ARM 芯片的 Python + AI 算力优化 主讲人： 朱宏林 – 阿里云程序语言与编译器团队 简介 • 当今开发者们大量使用 Python 语言编写的 AI 程序。过去这些程序总跑在 GPU 或者 x86 架构的 CPU 上。然而综合考虑到功耗、成本、性能等因素，云厂商们开始建设 ARM 架构的服务平台，如 何整合 Python + AI 的相关软件并使其在该平台上发挥最高的性能成为了工程师们关注的焦点。 提升明显。目前，该成果已经被集成进 OpenBLAS 和 PyTorch 中。 • 本次演讲，将向大家介绍我们在倚天 710 ARM 芯片上开展的 Python + AI 优化工作，以及在 ARM 云平台上部署 Python + AI 任务的最佳实践。 深度学习 • 广泛使用的深度学习框架 • TensorFlow、PyTorch • 结合硬件（ARM 服务端芯片） • 倚天 710 • exec -ti torch_bm bash -c "cd /tmp/resnet50 && python3 performance.py” • docker exec -ti torch_bm bash -c "cd /tmp/resnet50 && python3 performance.py --bf16" 首次执行会下载预训练模型 输出执行耗时 PyTorch BF16 加速演示 •

Domain-specific Knowledge Graph Construction Preliminaries Preliminaries AI system = Knowledge + Reasoning Preliminaries How dose the AI system works? Q: 1M = ? B 1024 Q: Where was David Beckham born? （机器的潜台词：“我”会推理，so easy ！）。 所以，通俗的来说，在AI system中：要么从原有的知识体系中直接提取信息来使用，要 么进行推理。 将知识融合在机器中，使机器能够利用我们人类知识、专家知识解决问题，这就是早期 知识工程（Knowledge Engineering）的核心内涵。 Preliminaries Explaining AI system from the perspective perspective of KE – Symbolism AI System = Knowledge + Reasoning  符号主义的主要观点 • 认知即计算 • 知识是信息的一种形式，是构成智能的基础 • 知识表示、知识推理、知识运用是人工智能的核心  Physical Symbol System • A physical symbol system has the necessary and

threads are not appropriate. Platforms: Tornado is designed for Unix-like platforms, with best performance and scalability on systems supporting epoll (Linux), kqueue (BSD/macOS), or /dev/poll (Solaris) a way that is transparent to its callers (systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous). Asynchronous faking a root apple-touch-icon.png by using the appropriate tag in your HTML. To improve performance, it is generally a good idea for browsers to cache static resources aggressively so browsers won’t

not appropriate. Platforms: Tornado should run on any Unix-like platform, although for the best performance and scalability only Linux (with epoll) and BSD (with kqueue) are recommended for production deployment deployment (even though Mac OS X is derived from BSD and supports kqueue, its networking performance is generally poor so it is recommended only for development use). Tornado will also run on Windows, although a way that is transparent to its callers (systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous). Asynchronous

needed in Python 3. Platforms: Tornado should run on any Unix-like platform, although for the best performance and scalability only Linux (with epoll) and BSD (with kqueue) are recommended for production deployment deployment (even though Mac OS X is derived from BSD and supports kqueue, its networking performance is generally poor so it is recommended only for development use). Tornado will also run on Windows, although a way that is transparent to its callers (systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous). Asynchronous

in Python 3.3. Platforms: Tornado should run on any Unix-like platform, although for the best performance and scalability only Linux (with epoll) and BSD (with kqueue) are recommended for production deployment deployment (even though Mac OS X is derived from BSD and supports kqueue, its networking performance is generally poor so it is recommended only for development use). Tornado will also run on Windows, although to its callers (systems like gevent [http://www.gevent.org] use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous). Examples

in Python 3.3. Platforms: Tornado should run on any Unix-like platform, although for the best performance and scalability only Linux (with epoll) and BSD (with kqueue) are recommended for production deployment deployment (even though Mac OS X is derived from BSD and supports kqueue, its networking performance is generally poor so it is recommended only for development use). Tornado will also run on Windows, although a way that is transparent to its callers (systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous). Examples

threads are not appropriate. Platforms: Tornado is designed for Unix-like platforms, with best performance and scalability on systems supporting epoll (Linux), kqueue (BSD/macOS), or /dev/poll (Solaris) a way that is transparent to its callers (systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous). Asynchronous faking a root apple-touch-icon.png by using the appropriate tag in your HTML. To improve performance, it is generally a good idea for browsers to cache static resources aggressively so browsers won’t

threads are not appropriate. Platforms: Tornado is designed for Unix-like platforms, with best performance and scalability on systems supporting epoll (Linux), kqueue (BSD/macOS), or /dev/poll (Solaris) a way that is transparent to its callers (systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous). Asynchronous faking a root apple-touch-icon.png by using the appropriate tag in your HTML. To improve performance, it is generally a good idea for browsers to cache static resources aggressively so browsers won’t

threads are not appropriate. Platforms: Tornado is designed for Unix-like platforms, with best performance and scalability on systems supporting epoll (Linux), kqueue (BSD/macOS), or /dev/poll (Solaris) a way that is transparent to its callers (systems like gevent use lightweight threads to offer performance comparable to asynchronous systems, but they do not actually make things asynchronous). Asynchronous faking a root apple-touch-icon.png by using the appropriate tag in your HTML. To improve performance, it is generally a good idea for browsers to cache static resources aggressively so browsers won’t