Effective Neural Networks Without GPU (SOL)
Full title: Effective Neural Networks Without GPU. SOL: Transparent Neural Network Acceleration on NEC SX-Aurora TSUBASA
In 2019, ICM University of Warsaw expanded its HPC infrastructure with a specialized vector computer, NEC SX Aurora TSUBASA, with eight vector processors. Aurora TSUBASA is used at ICM UW for calculations in physics, chemistry, AI, as well as development work intended to adapt and optimize the existing software to work on the new computer architecture.
Distinctive features of NEC SX-Aurora TSUBASA are:
High memory bandwidth (48 GB, HBM2) of the Vector Engine (>1 TB/s) at < 300 W, 64 fully functional vector registers and 192 double precision FP operations per cycle, Works within the GNU/Linux environment – natively or in the accelerator mode.
The Workshop is intended as an introduction to two software frameworks designed specifically for NEC SX-Aurora TSUBASA:
NEC SOL – Transparent Neural Network Acceleration – an AI acceleration middleware enabling wide range of optimizations for neural network workloads. It integrates with existing Machine Learning frameworks such as PyTorch and TensorFlow. SOL offers broad support for hardware architectures including CPUs (x86, arm64), GPUs (NVIDIA), and NEC SX-Aurora TSUBASA. It does not require modification of the original source code allowing the user to focus on solving the problem rather than on the specifics of the hardware architecture;
Frovedis – FRamework Of VEctorized and DIStributed data analytics – data analytics software primarily targeting the NEC SX-Aurora TSUBASA architecture.
- SOL: Transparent Neural Network Acceleration
- Integration with PyTorch
- Integration with ONNX -Deployment
- Frovedis: FRamework Of VEctorized and DIStributed data analytics
- Hands-on session: SOL at ICM infrastructure