How did 1,500 engineers in 4 years transform the entire FPGA industry?

On January 29, 2018, Xilinx, the world's largest FPGA manufacturer, welcomed its fourth CEO in history, Victor Peng, who is also the first Chinese CEO in the company's history. For Xilinx, Mr. Victor Peng is not a newcomer.

Since joining Xilinx in 2008, Victor Peng has led the strategic transformation and technology transfer of Xilinx's products and services, achieving leadership in three generations of process products from 28nm, 20nm to 16nm and major technological breakthroughs in integration and programming.

It can be said that Victor Peng has established one industry first after another by relying on his outstanding personal abilities and industry insights, combined with Xilinx's excellent R&D and technical capabilities.

However, Victor Peng's ambitions do not stop there.

Now, as the new leader of Xilinx, Victor Peng came to China for the first time after taking office as CEO and shared his views on Xilinx's future plans, changes, and the huge demands and challenges the company is facing.

In Victor Peng's view, Xilinx is in an era of great change.

This judgment is based on the vigorous development of big data and the rapid rise of artificial intelligence in the past two years, which have led to the rapid development of wave of technological waves, with new processes, new technologies, and new chips emerging in an endless stream.

Among them, the chip market, as the core and source of technology and application innovation, not only provoked a competition for the trend of artificial intelligence, but also triggered a chip war among GPU, FPGA, ASIC, etc.

On the smoke-filled battlefield, system, chip, and application companies are fighting one after another. We have not only witnessed the cruelty of the chip war, but also foreseen a future full of infinite possibilities.

In this era, Victor Peng, the new CEO of Xilinx, pointed out three major trends for future development.

The first trend is the explosion of data.

Life is full of all kinds of data, but what is lacking is the means to discover and collect data. The development of the Internet of Things industry and the advancement of sensor technology have not only proposed the possibility of collecting massive amounts of data, but also provided markets and application scenarios for the application of data.

Victor Peng believes that application scenarios such as cities, homes, factories and cars generate massive amounts of video data and image data. When the technology is not mature, these data are like chicken ribs, which are tasteless to eat and a pity to throw away.

However, when technical means are able to fully process these scattered, disordered and unstructured data, massive amounts of data will generate unlimited value.

However, in order to fully tap the value of this, Victor Peng believes that traditional CPU chips are obviously not able to support it. "We need a completely new architecture that can support greater data throughput, shorter latency, and provide the best real-time computing capabilities."

The second trend is the dawn of artificial intelligence.

History tells us that the arrival of new technologies will inevitably change our originally peaceful times.

Just as the Industrial Revolution promoted industrial transformation, artificial intelligence is changing many traditional industries and creating many new businesses and business models.

Victor Peng clearly realizes that artificial intelligence is not an application, but an emerging technology that can be perfectly combined with multiple applications. Although this technology is still in its early stages of development, it does not affect its ability to drive the creation and development of new applications.

From point to surface, from application to cloud and then to edge, artificial intelligence is becoming the most promising driving force for development in the next few decades with its unimaginable magic.

The third trend is computing in the post-Moore's Law era.

Moore's Law is more of an economic law, but even from an economic perspective, Moore's Law has become as old as the years since its birth, and its development speed has slowed down significantly. If we want to continue development, we must find new development directions from the pulse of the times, which is what we call the post-Moore's Law era.

In Victor Peng's view, the demand for computing power in the post-Moore's Law era has not only not weakened compared to the Moore's Law era, but has become increasingly intense.

"Especially in the Chinese market, the pace of industry innovation is getting faster and faster, and companies from all walks of life are increasingly demanding new products, new services, and new business models. Traditional chip solutions can no longer meet the needs of the industry." Victor Peng even lamented that this is a situation he has never seen in his decades of working in the industry.

However, as a senior industry expert who has been in the industry for decades, I believe that there are greater opportunities in this change.

Victor Peng believes that the high computing demands in the post-Moore's Law era mean that more heterogeneous systems are needed, and the traditional single CPU is difficult to meet all applications.

These three trends seem to have been emphasizing that in the era of artificial intelligence, heterogeneous architecture will be the mainstream trend in the future.

However, just as we study the industry, we need to look beyond the phenomenon to the essence. What does the arrival of the era of heterogeneous architecture mean?

As the CEO of Xilinx, Victor Peng is not limited to a single product and market, but has a longer-term vision.

"This is an exciting era. Our expectations from decades ago are gradually becoming a reality, and a smart, connected world is being shaped." Victor Peng emphasized that in this process, the most important contribution Xilinx can make is to make the world's transformation more adaptive, "because only by being adaptive can we build a global-scale smart, connected infrastructure faster."

Realizing an intelligent, connected, flexible and ever-changing world is not only Victor Peng’s long-term vision, but also Xilinx’s long-term vision as an FPGA company.

Over the past three decades, Xilinx has always held on to this expectation, adhered to its beliefs, adopted new strategies when faced with new opportunities, and strived to build an adaptive world.

To achieve this vision, Xilinx has formulated three major development strategies.

