Scientific applications are increasingly adopting Artificial Intelligence (AI) techniques to advance science. There are specialized hardware accelerators designed and built to run AI applications efficiently. With a wide diversity in the hardware architectures and software stacks of these systems, it is challenging to understand the differences between these accelerators, their capabilities, programming approaches, and how they perform, particularly for scientific applications.
We will cover an overview of the AI accelerators landscape with a focus on SambaNova, Cerebras, Graphcore, and Groq systems along with architectural features and details of their software stacks. We will have hands-on exercises that will help attendees understand how to program these systems by learning how to refactor codes written in standard AI framework implementations and compile and run the models on these systems.
- Intro to AI Series: AI Accelerators
Slides will be uploaded shortly after the talk.
To gain access to AI Testbeds at ALCF after current allocation expires apply for Director’s Discretionary Allocation Program
The ALCF Director’s Discretionary program provides “start up” awards to researchers working to achieve computational readiness for for a major allocation award.
You need to submit either Theory Homework or Hands-on Homework.
- What are the key architectural features that make these systems suitable for AI workloads?
- Identify the primary differences between these AI accelerator systems in terms of their architecture and programming models.
- Based on hands-on sessions, describe a typical workflow for refactoring an AI model to run on one of ALCF's AI testbeds (e.g., SambaNova or Cerebras). What tools or software stacks are typically used in this process?
- Give an example of a project that would benefit from AI accelerators and why?
- Overview of AI Testbeds at ALCF
- ALCF AI Testbed Documentation
- Director’s Discretionary Allocation Program
- ALCF Events Page
Contributors: Siddhisanket (Sid) Raskar, Murali Emani, Varuni Sastry, Bill Arnold, and Venkat Vishwanath.
This research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC02-06CH11357.