5. CHASING 1060 CHEMICALCOMPOUNDS
Identifying molecules with desirable chemical properties is centralto many
industries. In the chemicalspace of 1060 conceivable compounds, only
108 have been synthesized.
Screening even a small fraction of the remaining compounds with legacy
methods would take 100 node-seconds per compound.
Researchers at Dow are using GPU-powered deep learning to deliver
completely novelmolecular structures with specific properties.
The AI produced 3M promising chemicalleads in 1 day on an NVIDIA DGX.
6. AI IS SPEEDING THE PATH TO FUSION ENERGY
Fusion, the future of energy on Earth, is a highly sensitive process where small environmentaldisruptions can stall reactions and damage
multi-billion machines. Current models predict disruptions with 85% accuracy — ITER will need something more precise.
Researchers at Princeton University developed the GPU-powered Fusion Recurrent NeuralNetwork (FRNN) to predict disruptions. FRNN
has achieved 90% accuracy and is on the path to achieving 95% accuracy necessary for ITER’s tests.
Visualization courtesy of Jamison Daniel, Oak Ridge Leadership Computing Facility
33. 33
全結合層 (FULLY CONNECTED LAYER)
行列ベクトル積 (y = Wx)
Ny
計算量:
2 * 出力ノード数 (Ny)
* 入力ノード数 (Nx)
入力 x 出力 y重み W
Nx
重み W
入力 x
出力 y
Ny
Nx
Nx
[性能律速点]
メモリバンド幅
34. 34
全結合層 (FULLY CONNECTED LAYER)
ミニバッチ学習で、行列積に (Y = W・X)
重み W
計算量:
2 * 出力ノード数 (Ny)
* 入力ノード数 (Nx)
* ミニバッチサイズ (Nb)
入力 X 出力 Y重み W
Nb * Nx Nb * Ny
入力 X
出力 Y
Ny
Nx
Nx
Nb
[性能律速点]
演算性能
87. 87
量子化とモデル精度
8-bit INT に量子化しても、同程
度の精度を維持
B. Jacob, et. al., “Quantization and Training of Neural Networks for
Efficient Integer-Arithmetic-Only Inference”
Image classification Object detection