Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about CMCC's cross-architecture software stack designed to solve the challenge of AI application migration across heterogeneous accelerators in this 24-minute conference talk. Discover how this unified accelerator abstraction model eliminates the need for developers to rewrite application code and recompile for different AI accelerator architectures. Explore the stack's key components including a source code converter, cross-architecture compiler that works with existing compilers like NVCC as backends, and an adaptive runtime featuring dynamic device type identification and flexible just-in-time compilation. Understand how this "code once, compile once, run everywhere" approach streamlines AI development workflows across diverse hardware platforms. Examine the technical capabilities and underlying technologies of the software stack, review its open source progress under the OpenInfra Foundation's Cloud Native Function Working Group, and analyze current gaps when integrating with Kubernetes and OpenStack environments.
