INDUSTRIES, INCORPORATED ACCELERATED VISUAL PERCEPTION LIANGFU CHEN 11/16/2019 DEPLOY VTA ON INTEL FPGA©2019 HARMAN INTERNATIONAL INDUSTRIES, INCORPORATED 2 Moore’s Law is Slowing Down MOTIVATION©2019 Terasic DE10-Nano DEPLOY VTA ON INTEL FPGA©2019 HARMAN INTERNATIONAL INDUSTRIES, INCORPORATED 5 Software - CMA Contiguous Memory Allocation – Linux Kernel DEPLOY VTA ON INTEL FPGA https://pynq.readthedocs INCORPORATED 6 Software - CMA Contiguous Memory Allocation – Linux Kernel Module DEPLOY VTA ON INTEL FPGA Setup Environment Variables Navigate to 3rdparty/cma and build kernel module Copy kernel module

Target-independent Relay passes Target-optimized graph Target-dependent Relay passes Intel x86 ARM CPU Nvidia GPU ARM GPU Schedule templates written in TVM Tensor IR .. More targets AutoTVM – Tuning Target-independent Relay passes Target-optimized Int8 Relay Graph Intel x86 schedule ARM CPU schedule Nvidia GPU schedule ARM GPU schedule Relay Int8 Graph Target-dependent Relay layout opt© 2019 its Affiliates. All rights reserved. Outline • QNN Dialect • Design • Operators • Results on Intel Cascade Lake© 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Quantized Operators

人 人 e 人 e@ TVM Q@ AliOs Overview TVM @ AliOs ARM CPU TVM @ AliOos Hexagon DSP TVM @ Alios Intel GPU Misc /NiiOS ! 驱动万物智能 PART ONE TVM Q@ AliOs Overview AiOS 1驱动万物智能 AliOs overview 。 AliOs (www AN 2X MobilenetV2 TFLite 1.34X MobilenetV2 QNNPACK AliOs @ Roewe RX5 MAX OpenVINO @ Intel GPU AliDS AR-Nav Product @ SUV Release and adopt TVM (Apollo Lake Gold) Model 1.6X Intel AliOs TVM Arch Model 。 Facelandmark Pedestrian & Vehicle Detection Voice-GUI Gesture Lanenet NLU DMS FacelD Multimodal Interection CPU (ARM、Intel) 1驱动万物智能 Accelerated

© 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon/Intel Confidentia Presenter: Zhi Chen, Cody Yu Amazon SageMaker Neo, Deep Engine Science Bring Your Own Codegen to TVM Chip© 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Example showcase: Intel MKL-DNN (DNNL) library 1. Import packages import numpy as np from tvm import relay 2. Load a pretrained Relay Runtime (VM, Graph Runtime, Interpreter) Your Dispatcher Target Device General Devices (CPU/GPU/FPGA) Mark supported operators or subgraphs 1. Implement an operator-level annotator, OR 2. Implement

Impressive61 NVIDIA AI Ecosystem Tells Over Four Years = >100% Growth in Developers / Startups / Apps Note: GPU = Graphics Processing Unit. Source: NVIDIA (2021 & 2025) NVIDIA Computing Ecosystem – 2021-2025, per Cloud vs. AI Patterns105 Tech CapEx Spend Partial Instigator = Material Improvements in GPU PerformanceNVIDIA GPU Performance = +225x Over Eight Years 106 1 GPT-MoE Inference Workload = A type of workload Source: NVIDIA (5/25) Performance of NVIDIA GPU Series Over Time – 2016-2024, per NVIDIA Tech CapEx Spend Partial Instigator = Material Improvements in GPU Performance Pascal Volta Ampere Hopper Blackwell

performance gain by autotuning 3. TVM can support many kinds of hardware platform: Intel/arm CPU, Nividia/arm GPU, VTA…5 �������������� 1. Get a .log file from the autotvm on Ubuntu 2. Use the .log

