Effective neural networks *without* GPU – workshop with Nicolas Weber, NEC Laboratories Europe
Discover two software frameworks designed specifically for NEC SX-Aurora TSUBASA: SOL – Transparent Neural Network Acceleration and FROVEDIS – FRamework Of VEctorized and DIStributed data analytics.
Date: Wed, Sept 30, 2020
Time: 09.00 am to 12.30 pm CEST
Limited access: 15 participants – first come first served
Payment: Free of charge
Note! Registration deadline, Sept 27, 2020. Participants will be informed about their qualification for the workshop by September 28th.
In 2019, ICM University of Warsaw expanded its HPC infrastructure with a specialized vector computer, NEC SX Aurora TSUBASA [1], 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 [2] – 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 [3] – data analytics software primarily targeting the NEC SX-Aurora TSUBASA architecture.
About SX-Aurora TSUBASA
NEC SX-Aurora Tsubasa, introduced to the market in 2018, is a vector processor (Vector Engine, VE) belonging to the SX architecture line which has been developed by NEC Corporation since mid-1980s [1]. Unlike its stand-alone predecessors, Tsubasa has been designed as a PCIe attached card working within and being operated by an x86_64 host server (Vector Host, VH) running a distribution of the GNU/Linux operating system. The latter provides a complete software development environment for the connected VEs and runs Vector Engine Operating System (VEOS) which, in turn, serves as operating system to the VE programs. [4].
[1] https://www.nec.com/en/global/solutions/hpc/sx/index.html
[2] http://sysml.neclab.eu/projects/sol/
[3] https://github.com/frovedis/frovedis
[4] https://github.com/veos-sxarr-NEC/veos
About presenter
Nicolas Weber is Senior Researcher at the NEC Laboratories Europe. He received his PhD for work on automated memory access optimizations for GPU in 2017 from TU Darmstadt. Since then he focusses on the efficient mapping of artificial intelligence workloads onto various accelerator processors such as NEC SX-Aurora or GPUs, to transparently increase performance and efficiency.
SEE ALSO
10 recordings of the Virtual ICM Seminars In Computer And Computational Science – worldwide Open Science online meetings in HPC, AI, Quantum Computing, BigData, IoT, computer and data networks held in the past few months after SCFE20 conference.
The series of Virtual ICM Seminars in Computer and Computational Science will return in the new academic year 2020/2021.
Details soon on supercomputingfrontiers.eu/2020/seminars/