Modal GPU Glossary Maps the Full GPU Stack in Four Chapters
The takeaway is that GPU literacy for AI engineers can be structured as one coherent hierarchy — threads, memory and hardware map onto each other layer by layer — rather than scattered vendor docs. Practitioners writing or tuning CUDA kernels for inference and training workloads can use it as both a reference and a linear course. It is a free web resource from Modal, an infrastructure provider, and reads more accessibly than the CUDA C++ Programming Guide.
Modal has published a GPU Glossary at modal.com/gpu-glossary that organises the entire GPU technology stack into four cross-linked chapters: device hardware, device software (the CUDA execution model), host software (drivers and toolchains), and performance profiling.
The glossary starts from CUDA and explains its three meanings: a device architecture, a parallel programming model, and a software platform.
Entries are densely cross-linked, so readers can look up single terms on demand or read the whole document linearly.
The profiling chapter lays out a diagnostic chain from bottleneck identification through roofline analysis, utilisation, warp mechanics, memory access patterns and resource constraints.
Modal @modal 这份「GPU 术语手册」做的太友好了，它把整个 GPU 技术栈拆成四个相互勾连的层面：设备硬件、设备软件（执行模型）、主机软件（驱动与工具链）、性能分析，每个词条都密集交叉链接，既可以按需查阅单个术语，也可以从头到尾线性通读。 理解这份手册从「CUDA」开始，CUDA 有三重含义：一种设备架构、一个并行编程模型、一个软件平台。这份手册的四个章节恰好分别展开这三个含义加上性能分析。 GPU Glossary 四个章节： 1. 设备硬件：GPU 的物理解剖 - CPU 用复杂核心避免延迟，GPU 用海量简单线程隐藏延迟 2. 设备软件：CUDA 的执行模型 - 线程层级、内存层级、硬件三层严格同构，程序员亲手管理内存 3. 主机软件：驱动、工具链与库 - Runtime 包着 Driver，闭源库拿来即用，CUTLASS 一系负责构造自己的 kernel 4. 性能分析：完整诊断链 - 定瓶颈 → roofline 定性 → 利用率定位 → warp 机制 → 访存模式 → 资源约束 这三句总结很到位 GPU 的本质是“单周期切换海量简单线程来隐藏延迟”；编程的本质是“沿线程-内存-硬件的同构层级，把数据在正确的层级间搬运，把算术强度做到岭点以上”；性能工程的本质是“沿诊断链逐层定位，用 Little 定律算够并发，不为中间指标（如占用率）本身而优化。 modal.com/gpu-glossary📊 2 ⚡
This report is categorized under AITOP's "How-to" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Modal, CUDA, GPU, performance profiling.
Source attribution: curated from shao__meng (@shao__meng).
