After the wave of acquisitions, we are entering an era of heterogeneous data processing platforms. This platform will include SmartNIC, multi-core processors and hardware accelerators. Users will have to choose a complete solution instead of a mashup. Intel, AMD and Nvidia will provide a complete computing platform, including their own proprietary accelerators and smart chips.
The benefits of this may be easier deployment and better integration, but choosing each of the best options will no longer be feasible, such as choosing Xilinx FPGAs, Intel Xeon processors and Nvidia GPUs.
For Intel and AMD, it will be difficult to promote FPGAs in the field of deep learning, and other platforms will appear under their own brands. Perhaps FPGAs will be mainly used in their basic markets (networks and telecommunications, such as vRAN and 5G), while GPU, asic and other platforms will be promoted for deep learning.
The main advantage of FPGAs is programmability and can support custom architectures. This means they can adapt to new algorithms or applications faster. This competitive advantage is crucial, especially in the field of deep learning where new models are developed by ML engineers and data scientists. FPGA can be programmed with new customized models/algorithms, which has better performance than other platforms. Especially in applications that require bit-level processing such as packaging processing, genomics, and Bitcoin mining, FPGAs therefore show better performance than other platforms. In deep learning applications, FPGAs can provide lower latency and high performance, especially when using fewer bits. Therefore, FPGAs can play an important role in the new processing ecosystem, not just for smart network card applications.
in any case, in order to make FPGA a more attractive accelerator platform, we also need to provide the required framework to make its deployment simple and scalable. This is why a vendor-agnostic framework is needed to make FPGAs as easy to deploy as GPUs or CPUs.
In the field of embedded systems, FPGAs still dominate. Xilinx and Intel both provide socket-based FPGAs, and their ARM cores are widely used in embedded applications. Nvidia said that third parties will still allow the use of ARM cores, so Xilinx and Intel will continue to use ARM in their socket-based FPGAs until they switch to RISC-V processors in a few years.
The FPGA community is quite large and growing. Conferences like FCCM, FPGA and FPL show that there are many communities that support and promote the use of FPGAs. However, it remains to be seen whether FPGAs will continue to be used as general-purpose accelerators, or how much their applications will be restricted in networking and telecommunications applications.