Initiative
Initiative
Market opportunities
-
Merrill Lynch estimates that the global DNN training system market size grows from $650M in 2017 to $7B in 2020. Gartner estimates that 60% of DNN training systems are private and the annual growth rate is 30% ; so the total market size of private DNN training systems is expected to reach $12B in 2024.
-
Private DNN training appliance = X86 server + GPUs + DNN training software stack.
Open AI Training System
-
An open hardware architecture supporting a meshed PCIe network and multiple types of GPUs
-
A DNN training computation compiler that features efficient scheduling and mapping, and advanced training algorithms
-
A DNN Integrated Design Environment (IDE) that helps users to reduce unproductive training rounds
-
A DNN model optimization tool that minimizes the space/time requirement of already trained DNN models.
Field Trial & Business Case
-
System verification in AI Datacenter
-
Business cases as the references
New Venture
-
An innovative startup business as the open source-based AI system software provider.
Objective
Membership
Membership
*1:Members are eligible to participate in TASA workshops. (At least 6 workshops in one year)
*2:Members are eligible to provide one DNN training appliance for DNN training software install and test.
*3:Members are eligible for system verification at ITRI DNN Farm and, if verified, consulting service as to operation in AI Datacenter.
*4:Members are eligible for joint promotion with TASA at Computex, Cloud Computing Day Tokyo, and Big Data Expo. Expense is NOT included.
Appendix
OATS Architecture
Processor Type:
-
Nvidia’s Tesla P100 and V100 (12GB, 4.7TFLOPs of FP64)
-
Nvidia’s GeForce GTX 1080Ti (11GB, 11.3TFLOPs of FP32)
-
AMD’s Radeon Instinct MI25 (12.3 TFLOPS of FP32)
-
AMD Radeon RX Vega 64 (12.6 TFLOPS of FP32)
-
Intel multi-node Xeon
-
FPGA
Number of “GPU”s: 16+
System Interconnect: Meshed PCIe network supporting disaggregate rack architecture.
Intelligent thermal load management
Graphics driver API:
-
CUDA
-
OpenCL
DNN training framework:
-
Tensorflow
-
Caffe/Caffe2/NVCaffe