Victor Peng said, "Our strategy is to prioritize data centers and continue to focus on developing mainstream market applications including communications, automotive, and industry."

Different companies will adopt different development strategies for the same trend. Strategic analysis of different companies will be more conducive to our in-depth observation of this market.

Next, let us analyze these three development strategies and analyze Xilinx's development ideas.

Xilinx's first strategy is "data center first".

Regarding why this strategy was adopted, Victor Peng said that the reason why Xilinx insisted on focusing on data centers was not only because data centers are a very broad and rapidly developing market, but more importantly, Xilinx saw that the disruption of many industries occurred in data centers, and the popularization of many technologies also originated from data centers.

As Victor Peng said before, heterogeneous architecture will be one of the future trends, and this is also true in the data center field.

Many new architectures, whether NVIDIA or Intel, are no longer CPU-centric. A very obvious phenomenon is that the development trend led by CPUs in the past few decades is shifting to heterogeneous computing with many accelerators.

But for Xilinx, the data center market is a new market after all, and it is very different from other markets in terms of demand and development direction.

Adhering to the principle of seeking truth from facts, Xilinx has adopted a completely different support approach in the data center market. Victor Peng said that Xilinx has changed its past technical direction of mainly supporting hardware developers, and in the data center field, it supports software developers more so that more users can use Xilinx's adaptive computing acceleration platform.

Specifically, Xilinx mainly provides a software stack development environment in this field. It provides a development environment based on the standard PCIe FPGA acceleration card at the bottom to support software engineers to program with CC++ and OpenCL.

Xilinx also cooperates with third parties to optimize acceleration libraries and provide them to customers, thereby helping customers solve a series of tasks such as machine learning. For example, we have very good cooperation with DeePhi Technology, Alibaba, Amazon, Huawei, etc.

In addition, Xilinx also supports a wide range of AI applications. DeePhi Technology provides an efficient FPGA-accelerated speech recognition engine, which has 43 times the performance of the CPU and 40 times the performance per unit power consumption. It can not only accelerate machine learning and artificial intelligence, but also accelerate video processing, video transcoding, data retrieval, and compression. In addition, some companies have used Xilinx products to make major breakthroughs in genome sequencing and analysis.

Of course, in addition to the artificial intelligence market, Xilinx has not given up on other markets.

Victor Peng said that there will be some major changes in markets such as automobiles, wireless infrastructure, and wired communications, which will have a huge impact on our lives, and these markets are the mainstream markets that Xilinx is focusing on.

For example, through the widespread use of sensors and cameras, we can enhance urban security and factory inspections, equipment testing, communications, and even consumer electronics products.

Among them, Xilinx can participate in the process from endpoint to edge to cloud. Victor Peng believes that this is also a unique advantage of Xilinx, "it can truly achieve full coverage from endpoint to edge to cloud."

To this end, Xilinx has launched a new series of product types, from FPGA to ACAP (Adaptive Compute Acceleration Platform).

Victor Peng said that ACAP is a new product type, which is a major technological disruption to the industry and the most outstanding engineering achievement since the invention of FPGA. "It can cover a wider range of market applications and meet the needs of flexible workloads."

According to Victor Peng, ACAP, the full name of which is Adaptive Compute Acceleration Platform, translates to “Adaptive Compute Acceleration Platform”.

This is a highly integrated multi-core heterogeneous computing platform that can flexibly change the hardware layer according to the needs of various application workloads.

"ACAP is an ideal choice for adaptive and optimized design," Victor Peng believes that it can even cope with new changes that will appear in the future, such as interconnection and acceleration with new applications and new devices. "In addition, ACAP can not only be programmed in software, but also support the development of hardware developers like traditional products."

Xilinx announces a new breakthrough product called ACAP

To illustrate this development model, Victor Peng demonstrated a product use case for ACAP, which is the action code-named "Everest" currently being promoted by Xilinx. "Everest" will be Xilinx's first ACAP series.

It is understood that the "Everest" operation uses TSMC's 7nm process. Xilinx has spent four years developing the main architecture, with a cumulative R&D investment of more than US$1 billion. A total of 1,500 engineers are involved. The products are mainly mid-to-high-end products, among which high-end products have more than 50 billion transistors.

Compared with 16nm products, Victor Peng said that ACAP can achieve dynamic optimization for workloads.

"ACAP is a new product type. Xilinx is the first company to launch ACAP products, and this is also our core competitiveness." Victor Peng believes that as time goes by, ACAP will become a mainstream platform and will become a basic platform for building the future world like CPU and GPU.

As an FPGA company, why did Xilinx launch such a computing platform?

Victor Peng explained that although it will take a long time for the new computing platform to become popular, it also marks Xilinx's ambition to become a completely different company that can launch new and different types of products, rather than just being an FPGA company.

From this perspective, Victor Peng and Xilinx made the right choice. Historically, many companies have transformed themselves by launching different products and new series, and have embarked on a new development path and achieved glory.

Therefore, Victor Peng believes that the launch of this new product is a very important moment for Xilinx today, and this platform will help Xilinx become a brand new company.

