CHASING 1060 CHEMICALCOMPOUNDSIdentifyingmolecules with desirable chemical properties is centralto manyindustries. In the chemicalspace of 1060 conceivable compounds, only108 have been synthesized.Screening even a small fraction of the remaining compounds with legacymethods would take 100 node-seconds per compound.Researchers at Dow are using GPU-powered deep learning to delivercompletely novelmolecular structures with specific properties.The AI produced 3M promising chemicalleads in 1 day on an NVIDIA DGX.
6.
AI IS SPEEDINGTHE PATH TO FUSION ENERGYFusion, the future of energy on Earth, is a highly sensitive process where small environmentaldisruptions can stall reactions and damagemulti-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. FRNNhas 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
82インファレンスに必要な計算精度• FP32はインファレンスには過剰、FP16/INT8で十分• 初代Google TPU (インファレンス専用) は INT8 チップ• ウェイトは2 or 3値で十分と主張している研究もある• 2 or 3値だと、一般的にモデル精度が低下• 精度確保のためモデル変更が必要になることが多い(*) “In-Datacenter Performance Analysis of a Tensor Processing UnitTM”
83.
83インファレンスに必要な計算精度• FP32はインファレンスには過剰、FP16/INT8で十分• 初代GoogleTPU (インファレンス専用) はINT8チップ(*) Matthieu Courbariaux, et al., “BinaryConnect: Training DeepNeural Networks with binary weights during propagations”• ウェイトは2 or 3値で十分と主張している研究もある• 2 or 3値だと、一般的にモデル精度が低下• 精度確保のためモデル変更が必要になることが多いBetter
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 forEfficient Integer-Arithmetic-Only Inference”Image classification Object detection