What are the three eras of FPGA development?
Field programmable gate array (FPGA) has always been an alternative species in the semiconductor market. It has evolved from only a few logic gates that can be used to simulate other hardware and then run software; while FPGAs have been following their own evolutionary path, they are also being used as general-purpose central processors ( CPU) and other technologies developed by customized application specific integrated circuits (ASIC).
At the "NextNext FPGA Platform (TheNextFPGAPlatform)" event held at GlassHouse in San Francisco on January 22, 2020, this became one of the topics of our conversation with Achronix Vice President of Product Planning and Business Development ManojRoge. Founded in 2004, Achronix is a relatively new player in the FPGA market, and it is currently gaining stronger development momentum than other larger competitors.
Roge is no stranger to the products and strategies of important FPGA players. From the mid-1990s to the mid-2000s, Roge was the product line manager of Cypress Semiconductor, responsible for managing the corresponding product lines in different management positions. In 2005, Roge served as senior manager of product planning for Altera, in charge of the company's 65nm and 40nm products and marketing strategies. At that time, Achronix was only recently established. Five years later, Roge became Xilinx's product planning director, responsible for the company's product roadmap and ecosystem. Four years later, he became the director of planning and marketing driving his data center and communications product lines. In the summer of 2017, Roge got his current position at Achronix, where he was responsible for the strategy and roadmap of another FPGA supplier. This experience gave him a unique perspective, and Roge was very excited about it, instead of feeling unintentional repetition.
"This field can bring you a lot of excitement, for several reasons," Roge told TheNextPlatform magazine. "The first point is that the key workloads in the data center are changing at a rapid and new pace, and new features such as machine learning are being added. Everyone is looking forward to exponential power reduction and cost-effectiveness improvements. This is because you can't just expand the size of the data center by deploying more and more CPUs. I think the current industry consensus is that you need a heterogeneous accelerator. FPGA is an option to keep your data center in the future. At the time, deploying FPGA technology can bring many advantages."
This kind of plasticity-we have heard many people describe FPGAs as "dynamic ASICs", and we prefer to call them "dynamic virtual ASICs" because they usually behave as programmable virtual instruction sets, A certain number of CPUs, these are also the key functions of FPGA. Whenever you have workloads that require frequent changes, and they require low-latency and parallel computing, FPGAs are the computing platform you should consider.
From Roge's perspective, FPGAs have gone through three eras, and we are still in the third era that began only a few years ago.
"The FPGA1.0 stage from the 1980s to the mid-1990s was mainly about logic units connected together," Roge said. "The products of this era include low-end CPLD and PLD devices, and customers often used them as logic units connected together, or some programmable I/O devices. Generally speaking, their total market demand is 10- Between 2 billion US dollars. And from the mid-1990s to about 2017 is the FPGA2.0 stage, we call it the connection wave period, in this stage many FPGAs are used to implement interfaces for networks and storage. Everyone has always loved Ethernet and DDR memory, as FPGA density and performance continue to increase, they are used for some complex functions, not just logic connected together. These applications expand the total market demand of FPGA to about 50 One hundred million U.S. dollars."
Depending on how you calculate, the total sales revenue of major FPGA vendors including Xilinx, Altera/Intel, Lattice Semiconductor, and Achronix has reached approximately $6.5 billion, so they now achieve a much larger market demand. How big is the subject of the discussion? This is one of the reasons why we host this FPGA-centric event.
Roge said: "From about 2017, we entered the new era of FPGA3.0, and its core content is data acceleration. Now, FPGAs are no longer just used for some connected logic or prototyping, they have become themselves The computing engine has been deployed in large quantities in data centers for applications such as Microsoft Azure and Amazon WebServices." Now the conservative estimate of total FPGA market demand is $10 billion, and it may also be 2-3 times this number.
The question now is how to convert this total market demand into real demand, and this requires rebuilding hardware and software.
"Whether you are building systems such as high-performance computing (HPC) clusters or analytical functions for very large-scale data centers, they can be simplified in principle into these three elements: very efficient computing, tiered storage architecture and Bandwidth, and efficient data transfer, and we have been seriously studying how to optimize them. Now, we have adopted TSMC's 7nm process technology, so we will benefit from it and use the same as Xilinx's latest products. Technology; and Intel uses 10nm technology in the latest Agilex products, which is roughly the same as TSMC’s 7nm technology. However, the difference between our products lies in the architecture and the performance we use. We decided to focus on several types of application work types. At the same time, we Hope to provide industry-leading products, that is, the best performance in unit power consumption and unit cost." Roge said.
"We have repeatedly considered all aspects of these applications, and innovated from the architecture, so that customers are more inclined to choose our products even in the face of Xilinx and Intel. We are not worried about direct comparison with their products on individual items- —Logic unit, DSP unit block, memory unit block, etc. Teraops is a number used for marketing, it really does not mean anything. What really matters is how you can accelerate The entire application is precisely because you already have the appropriate storage bandwidth and efficient machine learning computing engine, so we can provide an excellent end-to-end application performance, such as the best graphics recognition performance." Roge added.
We can say that storage bandwidth, I/O bandwidth, and Teraops (or Teraflops) are the landmark indicators of any device, and how you can use hardware and software to form a unique combination to achieve these indicators is really the ultimate event. This is our alternative view when defining FPGA devices, so that these devices can find long-lasting but constantly changing applications in data center and other fields.
If you want to know more, our website has product specifications for FPGA development, you can go to ALLICDATA ELECTRONICS LIMITED to get more information