NAVER and SAMSUNG signed a MOU to develop AI semiconductors. A year later, they unveiled the result that AI semiconductor they developed is focused on processing data faster than competing AI semiconductors when running LLMs. Their choice of the AI semiconductor is an “FPGA”, a type of programmable non-memory semiconductor that is optimized to be more power efficient. Why NAVER and SAMSUNG choose “FPGA” for AI semiconductor? In this post, I’d like to introduce what is FPGA, which are less familiar than CPUs and GPUs, what advantages they have.

What is FPGA?

FPGA (Field Programmable Gate Array) is a type of non-memory semiconductor that can be programming. Unlike general semiconductors that cannot be changed or are restricted from change, FPGAs can use and delete logic and semiconductor circuits like programs, and it has all of the advantages of quick and efficient ASICs (Application Specific Integrated Circuit) and with excellent versatility, it also has the advantage of CPUs and GPUs in that it can run various programs.

However, FPGA still less well-known than CPUs and GPUs because it was being used as a tool to verify that the circuits worked before you design ASIC. ASIC is very expensive to manufacture, so it needs to verify working properly before produce in large quantities. FPGA was used for verification for a while and in a limited type of industries, so it is less well-known than CPUs and GPUs.

As the usage scope of everyday computers has expanded, more users wanted efficient computer resources in more places, and while using ASIC is efficient, it was difficult to satisfy the needs of “various purposes” and “various applications.” Furthermore, while ASIC has the versatility of CPUs and GPUs, effective calculations were difficult, so FPGAs were gaining attention for their versatility to run several programs, along with their speed and efficiency.

“FPGA", a rising AI semiconductors.

All cloud companies including AWS and MS Azure started to provide FPGAs as computing instances, and Intel and AMD have recorded the highest quarterly sales after acquiring the world’s top FPGA manufacturers, Altera Corporation and Xillinx respectively.

However, from the perspective of those who design semiconductors, there are apparently some slightly unfortunate aspects of FPGA gaining attention. Designers want to add more features and faster features to FPGAs to provide a better product to users, but as it becomes an established product that is mass-produced, features such as the size of circuits and operating speed become limited, leaving designers wanting more. However, just as “next generation” CPUs and GPUs have become upgraded with more features, it is predicted the FPGAs will also follow a similar route.

BLUEDOT, where the Top FPGA semiconductor designers are gathered.

BLUEDOT develops semiconductor IP for video solutions, specifically trained video AI by self and designed on FPGA for fast, a level of real-time processing speed. Video AI, which is nearly impossible to process in real-time due to its massive computation, but BLUEDOT developers came together to design semiconductor IP to accelerate processing without sacrificing performance. As a result, BLUEDOT won first place in the world FPGA competition held by Xillinx, leading global FPGA manufacturer (now acquired by AMD), it is recognized how BLUEDOT’s semiconductor IP is remarkable.

In currently, BLUEDOT showcases technologies that enhance video quality and increase encoding efficiency. They have developed a total of five video technologies in semiconductor IPs from “DeepField-SR” that can be upscaled quickly in real-time, “DeepField-PQO” that reduce data size while maintaining picture quality, “Pulsar-AV1” and “Pulsar-AVIF” which are gaining attention as next-generation codecs for their compression efficiency and increased encoding speed of AV1 and AVIF, and “Pulsar-VMAF” which increases the efficiency of VMAF quality assessment solution by 4 times, and they are continuously researching and developing faster and more efficient media solutions.