As the growth of AI industry is continued, NVIDIA is streets ahead in the AI semiconductor industry. Especially as generative AI has democratized faster, NVIDIA’s golden age is going to last for a long time. NVIDIA’s “H100” is expected to be in high demand going forward, and other semiconductor companies such as Intel and AMD are attempting to expand their share of the AI semiconductor market by introducing semiconductors that can replace NVIDIA.

Intel, AMD’s hidden cards in AI semiconductor industry.

It looks like Intel and AMD used “FPGA” cards to target AI semiconductor market. In September, Intel announced that it had “Expanded FPGA portfolio”. Intel unveiled its new FPGA line-up at Intel FPGA Technology Day (IFTD), saying they have enhanced capabilities to help developers build solution faster, and that they meet all levels of programmable logic requirements using FPGAs. Intel has demonstrated that it is increasing its investment in FPGA, hoping to meet enhanced AI capabilities with an expanded FPGA portfolio.

AMD also developed AI semiconductors using FPGA from its subsidiary Xilinx, and Microsoft has developed an FPGA-based AI accelerator which is applied to MS Azure(Microsoft’s cloud service). In addition, not only global IT companies, but also Samsung and Naver have chosen FPGAs as AI semiconductors to drive their AI models, and have announced that they satisfied performance they expected.

Can FPGA be used as AI semiconductors?

Why are so many semiconductor companies, including NVIDIA, focusing on developing “AI semiconductor”? it makes possible to perform the huge computations required by AI very efficiently. Although AI semiconductors are not universality type like CPU or GPU, they are specialized in AI computation and processing efficiently, so the demand for AI semiconductors will continue to grow as AI becomes more commonplace.

FPGS is estimated to be a suitable AI semiconductor because of ‘latency’. When you designed AI network directly in hardware, you can apply “dataflow architecture” in FPGA, so the lower latency, the better. This is why FPGA is a good alternative for AI semiconductors. In fact, AMD announced the comparison performance of NVIDIA’s GPU vs. AMD’s FPGA is primarily focused on “low-latency” applications.

The possibility of FPGA, Field Programmable Gate Array.

FPGA’s full name is “Field Programmable Gate Array”, it is a programmable and non-memory semiconductor. They can be reprogrammed logic circuits, semiconductor circuits, and they handle specific functions quickly and effectively. So FPGA were being used as a tool which verify the circuits designed on ASIC is working properly. However, many semiconductor companies have begun to mass-produce FPGA in recently because they not only have the versatility to run a wide variety of programs like CPU and GPU but also provide fast and efficient processing for specific functions.

BLUEDOT developed video solutions that runs on FPGAs. We focused on “Fast and Efficient processing of specific functions” in FPGA, so our solution can process video data fast and efficiently enable to the level of real-time processing. The solutions are AI based video quality enhancement & upscaling, pre-processor technology to increase video compression efficiency, and high performance AV1-AVIF codecs. In 2020, BLUEDOT’s FPGA were recognized by winning first place in an FPGA competition held by the NO.1 FPGA company, Xillinx (now subsidiary of AMD).

I think FPGA will make more use cases. Especially, the grow of AI semiconductor market fast, alternatives to NVIDIA’s GPU will be appeared continuously include FPGA. It is not be an absolute powerful substitute, but it have the potential to play the role of a hidden card in the system semiconductor industry.