Photo by Harrison Broadbent on Unsplash
THE POSITIVES OF SEMICONDUCTOR POWERED FPGA
Year on year, for decades, Moore’s law has inspired semiconductor designers and manufacturers to pack more devices in the same silicon area. It has thus provided the computing world with different types of semiconductor products to process the data as per the power-performance goals.
One semiconductor product that has benefitted the most from Moore’s law is the XPU. XPU are of different types, and the widely known ones are CPU (Central Processing Units) and GPU (Graphical Processing Units). Apart from traditional XPUs, ASIC (Application Specific Integrated Circuits) have also enabled elegant computing solutions for decades.
ASICs are also of different types, and the most advanced ASIC is the FPGA (Field-Programmable Gate Array). FPGA is not new to the semiconductor market and has played a crucial role in industry and academia. One of the advantages of FPGA is the ability to get programmed/configured to prove-in different methods and that too after manufacturing. It has thus enabled several industrial applications from audio processing to wireless communication to rovers on Mars.
Adaptive: The design features of FPGA make it highly adaptive, and this provides the ability to deploy FPGA as per the target application requirements. It is possible due to the array of programmable blocks, which empowers developers to use (and also enhance) the FPGA as per the target application requirement.
Scalable: FPGA comes equipped with silicon features that enable them to be more scalable. Such features make them more adaptive than any traditional XPU. Such reconfigurability ability is possible due to the extensive libraries and IPs that the FPGA ecosystem has.
Embeddable: Several advanced computing systems often require dedicated ASIC that can take over the specialized task. This ASIC is often embeddable and comes equipped with the XPU system. In the same line, FPGAs are also embeddable and often carry out the specialized task on behave of the host system. FPGA being configurable are thus becoming a de-factor for the data learning world.
FPGA’s importance is growing further due to the adaptive/programming features that make it a perfect silicon for training a large amount of data on the go. ML/AI/DL is another area where FPGA has found a growing use. It is also one of the reasons why both Intel and AMD have invested billions of dollars to acquire FPGA-focused assets. These assets have enabled Intel and AMD to drive next-gen XPUs with embedded FPGA for specific post-manufacturing programmable requirements. Example applications can range from neuromorphic computing to inference-based processing to real-time metaverse.
Today, the FPGA market is approximately ten billion dollars. With the demand for real-time adaptive silicon increasing, the FPGA market will grow further and thus presents opportunities for both established and emerging XPU focused semiconductor companies.
THE NEGATIVES OF SEMICONDUCTOR POWERED FPGA
The features of FPGA do not make it an alternative for traditional XPUs. One of the fundamental reasons for the fast growth of XPUs like CPU and GPU is the ease of use, and this is something that is not possible with FPGA due to the target application and the design/IP complexity.
However, in the last couple of years, the story has changed. Today, performance-oriented XPU comes equipped with FPGA (embeddable-FPGA or eFPGA). One important reason is to allow training/testing of learning algorithm on the device itself. Such silicon features are going to form the base for advanced X-Reality products.
Like any other computing system, FPGAs also have drawbacks. The majority of these drawbacks are not going to stop the growing FPGA market. However, overcoming these fundamental challenges can speed up the deployment of FPGA on critical infrastructure.
Programming: Programming FPGA is not easy and demands detailed in-architecture understanding. Companies developing FPGA often provide a set of software libraries and tools to make the programming experience easier. However, for the FPGA to become an attractive platform, the programming language needs to evolve. It might enable faster algorithm development to drive cost-effective solutions.
Dependence: There are several application areas where FPGA cannot get deployed without a host system. It makes them dependent on the host XPUs. FPGA developers have tackled this issue by utilizing low-power SoC, but the cost and complexity it comes with are high.
Manufacturing: FPGA is very complex, and embedding them with any other XPU makes the manufacturing process harder. Semiconductor manufacturing has advanced a lot, but driving high yield is still a big challenge for advanced nodes with highly complex designs like SoC FPGA. As FPGA becomes part of a more complex computing system, the manufacturing methodology will have to evolve too.
Neural Processing Units (NPU) are becoming a given feature for every new XPU. Intel, AMD, Apple, Microsoft, Qualcomm, NVIDIA, and several other companies are increasingly adapting their XPU to have NPU blocks. The building block of NPU is still FPGAs, and this makes it critical to overcome all the FPGA-related challenges to ensure the users can make the most of the FPGA ecosystem and are driving the adoption at a faster rate.
The semiconductor giants have already acquired long-term assets to make their XPUs FPGA friendly. These semiconductor companies are already delivering FPGA-inspired solutions and are pushing the computing industry to the next level. It showcases that FPGA will keep growing, and the semiconductor industry will keep coming up with advanced FPGA silicon.