At the same time, Victor Peng also believes that ACAP can achieve superior performance and power consumption compared with traditional CPU and other solutions.

However, people who are familiar with fields such as artificial intelligence know that in these fields, in addition to FPGA, there are also many other types of chips such as traditional CPUs and GPUs.

Different companies have different views on different types of chips. Some companies believe that GPUs can be successful, while others believe that CPUs can be successful.

It can be said that each person has his own opinion and it is difficult to reach a consensus.

In response to this issue, Victor Peng gave us a detailed analysis of the advantages and disadvantages of chips on various platforms.

Taking CPU and GPU as an example, although GPU is more applicable to many emerging applications in some aspects, it still cannot adapt to all application environments.

As mentioned before, the future is a world of intelligent interconnection, and such a world cannot be handled by a single architecture. Heterogeneous architectures will prevail, and the architecture that can adapt to the flexible and changing market will be able to dominate the entire general market.

"The key point is that it is highly flexible and has a high degree of adaptability." Victor Peng believes that the so-called high adaptability means that it can accelerate different workloads rather than just a single application.

However, we can see that the CPU can only handle some fixed tasks, and the acceleration performance is not obvious. Although the GPU performs well, it can only show impressive results in the acceleration process of machine learning. Of course, it is still far from the ideal effect.

The effect of GPU on other applications is even worse. If ASIC architecture is adopted, "Victor Peng said that although such fixed-function chips can be customized, when faced with a new application, customized design must be carried out from scratch before entering the market.

The so-called adaptive chips need to be able to adapt to a variety of new workloads, and even many new applications that have not yet appeared. Only chips that meet this requirement can be truly adaptive chips.

Victor Peng also emphasized that although in some applications, the effect of FPGA acceleration can be seen to be better than GPU or even ACAP, the disadvantage of doing so is that after a period of use, the performance of the chip is fixed. When a new algorithm emerges, the chip needs to be replaced for deployment. The core of ACAP is a new generation of FPGA architecture, which combines distributed memory with hardware-programmable DSP modules, a multi-core SoC, and one or more software-programmable and hardware-adaptive computing engines, all interconnected through an on-chip network (NoC). ACAP also has highly integrated programmable I/O functions, which range from integrated hardware-programmable memory controllers to advanced SerDes transceiver technology, cutting-edge RF-ADC/DAC, and integrated high-bandwidth memory (HBM) depending on the device model.

Software developers will be able to use C/C++, OpenCL and Python software tools to apply ACAP systems. At the same time, ACAP can still be programmed from the RTL level using FPGA tools.

In the era of the rapid rise of big data and artificial intelligence, ACAP is suitable for accelerating a wide range of applications, including video transcoding, databases, data compression, search, AI inference, genomics, machine vision, computing storage and network acceleration. Software and hardware developers will be able to design ACAP-based products for endpoint, edge and cloud applications.

As mentioned before, the first ACAP product series will be a product series code-named "Everest" developed using TSMC's 7nm process technology, which will be taped out at the end of this year.

Victor Peng said that over time, the ACAP product line will gradually sink into low-end market applications, especially those markets that are more sensitive to cost and power consumption.

Victor Peng cited the example of the automotive industry, where Xilinx has served for more than a decade. In Xilinx's view, the automotive industry is very sensitive to price and power consumption, but the ACAP product series can already serve the automotive industry very well, so it can also solve the problems of other industries that are also sensitive to price and power consumption.

However, Victor Peng admitted that there are still many uncertainties and challenges in future market changes, and Xilinx needs to continue to develop more complete product solutions rather than implement closed product solutions.

"Through continuous contact with customers, we found that customers like our products because they can differentiate their own products through Xilinx products, rather than launching products that are essentially the same based on closed products."

Victor Peng emphasized that Xilinx hopes to continue to serve mainstream customers by continuously improving open product solutions, providing more complete platforms and easier-to-use solutions to support more hardware and software developers. This idea will continue to transform the FPGA market pioneered by Xilinx!

Text/Semiconductor Industry Observer Liu Yi

Today is the 1533rd issue of content shared by "Semiconductor Industry Observer" for you, welcome to follow.

Follow the WeChat public account Semiconductor Industry Observation , reply to the keyword in the background to get more content

Reply Panel , read more articles about the panel industry

Reply Bitcoin , read more articles related to Bitcoin and mining machines

Reply Wafer , read articles related to wafer manufacturing

Reply Silan Micro , read more articles related to Silan Micro

Reply to ISSCC , see "The Trend of Semiconductor Industry from ISSCC Papers"

Reply Huawei , read more articles related to Huawei

Reply Packaging , see more articles related to packaging technology

Reply to A shares to see more articles related to listed companies

Reply Exhibition , see "2017 Latest Semiconductor Exhibition and Conference Calendar"

Reply Submit your article and read "How to become a member of "Semiconductor Industry Observer""

Reply Search and you can easily find other articles that interest you!

Click to read the original article to learn more about Moore's elite