| Popis: |
The SX-Aurora Tsubasa Vector Engine (VE) is NEC’s latest instantiation of their long vector architecture for high performance computing and AI with large HBM2 memory of 48 GB and high memory bandwidth of 1.2 TB/s. It is completely different from the previous mainframe-sized product generations and comes as a PCIe card pluggable into normal Linux servers, where the VE integrates seamlessly into the Linux environment and runs native VE programs compiled with C, C++ or Fortran transparently, from the command line. This report introduces Vector Engine Offloading, VEO, the base mechanisms used to extend the programming model of the VE to an accelerator style offloaded model somewhat similar to OpenCL or CUDA. This programming model extends the scope of the VE and simplifies the porting of applications which have already been adapted to using accelerators like GPGPUs. The PyVEO Python bindings furthermore simplify accessing the power of the VE even from scripts and interactive notebooks. |
