FPGAs (FPGAs) are not as versatile and easy to program as CPUs, but they can significantly speed up specific tasks. In the hands of a skilled engineer, FPGAs can free the CPU from a wide variety of workloads. Intel back in 2014 started integrating Xeon and FPGA for customers, but this did not lead to any open source product of this kind. The new patent shows that AMD is also thinking in this direction, but has chosen a different approach.
On December 31, 2020, the United States Patent and Trademark Office (USPTO) published AMD’s application for FPGA integration with the central processor, which was filed back in the summer of 2019. AMD included 20 points in its patent application, but the gist of them, in general, boils down to the fact that the processor can include one or more execution units that can be programmed to process different types of user instruction sets. This is exactly what FPGA does.
AMD has already made a splash with its chiplet-based processors, but in this case we are talking about something completely different. The programmable FPGA module in the AMD patent uses shared registers with CPU integer execution units and floating point modules. Therefore, we are talking about integrating an FPGA with a CPU on a single die – otherwise it would all work too slowly. Such programmable blocks will be able to process special types of data like FP16 used to accelerate AI.
In the case of several FPGAs on a chip, each module can be programmed using a different set of specialized instructions, and a specific AMD client can reconfigure these programmable blocks to suit their needs. Moreover, the processor will be able to reconfigure the FPGA to accelerate specific tasks for specific software.
AMD has been working on various ways to accelerate computing in AI for years. First of all, we were talking about specialized accelerators based on Radeon video cards. Moving to FPGAs after acquiring Xilinx makes sense, and I would like to see what this patent can turn out to be in reality. Perhaps it will be some time before we see products based on this design – it is too early to think that such chips will become part of the EPYC family in the foreseeable future.
If you notice an error, select it with the mouse and press CTRL + ENTER.