3. DNN Processing Units
効率性柔軟性
Soft DPU
(FPGA)
Contro
l Unit
(CU)
Registers
Arithmeti
c Logic
Unit
(ALU)
CPUs GPUs
ASICsHard
DPU
Cerebras
Google TPU
Graphcore
Groq
Intel Nervana
Movidius
Wave Computing
Etc.
BrainWave
Baidu SDA
Deephi Tech
ESE
Teradeep
Etc.
4.
5. 0 1 2
784 x 100 + 100 x 10 = 785000本
9
60000個の
教師あり学習データ
785000 x 60000 = 47100000000回の足し算をします 471億回
6.
7.
8. 14 days 1 hour 31 mins 15 mins
Before
2017
Apr Sept Nov
ResNet-50
NVIDIA M40 GPU
ResNet-50
32 CPU
256 Nvidia P100 GPUs
ResNet-50
1,600 CPUs
ResNet-50
1,024 P100 GPUs
Facebook
UC Berkeley, TACC,
UC Davis
Preferred Network
ChainerMN
1018 single precision operations
2017
26. • Local tools
• Local Debug
• Faster
experimentation
Single VM
Development
• Larger VMs
• GPU
Scale Up
• Multi Node
• Remote Spark
• Batch Nodes
• VM Scale Sets
Scale Out
29. Trained AI
Model
score.py
{ JSON
}schema.json conda_dependencies
.yml
Azure Machine Learning
Model Management
Run Time
Model
Registry
Image
Registry
Manifest for
Image Generation
Single
Machines
(e.g.. DSVM, IoT
Devices, local
PC)
Azure Container Service
(AKS) – Kubernetes clusters
4
1 2 3