For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2023/06/state-of-the-art-model-quantization-and-optimization-for-efficient-edge-ai-a-presentation-from-deepx/ Hyunjin Kim, Senior Staff Engineer at DEEPX, presents the “State-of-the-art Model Quantization and Optimization for Efficient Edge AI” tutorial at the May 2023 Embedded Vision Summit. Extremely efficient edge AI requires more than efficient processors; it also requires tools capable of generating superefficient software. In this talk, Kim explains and demonstrates how DEEPX’s DXNN SDK utilizes state-of-the-art optimization techniques to generate extremely efficient, accurate code for DEEPX’s new M1 neural processor. Kim begins by describing how the DXNN SDK uses hardware-aware, selective quantization to maintain high accuracy while achieving efficient DNN implementations. Next, he explains how the SDK maps DNN layer operations into processor micro-operations to provide both efficiency and flexibility. Kim also shows how the DEEPX SDK conserves memory by utilizing tiling, layer fusion and feature reuse. Finally, he illustrates the ease of use of the SDK by demonstrating the use of the DXNN SDK to implement a state-of-the-art model on the M1 NPU.

1. State-of-the-Art Model
Quantization and
Optimization for
Efficient Edge AI
Hyunjin Kim
Compiler Team Lead
DEEPX AI

2. DEEPX: Edge AI Solution
2
© 2023 DeepX AI
NPU
Technology
NPU based SoC
Technology
AI HW/SW
Optimization
Developing one of the most efficient NPU technologies
Developing SoC ASIC with NPU for commercial products
Optimizing both AI HW & SW to provide the highest NPU efficiency

3. DEEPX: Edge AI Solution
3
© 2023 DeepX AI
NPU
Technology
NPU based SoC
Technology
AI HW/SW
Optimization
Developing one of the most efficient NPU technologies
Developing SoC ASIC with NPU for commercial products
Optimizing both AI HW & SW to provide the highest NPU efficiency
This Talk!

4. • Satisfy the EDGE AI inference requirement
• EDGE AI inference requirement
• Maintain high accuracy with quantization
• Support various AI models (operation coverage)
• Minimize HW Cost: SRAM
DEEPX Mission
4
© 2023 DeepX AI

5. • Satisfy the EDGE AI inference requirement
• EDGE AI inference requirement
• Maintain high accuracy with quantization
• Support various AI models (Operation Coverage)
• Minimize HW Cost: SRAM
DEEPX Mission
5
© 2023 DeepX AI
Our Solution:
HW/SW Co-design

6. • HW/SW Co-Design
• Selective quantization to avoid accuracy drop
• 8-bit quantized ops + FP32 ops
High Accuracy with Quantization
6
© 2023 DeepX AI
DXNN-Lite DXNN-Pro DXNN-Master
Norm. Accuracy
Over FP32 (avg.)
99.01% 99.42% 100.54%
Time Cost ~ 1 min - 1~2 hours 50+ GPU Hours
* Training env.

7. • High operation coverage with minimum implementation cost
• HW/SW Co-Design
• Increase op coverage via combinations of supported ops
• TransConv: Resize + Padding + Conv
• DilatedConv: multiple Convs
• ArgMax: Two-stage implementation
Operation Coverage
7
© 2023 DeepX AI

8. • DEEPX NPUs have small SRAM sizes.
• HW/SW Co-Design
• DEEPX compiler’s memory optimization maximizes SRAM
utilization to reduce DRAM accesses
Minimize HW Cost: SRAM
8
© 2023 DeepX AI
SRAM: 1 MB SRAM: 8.25 MB
SRAM: 2.25 MB

9. Memory Optimization: DRAM Accesses
9
© 2023 DeepX AI
49%
0%
20%
40%
60%
80%
100%
120%
MobileNetv1 MobileNetv2 MobileNetv3-Large ResNet50 EfficientNet-B0 YOLOv4 (telechips) YOLOv5s Average
Normalized DRAM Accesses Over Base [M1-4K, 2.25 MB]
Base #DRAM

10. Memory Optimization: Performance
10
© 2023 DeepX AI
137%
0%
20%
40%
60%
80%
100%
120%
140%
160%
MobileNetv1 MobileNetv2 MobileNetv3-Large ResNet50 EfficientNet-B0 YOLOv4 (telechips) YOLOv5s Average
Normalized Performance (FPS) Over Base [M1-4K, 2.25 MB]
Base MemOpt

11. 11
Case Study: Fused Operations
© 2023 DeepX AI

12. • Segmentation Models
• Resize → Argmax at the end for pixel-wise classification
Fused Operation: Resize + Argmax (1/4)
12
bisenet_v2
© 2023 DeepX AI

13. • Segmentation Models
• Resize → Argmax at the end for pixel-wise classification
Fused Operation: Resize + Argmax (2/4)
13
bisenet_v2
1 x 171 x 640 x 640 x 4 (FP32) = 267 MB
1 x 640 x 640 x 2 (INT16) = 800 KB
© 2023 DeepX AI

14. • Segmentation Models
• Resize → Argmax at the end for pixel-wise classification
Fused Operation: Resize + Argmax (3/4)
14
bisenet_v2
1 x 171 x 640 x 640 x 4 (FP32) = 267 MB
Fusing Resize and Argmax by performing Argmax during
Resize via 2-stage Argmax (171 → 1)
1 x 640 x 640 x 2 (INT16) = 800 KB
© 2023 DeepX AI

