This website uses cookies to manage authentication, navigation, and other functions. By using our website, you agree that we can place these types of cookies on your device.

View e-Privacy Directive Documents

The DEEP-EST project provides the partners with the opportunity to explore new technologies. Here, we want to present ASTRON's Approach on a Technology called tensor cores.  

 

Figure 1 illustrates how disruptive the tensor core technology is. While the benefits of Moore's law and improved cores are clearly levelling off, the innovative Approach of using Tensor cores can provide even higher Performance.

 

Tensor cores are mixed-precision matrix-multiplication units that were added to the latest generations of GPUs. They perform matrix multiplications up to eight times faster than regular GPU cores, albeit with lower accuracy. Tensor cores are designed to speed up deep learning, where 16-bit numbers are sufficiently precise to support training and inference. However, as tensor cores are directly programmable and just perform matrix multiplications, their use is not limited to deeplearning.

ASTRON explores Tensor cores for signal processing, for which 16 bits precision is also sufficient. Several signal-processing algorithms, like correlations and beam forming can be expressed as matrix multiplications and are tested. These signal-processing algorithms are commonly found in the processing pipelines of radio telescopes.

This is not a straightforward process though, as signal-processing algorithms operate on complex numbers, while tensor cores operate on real numbers. Additionally, tensor cores compute so quickly that providing input data is not trivial: efficient fetching and caching of data throughout the memory hierarchy is essential. Yet, unprecedented processing speeds of 70-80 (!) TFLOPS on key signal-processing algorithms are achieved on a single GPU. In contrast, regular GPU cores achieve no more than 12-13 TFLOPS.

Figure 1 illustrates how disruptive the tensor core technology is. While the benefits of Moore's law and improved cores are clearly levelling off, the innovative Approach of using Tensor cores can provide even higher Performance.


Figure 1. Innovation beyond Moore's law: The graph shows correlation performance with 576 receivers, for five consecutive GPU generations (2013-2019), both on regular GPU cores and tensor cores.