[單元9] AI深度學習CNN硬體加速器(2)-PCIe 8位元量化CNN加速器
實作8bit quantization 及PCIe CNN軟硬體協同加速器
TWD $5200
5 hours 6 mins
作者
ChipSkywalker
章節及內容
1 堂課
18 分鐘
更詳細簡介軟硬體設計與運作流程 (免費預覽課堂)
18 分鐘
6 堂課
51 分鐘
Lab準備-Quantization Training Framework 下載安裝
8 分鐘
硬體設計工具Vivado及HLS安裝
註:Vitis (軟體SDK可以不用安裝), 裝Vivado(會順利裝Vitis HLS), 再裝HLS Patch
12 分鐘
Lab Source Zip 下載 (含參考答案)
請由[課堂內容]下方Link下載, 開於本課程智財權部份,請勿對外分享
Lab 1-1 量化訓練平台的建立及Python Code
Install Python Anaconda, TensorFlow/QKeras and requirements
20 分鐘
Lab 1-2 執行Python Code導出8bits量化參數
1-2 Running Python code to train CNN and export (FP8/INT8)
8 分鐘
Lab 1-1/1-2 實際操作示範影片
3 分鐘
4 堂課
57 分鐘
Lab 2 HLS Fixed8 推論, C-Simulation, C-Synthesis, C to RTL- Part 1
18 分鐘
Lab 2 HLS Fixed8 推論, C-Simulation, C-Synthesis, C to RTL- Part 2
19 分鐘
Lab 2 實際操作示範影片
16 分鐘
Lab2 Homework
4 分鐘
6 堂課
155 分鐘
Lab 3 - Part 1 - FPGA PCIe 電路板安裝
19 分鐘
Lab 3 - Part 2 - PCIe CNN硬體電路設計(RTL, HLS IP, Block Design 整合)
54 分鐘
Lab 3 - Part 3 - PCIe CNN實際操作影片
37 分鐘
Lab 3- Part 4 - PCIe CNN C++軟體設計
35 分鐘
Lab 3 - Part 5 - 驅動程式安裝
2 分鐘
Lab 3 - Part 6 - PCIe CNN C++軟體操作影片
8 分鐘
2 堂課
26 分鐘
Lab 4 int8量化訓練, HLS 推論測試
14 分鐘
Lab 4 實際操作示範影片
12 分鐘
產品介紹
單元8(Edge AI SoC)的後續相關課程,
- Lab 1 Training CNN model and Quantizing 8-bit weight
- 1-1 Install Python Anaconda, TensorFlow/QKeras and requirements
- 1-2 Running Python code to train CNN and export (FP8/INT8) Weight/Bias for Parallel Hardware Inference
- Lab 2 HLS Hardware FP8 Quantization CNN Inference
- HLS C++ Testbench to read MNIST dataset
- HLS C++ CNN design to load FP8 Weight
- C Simulation, C Synthesis and C/RTL Co-Simulation
- Compare the HLS C++ Inference accuracy with Qkeras training
- Further optimization to reduce IP logic area (keep accuracy)
- Compare FP8 vs float/half [AI NPU (1)] CNN IP : area/latency
- Exporting HLS AXI4 AXI-Stream(AXIS) FP8 CNN RTL IP
- Lab 3 PCIe AI CNN Accelerator
- PCIe DMA, AXI DMA and Memory controller IP Introduction
- (H/W) Integrating with AXIS FP8 CNN IP
- (S/W) Writing S/W C++ for Linux x86 PCIe FPGA CNN Acceleration
- Testing Accuracy & Performance of the PCIe CNN Accelerator
- Install DMA driver
- Compile S/W C++ for Linux x86 PCIe FPGA
- Setup/Configure PCIe FPGA H/W
- Execute CNN Inference
- Lab 4 INT8 Quantization training and HLS Inference Test
- Future Work
- Supporting more layers (eg BatchNormalization) and complex models
- Optimize H/W CNN by pure RTL coding for higher performance
課程注意事項
- 建議學習基礎單元5/6/7,與單元8合購有優惠,詳見首頁優惠或來信詢間問
- 課程有效期 : 購買後二年(730天)