15. • Segmentation Models
• Resize → Argmax at the end for pixel-wise classification
Fused Operation: Resize + Argmax (4/4)
15
Fused
bisenet_v2
1 x 171 x 640 x 640 x 4 (FP32) = 267 MB
Fusing Resize and Argmax by performing Argmax during
Resize via 2-stage Argmax (171 → 1)
1 x 640 x 640 x 2 (INT16) = 800 KB
© 2023 DeepX AI

16. 16
DXNN SDK Tutorial
© 2023 DeepX AI

17. DXNN™ – DEEPX NPU SDK
17
© 2023 DeepX AI
 Beta version of DXNN has been released to customers.

18. DXNN SDK Tutorial: Overview
18
© 2023 DeepX AI
Model ONNX file
Model Config file
- input shape
- calibration dataset
- preprocessing
NPU Binary
NPU binary config
Weight Binary

19. • DX-COM compiles the model onxx file
• It also performs the quantization as well.
• The config file includes the model meta data.
• input shape, input pre-processing parameter, …
$./dx_com/dx_com
--model_path sample/yolov5s.onnx
--config_path sample/yolov5s.json
--output_dir output/yolov5s
DXNN SDK Tutorial: Model Compilation
19
© 2023 DeepX AI

20. $run_model --model /dxrt/models/yolov5s # compilation output directory
--input /dxrt/models/yolov5s/input/npu_input_0.bin
--ref /dxrt/models/yolov5s/output/output_last/npu_output_0.bin
--output output.bin
--loop 5
DXNN SDK Tutorial: Run Inference
20
© 2023 DeepX AI
DX-RT
Model Inference
Option
Input
Tensors
Post Processing
Output
Tensors

21. DXNN SDK Tutorial: Post Processing
21
© 2023 DeepX AI
• User should configure the parameters for post-processing.
• The parameters are not included in the compilation output.
• Example: yolov5s
YoloParam yolov5s_512 = {
.imgSize = 512,
.confThreshold = 0.25,
.scoreThreshold = 0.3,
.iouThreshold = 0.4,
.numClasses = 80,
.layers = {
…
};

22. #include "dxrt/dxrt_api.h“
#1 Set Inference Options: 1 NPU, 1 WORKER THREAD, 2 MEMORY BUFFER
dxrt::InferenceOption option(1, 1, 2, dxrt::InferenceMode::MODE_SYNC);
DXNN SDK Tutorial: DX-RT API #1
22
© 2023 DeepX AI

23. #include "dxrt/dxrt_api.h“
#1 Set Inference Options: 1 NPU, 1 WORKER THREAD, 2 MEMORY BUFFER
dxrt::InferenceOption option(1, 1, 2, dxrt::InferenceMode::MODE_SYNC);
auto ie = dxrt::InferenceEngine("yolov5s", &option); #2 Load Model
DXNN SDK Tutorial: DX-RT API #2
23
© 2023 DeepX AI

24. #include "dxrt/dxrt_api.h“
#1 Set Inference Options: 1 NPU, 1 WORKER THREAD, 2 MEMORY BUFFER
dxrt::InferenceOption option(1, 1, 2, dxrt::InferenceMode::MODE_SYNC);
auto ie = dxrt::InferenceEngine("yolov5s", &option); #2 Load Model
auto inputs = ie.GetInput(); #3 Set inputs
DXNN SDK Tutorial: DX-RT API #3
24
© 2023 DeepX AI

25. #include "dxrt/dxrt_api.h“
#1 Set Inference Options: 1 NPU, 1 WORKER THREAD, 2 MEMORY BUFFER
dxrt::InferenceOption option(1, 1, 2, dxrt::InferenceMode::MODE_SYNC);
auto ie = dxrt::InferenceEngine("yolov5s", &option); #2 Load Model
auto inputs = ie.GetInput(); #3 Set inputs
auto outputs = ie.Run(); #4 Run Inference
DXNN SDK Tutorial: DX-RT API #4
25
© 2023 DeepX AI

26. #include "dxrt/dxrt_api.h“
#1 Set Inference Options: 1 NPU, 1 WORKER THREAD, 2 MEMORY BUFFER
dxrt::InferenceOption option(1, 1, 2, dxrt::InferenceMode::MODE_SYNC);
auto ie = dxrt::InferenceEngine("yolov5s", &option); #2 Load Model
auto inputs = ie.GetInput(); #3 Set inputs
auto outputs = ie.Run(); #4 Run Inference
auto result = yolo.PostProc(outputs); #5 Perform Post Processing
DXNN SDK Tutorial: DX-RT API #5
26
© 2023 DeepX AI

27. DXNN SDK Tutorial: Result
27
© 2023 DeepX AI

28. • DEEPX’s HW/SW codesign leads to
• Meeting the required specifications with minimum HW cost
• Selective quantization
• The combination of supported ops
• Minimize HW Cost: SRAM
• Case Study
• Reduced DRAM access by fused operation
Conclusion
28
© 2023 DeepX AI

29. Resources
2023 Embedded Vision Summit
Previous Talks from DEEPX
1. “Toward the Era of AI Everywhere”
(Lokwon Kim, CEO)
2. “DEEPX’s New M1 NPU Delivers
Flexibility, Accuracy, Efficiency and
Performance”
(Jay Kim, Vice President)
29
1. Demo Booth: #103
2. DEEPX Developers Page:
https://deepx.ai/developers
3. Linkedin & Youtube
© 2023 DeepX AI