blocks connect, scale, and become manufacturable

Most books explain one slice of the GPU.

This volume connects the slices.

Compute, tensor acceleration, ray tracing, synchronization, rendering, video, interconnect, performance, physical design, and future accelerator trends are treated as parts of one machine — not isolated buzzwords.

If Volume 1 is the graphics pipeline foundation, Volume 2 is the complete systems view.

Read it to understand how acceleration, integration, and physical design turn modern GPU architecture into real silicon.