Last week, during a conference for investors, AMD presented its long-term strategy in the field of software. In particular, the company’s CEO Lisa Su shared plans for the release of new drivers and middleware designed to increase the speed of applications running on machines with central processors, video cards, AI accelerators and network chips manufactured by AMD.
Currently, the company offers a wide range of software tools for its chips, which are mostly disconnected from each other.
The software stack using AI technologies includes optimized output models, compilers, libraries and runtime environments for Linux and Windows. The ROCm open source software stack is equipped with the same set of tools. In addition, AMD now has the Xilinx Vitis platform for specialized AI chips. The company’s next step, according to Su, will be to combine all these disparate software stacks on a single platform.
In particular, AMD announced a unified software strategy Unified AI Stack 2.0, combining ROCm, software stack for CPU and Vitis environment. An external interface focused on logical conclusions will support major industry platforms, including TensorFlow and Pytorch. The development and deployment stacks will include quantization and pruning tools that eliminate parts of the network in order to improve performance and reduce the resource requirements of the model.
“We intend to unify even more middleware, and now we will have commonality in terms of our machine learning graph compiler. We will have much more in common in library APIs and outputs. To do this, we will definitely release many more pre-optimized models,” said Victor Peng, head of AMD’s adaptive and embedded computing group.
AMD also intends to add more tools to the training stack, but the company’s representatives did not specify what it would be.