Extending WebAssembly for Deep-Learning Inference Across the Cloud Continuum

Recent advancements in serverless computing and the cloud-edge continuum have increased interest in WebAssembly (WASM). This technology enables portability and interoperability across diverse computing environments while achieving near-native execution speeds. Currently, WASM supports Single Instruction Multiple Data (SIMD), which allows for data-level parallelism that is particularly beneficial for vectorizable operations such as general matrix-matrix multiplication (GEMM) and convolutional layers. However, WASM lacks native integration with specialized hardware accelerators like GPUs, TPUs, and NPUs, as well as the ability to benefit from multi-core processing capabilities, which are critical for efficiently running Deep-Learning (DL) workloads. In contrast, despite these gains, WASM still lacks native support for heterogeneous accelerators such as GPUs, TPUs, and NPUs, as well as full multi-core parallelism capabilities that are critical for meeting the latency and throughput requirements of modern DL inference services. To bridge this gap, WASI-NN was developed, enabling WASM to integrate with external runtimes such as OpenVINO and ONNX Runtime, which leverage hardware acceleration. However, these current integrations often introduce performance overhead on certain devices, restricting their usability across the CECC. To address these challenges, we propose a new integration focusing on TVM as an external runtime for WASI-NN to enhance WASM’s performance and expand support to a broader range of devices. Additionally, we integrate this solution into Knative, a serverless framework, to provide a scalable and flexible platform for DL deployment. Using WASM technology, we evaluate our TVM-based solution through comparative studies. Results on AMD CPUs demonstrate the effectiveness of our approach, achieving 58% overall gain over other WASI-NN integrations (e.g., ONNX Runtime and OpenVINO) for CNN-based models while also achieving optimal performance on different platforms, such as Intel GPUs. These findings highlight the effectiveness of our solution.