TempoQuest collaborates with AMD to port AceCAST on AMD Instinct™ MI200 series accelerators

TempoQuest collaborates with AMD to port AceCAST on AMD Instinct™ MI200 series accelerators

By Christian Tanasescu, COO and CTO, TempoQuest

Pictured: AMD MI210 launched today

As high-quality weather forecasts require a tremendous amount of computational power, TempoQuest (TQI) aimed at accelerating the Weather and Research Forecast (WRF) model from NCAR by utilizing GPU in its product AceCAST™.

Compared to standard computing methods, AceCAST™ is currently the world’s fastest and highest resolution weather forecasting model. It encompasses an ample set of refactored common WRF physics and dynamics modules with CUDA and OpenACC programming techniques, allowing users to adopt AceCAST™ painlessly as a drop-in replacement for existing WRF configurations. Currently, it runs on NVIDIA A100 and V100 series accelerators.

TQI announces the partnership with AMD to port AceCAST™ on AMD Instinct™ MI200 series accelerators. The port will take advantage of the ROCm™ 5 platform to exploit the powerful capabilities of MI200 series accelerators.

The CUDA codebase in AceCAST™ has been already translated to HIP and tested on MI210. Initial benchmarks reveal that standalone physics schemes developed in CUDA, like Thompson achieve on MI210 more than 50% higher performance than the current accelerators available on the market . TQI is committed to optimize AceCAST™ on AMD Instinct™ MI250 and to further improve the performance leveraging the ROCm 5 development platform.
Conversion of the OpenACC codebase is in progress.

AceCAST™ supports multi-GPU systems, runs entirely on GPU and utilizes one core per GPU for job control and IO. It dramatically accelerates the compute time required to run the model by factor 10x or even higher depending on the number and type of GPUs used, the domain size, resolution, grid points and parameterization, IO frequency, etc.

By accelerating WRF model on GPUs, AceCAST provides tangible quantifiable improvements when compared to traditional CPU derived solutions for running WRF:
· Faster simulation time for the same simulation and compute cost
· Lower computation cost for the same simulation speed
· Higher spatial resolution/domain size for the same simulation cost; or any preferred distribution of the above.

AceCAST™ will be available on AMD Instinct™ MI200 series accelerators in Q2 2022.