## Go on GPU ## Changkun Ou changkun.de/s/gogpu GopherChina 2023 Session “Foundational Toolchains” 2023 June 10 ## Agenda - Basic knowledge for interacting with GPUs • Accelerate Go programs using • Conclusion and outlooks ## Agenda - Basic knowledge for interacting with GPUs o Motivation o GPU Driver and Standards Render and compute pipeline o Vulkan/Metal/DX12/OpenGL ☐ Accelerate Go programs programs using GPUs ☐ Challenges in Go when using GPUs ☐ Conclusion and outlooks ## Motivation of GPU Acceleration Improve system computation performance Increase amount of concurrency Processing large

## GPU Resource Management On JDOS 梁永清 liangyongqing1@jd.com ## 提供的服务 ## Experiment ## Training 1. 用于实验的 GPU 容器 2. 基于 Kubeflow 的机器学习训练服务 3. 模型管理和模型 Serving 服务 ## Serving 均基于容器，不对业务方直接提供 GPU 物理机 物理机 ## GPU 实验 JDOS 常规的容器服务， 使用 gpu 的 zone，自行设定相应的镜像即可， 有完善的周边服务 我的系统  三一键编译  public/tensor/now.1.4.1-ueve-gpu-vi  public/tensorflow:1.7.0-devel-gpu-py3-v1 ![Image](/uploads/documents/8/

CPU and GPU ## THOMAS MEJSTRIK ## DIMETOR  FWF ## Bridging the Gap: Writing Portable Programs for CPU and GPU using CUDA ROCm, Vulkan, ... ☐ You can tell me about afterwards ## Why write programs for CPU and GPU ## ☐ Difference CPU/GPU Algorithms are designed differently ☐ Latency/Throughput ☐ Memory bandwidth ☐ Number Problem ☐ Why it makes sense? ☐ Scope of the talk ## Why write programs for CPU and GPU ## ☐ Difference CPU/GPU ☐ Why it makes sense? Library/Framework developers ☐ Embarrassingly parallel algorithms

Issue: How to train MLP Chain Rules => Backpropagation  ## Backpropagation: First Spark - Derived in early 60's [1389.D4] [1635.52] [1897.128] ## Now theoretically solved:1989 Communicated by Dana Ballard # Backpropagation Applied to Handwritten Zip Code Recognition Y. LeCun B. Boser J. S. Denker D. Henderson discovery and genomics. Deep learning discovers intricate structure in large data sets by using the backpropagation algorithm to indicate how a machine should change its internal parameters that are used to compute

## FFmpeg在Intel GPU上的硬件加速与优化 赵军 DCG/NPG @ Intel ## 介绍FFmpeg VAAPI • Media pipeline review • 何谓FFmpeg VAAPI • 为什么我们需要FFmpeg VAAPI • 当前状态 • 更进一步的计划 · 附录 ## 典型的 media pipeline SOURCE libavformat radeon, nouveau (?), freedreno, … • 废弃的 API bridges • vdpau—va bridge • powervr—va bridge ## I ntel GPU简介 ## • Gfx Label • Gen3: Pinetrail (Pineview) • Gen4: G965 • Gen5: G4X, Ironlake (Piketon, Calpella) OpenGL） ## I ntel GPU media 硬件编程模型 GPU  ## FFmpeg & Intel GPU加速方案 - FFmpeg 作为最流行的开源多媒体框架; 集成Intel的GPU的硬件加速能为用户带来更多收益

## PyTorch ## 激活函数与GPU加速 主讲人：龙良曲    ## GPU accelerated ## ☐ ☐ ☐ device = torch.device('cuda:0') net = MLP().to(device) optimizer =

## CUDA 开启的 GPU 编程 by 彭于斌 (@archibate) 往期录播：https://www.bilibili.com/video/BV1fa411r7zp 课程 PPT 和代码：https://github.com/parallel101/course ## 前置条件 • 学过 C/C++ 语言编程。 - 理解 malloc/free 之类的概念。 • 熟悉 STL ## 编写一段在 GPU 上运行的代码 - 定义函数 kernel，前面加上 ___ global___ 修饰符，即可让他在 GPU 上执行。 - 不过调用 kernel 时，不能直接 kernel()，而是要用 kernel<<1, 1>>() 这样的三重尖括号语法。为什么？这里面的两个 1 有什么用？稍后会说明。 • 运行以后，就会在 GPU 上执行 printf kernel 函数在 GPU 上执行，称为核函数，用 ___ global___ 修饰的就是核函数。  ## 没有反应？同步一下！ - 然而如果直接编译运行刚刚那段代码，是不会打印出Hello, world! 的。 这是因为 GPU 和 CPU 之间的通信，为了高效，是异步的。也就是

learning tasks that run on real life devices like embedded devices, phone, etc, do not have access to GPU. To bridge that gap, micrograd++ let's any user train a neural network in C++ and ship that to and neurons, enabling users to construct complex network architectures. - Backpropagation: The implementation of backpropagation in micrograd++ allows for efficient training of models through gradient descent with web assembly to make it work on the web on client side. - Optional GPU support: Make microgradpp compatible with modern GPU frameworks. - CI/CD Pipeline: Establishing a continuous integration and

to(‘cpu’) $$ • GPU $$ x=x.to(‘cuda’) $$ ## Tensors – Device (GPU)  - Check if your computer has NVIDIA GPU torch.cuda • Why use GPUs? Parallel computing with more cores for arithmetic calculations ☐ See What is a GPU and do you need one in deep learning? ## Tensors – Gradient Calculation (1) >> x = torch.tensor([[1 to device (cpu/cuda) forward pass (compute output) compute loss compute gradient (backpropagation) update model with optimizer ## Neural Network Validation Loop model.eval() total_loss

machine learning algorithms over the past two decades. Stochastic Gradient Descent (SGD) and Backpropagation were the well-known algorithms designed for training deep networks. However, one of the critical are designed for Training Efficiency would help. For example, if model A takes 100 GPU hours, while model B takes 5 GPU hours, it might be worth preferring model B if training efficiency is a more important Advances in hardware are significantly responsible for the deep learning revolution, specifically the GPU (Graphics Processing Unit), since they made it possible to train deep models many times faster than