数 Windows 配置要求 Mac 配置要求 适⽤场景 1.5B - RAM: 4GB - GPU: 集成显卡/现代CPU - 存储: 5GB - 内存: 8GB （M1/M2/M3） - 存储: 5GB 简单⽂本⽣成、基础代 码补全 7B - RAM: 8-10GB - GPU: GTX 1680（4-bit量 化） - 存储: 8GB - 内存: 16GB（M2 Pro/M3） Pro/M3） - 存储: 8GB 中等复杂度问答、代码 调试 14B - RAM: 24GB - GPU: RTX 3090（24GB VRAM） - 存储: 20GB - 内存: 32GB（M3 Max） - 存储: 20GB 复杂推理、技术⽂档⽣ 成 32B+ 企业级部署（需多卡并联） 暂不⽀持 科研计算、⼤规模数据 处理 2. 算⼒需求分析 模型 参数规 模 2*XE9680（16*H20 GPU） DeepSeek-R1-Distill- 70B 70B BF16 ≥180GB 4*L20 或 2*H20 GPU 三、国产芯⽚与硬件适配⽅案 1. 国内⽣态合作伙伴动态 企业 适配内容 性能对标（vs NVIDIA） 华为昇 腾 昇腾910B原⽣⽀持R1全系列，提供端到端推理优化 ⽅案 等效A100（FP16） 沐曦 GPU MXN系列⽀持70B模型BF16推理，显存利⽤率提升

大模型框架有哪些特点： ：大模型开发框架通过提供高 层次的 API 简化了复杂模型的构建过程。这 些 API 抽象掉了许多底层细节，使开发者能 够专注于模型的设计和训练策略。 ：这些框架经过优化，以充分利用 GPU、TPU 等高性能计算硬件，以加速模型 的训练和推理过程。 ：为了处理大型数据集和大规模参 数网络，这些框架通常设计得易于水平扩展， 支持在多个处理器或多个服务器上并行处理。 ：它们提供工具来有效地加 Platform 和 Microsoft Azure Machine Learning 都是提供端到 端机器学习服务的云平台。 这些工具和库专门为加速机器学习模型的训练和推理而设计，通常利 用 GPU 或 TPU 等硬件。这类工具可以显著提高训练和推理的速度， 使得处理大规模数据集和复杂模型变得可行。NVIDIA CUDA 和 Google Cloud TPU 均是此类工具。 这类工具通常由 的算力指的是执行这些模型所需的计算资源。这包括用于训练和运行模型的硬件（如 GPU 或 TPU）、内存、存储空间以及处理 大量数据的能力。LLM 需要非常强大的算力来处理、理解和生成文本，因为它们涉及到数十亿甚至数万亿个参数的训练和推理。 LLM 的基石是算力，而算力的基石是硬件，硬件的性能直接影响着计算任务的速度、效率和能力。 是全球领先的 GPU 制造商，提供了强大的图形处理单元，专门用于深度学习和AI计算。

100 150 200 250 300 DeepSeek-V2 DeepSeek 67B saving 42.5% of training costs Training Costs (K GPU Hours/T Tokens) 0 100 200 300 400 DeepSeek-V2 DeepSeek 67B reducing KV cache by 93.3% KV Cache cluster, for training on each trillion tokens, DeepSeek 67B requires 300.6K GPU hours, while DeepSeek-V2 needs only 172.8K GPU hours, i.e., sparse DeepSeek-V2 can save 42.5% training costs compared with high demands on the training framework. It requires careful engineering optimization to manage the GPU memory and RAM pressure, and meanwhile maintain a fast training speed. For this goal, we implement

(bcm2837) -target=armv7l-linux-gnueabihf -mattr=+neon pynq -target=armv7a-linux-eabi -mattr=+neon GPU mali (midgard) firefly rk3399, rock960 (mali t860) N/A opencl bifrost hikey960 (mali g71) N/A FPGA closely in an organized way ○ Arm - Cortex-A/Cortex-M/Neoverse CPU, Mali GPU, Ethos NPU ○ Qualcomm - Hexagon DSP, Adreno GPU ○ Hisilicon, Xilinx, NXP, TI, ST, Fujitsu, Riken, and etc ● Collaborations