AMD spoke in yesterday’s quarterly report about the preparation of processors that integrate FPGA matrices – the first products of this type will appear as early as 2023. Such chips will perform better in tasks related to artificial intelligence. They are used in data centers and servers.
AMD recently acquired Xilinx, one of the largest FPGA developers, for tens of billions of dollars, so the appearance of unusual EPYCs with built-in programmable matrices was to be expected.
AMD’s decision to develop processors with integrated FPGAs is nothing entirely new. Intel already tried to develop such chips after buying Altera at the end of 2015. Back in 2014, Intel announced work on a CPU with an embedded programmable matrix and even showed a test chip. In the end, however, such chips came out in a limited experimental form, and over time the project stalled.
AMD has yet to reveal details about its FPGA CPUs, but the company’s patents suggest that integrating FPGAs on the same substrate with processor cores can be done in a number of ways.
While Intel used standard PCIe lanes and the QPI bus to connect its FPGA chip to the CPU, AMD is exploring 3D layout options. One possibility is to place an FPGA die over the chiplet with processor cores, similar to how additional 3D V-Cache crystals are installed in the EPYC Milan-X and Ryzen 7 5800X3D chips. But in this case the two crystals heat each other up and can accordingly offer lower frequencies. Therefore, the second option looks much better – AMD can install an FPGA on a crystal with IO interfaces.
In fact, AMD has many other options for using the FPGA. They can be placed on the processor textolite itself, like ordinary chiplets with CPU cores. AMD also has the ability to offer customers non-standard, semi-custom solutions assembled from CPU and FPGA chips upon request. Other types of chiplets can be added there, from GPUs and DPUs to specific DSPs and even ASICs.
Still, AMD is known for its semi-custom solutions, and the demand for custom chips in the data center industry continues to grow. So the new AMD proposals in this area have every chance of success. Let’s add that AMD can and will still release FPGA-based discrete accelerators – such solutions have been in the Xilinx range for a long time.
Overall, AMD’s decision to offer FPGAs for AI could give the company an edge over NVIDIA and Intel. The latter offer more or less universal solutions for AI, while FPGAs can be fine-tuned for specific tasks and reconfigured for other tasks if necessary. But as always, the program component will play a key role. And AMD’s management has confirmed that the company will leverage Xilinx’s software expertise to optimize the custom software stack.