Exploit Parallelism for AI Workloads with WASM and OpenMP - Atanas Atanasov & Andrew Brown, Intel

Exploit Parallelism for AI Workloads with WASM and OpenMP - Atanas Atanasov & Andrew Brown, Intel

With the staggering growth of AI in recent years, the importance of multithreaded functionality has become more prevalent. Many tech stacks will require the most optimal approaches to vector operations to meet the high demands of AI workloads at the edge. Web assembly has become an emergent technology in the cloud computing space due to the granular control and security it provides in containerised environments. Despite this, native support for multithreaded, shared memory and vectorized workloads is somewhat lacking. The aim of this talk is to highlight the potential benefits of Wasm for AI workloads using the wasi-threads interface in conjunction with modern parallel execution interfaces such as OpenMP. In the talk we will highlight the challenges of bringing some of the key OpenMP parallelisation and vectorization features to Wasm based on example of a CNN Kernel.