NeurIPS 2019 Expo Workshop
Hands-on Workshop: Implementing High-Performance AI Workloads with Habana AI Processors
Sponsor: Habana Labs
The ML community has seen demand spike for AI processors (AIPs), expressly designed for Machine Learning workloads. The workshop addresses practical aspects of implementing AIPs. Part One: 2.5-hour hands-on session. Leon Goldgeisser will provide details on how to implement an inference system with the Habana Goya Inference Processor. He will share foundations of the processor’s hardware and software and demonstrate how to use Goya to solve the most common and computationally extensive inference tasks. The full inference process will be demonstrated: starting from a trained model, converting the model into Habana’s Intermediate Representation model, reviewing Quantization, including methods and considerations, optimizing for performance or accuracy, debugging and profiling for bottlenecks, and demonstrating how the Habana Kernel library can be extended with User Specific Kernels. Part Two: 30-minute presentation. Whitney Zhao of Facebook will present industry advances to address the explosion of new types of hardware accelerators for Machine Learning, Deep Learning, and High-Performance Computing. Different implementations target similar requirements for power/cooling, robustness, serviceability, configuration, programming, management and debug, as well as inter-module communication to scale up and input/output bandwidth to scale out. The industry needs an open infrastructure to intercept rapid AI innovation. OAI is where open accelerator infrastructure meets open artificial intelligence. Zhao will discuss industry collaboration to standardize accelerator module form factors to OCP Accelerator Module (OAM) and build a modularly interoperable infrastructure around OAM. Attendees will be given access to the lab machine and may run inferences on supplied models or models from the internet. Advanced users may write kernels and run them on the platform. Users may bring a laptop with an SSH, VNC or NoMachine client to connect to the lab machine.