Tiny-SYCL-NN

• 

  Teaser

• 

  Our motivation

• 

  Network Architecture

• 

  Our development path

• 

  Component 1: Position encoding

• 

  Component 2: Network inference

• 

  Component 2: Network inference

• 

  Current results

NeRF-like implicit neural networks exhibit significant promise across various domains, but their application in real-time settings is hampered by the intensive computational costs of training and inference. To address this challenge, machine learning libraries such as Tiny-CUDA-NN have emerged, utilizing low-level GPU features to notably accelerate neural network inferences. Such libraries offer significant performance improvements over mainstream frameworks like PyTorch and TensorFlow, paving the way for effective 3D digitization. However, the proprietary nature of Tiny-CUDA-NN, tailored for NVIDIA GPUs, limits its broader applicability and it lacks several features vital for implicit neural networks. Our goal is to create a cross-platform version of Tiny-CUDA-NN using SYCL, and fine-tune it through the DPC++ Explicit SIMD API and Intel’s XMX extensions.

Built With

• c++
• python
• sycl