2 min

Tags in this article

, , ,

Google wants to make the development of AI models based on machine learning even more effortless. The tech giant recently released a new runtime of its TensorFlow platform. This makes the use of machine learning models easier for multiple devices.

According to Google, the main reason for the release of the new version of TensorFlow runtime is that the development of machine learning models and ecosystems is subject to significant change. This means that there are more and more complex models and deployment scenarios that require more computing power.

Runtime adjustment

Google has now adjusted and optimized its TensorFlow runtime. Instead of the current TensorFlow stack optimized for a graphical deployment, the machine learning platform and ecosystem now has a high-performance and low-lever runtime. This runtime should also reduce overhead in the execution of machine learning operations.

Among other things, the new runtime should make the use of multithreaded host CPUs more efficient. It should also support asynchronous programming models and focus on low-level efficiency. Furthermore, the new runtime will make it easier for kernels to run on the desired hardware. This makes the runtime more suitable for a larger variety of devices.

TensorFlow Runtime Features

The new TensorFlow Runtime (TFRT) features a lock-free graphic executor with low synchronisation overhead. TFRT also splits the device runtime and the host runtime, the basic component that takes care of the host CPU and I/O operations. In addition, the runtime is now integrated with MLIR’s compiler infrastructure to generate and optimise further improvements for runtime execution.

Google’s new TFRT is particularly suitable for researchers who want faster iteration time and better error reporting. Also, application developers can benefit from improved performance. The new runtime also allows hardware manufacturers to better integrate edge and data center devices modularly into TensorFlow.

TFRT Development Program

TFRT is initially integrated into TensorFlow as an opt-in feature. With this, Google wants to ensure that start-up problems are more easily repaired, and performance can be fine-tuned. However, TFRT will soon become the default runtime in TensorFlow.