This billion-dollar data center market is divided between Altera, Xilinx, and other FPGA vendors. This market became even more complex after Intel acquired Altera in June 2015.
In 2014, prior to the acquisition, 16 percent of Altera’s $1.9 billion in revenue came from its data center-related computing, networking and storage business, which was worth $304 million. Those communications and wireless device system manufacturers who have been in this space for a decade or two want higher energy efficiency, lower cost and higher scalability, all areas where FPGAs excel. Another point that needs to be mentioned is that in performing these functions, the use of FPGAs does not require an operating system and corresponding software as it does with CPUs. This segment accounted for 44 percent of Altera’s revenue, totaling $835 million.
Another 22 percent of Altera’s revenue, or $418 million, comes from areas such as industrial controls, military equipment, and automotive manufacturing. They face the same dilemma and therefore choose FPGAs to handle some of their workloads.
In fact, back in 2014, Intel looked at a potential market of $115 billion worth of various types of chips. Among them, editable logic devices (mainly FPGAs) accounted for about 4%, ASICs accounted for 18% and the rest was a hodgepodge of ASSPs.
In the field of editable logic devices, Intel predicted that Altera accounted for 39% of the $4.8 billion market, Xilinx accounted for 49%, and the remaining suppliers accounted for the remaining 12%.
The reason Intel did not acquire Altera at the time was that the FPGA business was growing almost as fast as its data center group, which provides chips, chipsets and motherboards to server, storage and switch manufacturers.
Further, Intel didn’t do so because Moore’s Law was slowing down to a crawl, posing a growing competitive threat to FPGAs.
In fact, if applied, there is more than one FPGA, GPU or DSP gas pedal installed in the data center, but not multiple Xeon CPUs. since Intel could not continue to provide more cores and gas pedals for Xeons, they came to the conclusion to use FPGAs as gas pedals.
Unless FPGAs can generate $500 million in revenue in the data center, or $1 billion or more in a few years. Otherwise Intel would rather sacrifice two to three times the Xeon revenue than give it away.
Deep learning, FPGA prospects are promising
According to Intel’s projections, they plan to ramp up their FPGA business at near-linear growth rates between now and 2023. We are always skeptical about this. But the FPGA business is growing more or less over time (about 2.5 times more than 15 years ago).
Intel also expects that FPGA revenue will double again between 2014 and 2023. According to Intel’s forecast, its revenue should be slightly less than the projected $8.9 billion with a compound annual growth rate of 7 percent between 2014 and 2023. Interestingly, since Intel’s forecast does not include the share of FPGA revenue from data computing centers (servers, switching and networking) in the plan, this will change significantly. Let’s analyze.
If Altera’s and Xilinx’s market shares do not change, and assuming Altera’s revenue stays the same in the networking, compute, and storage segments, Altera’s revenue from this part of the business will reach about $560 million by 2023. We think Intel’s figures underestimate the pressure on data centers to deliver more efficient and flexible computing. But we think the outlook for FPGAs is far better than this forecast. That said, many proponents of FPGA technology have been expecting the day to come soon when FPGAs will gain computing legitimacy in the data center.
It is ironic that Intel itself, an expert in programming FPGAs, a user of hardware description languages, and a well-known ASIC manufacturer, should be a major player in the push to make FPGAs the preferred choice for gas pedals. Such gas pedals can be used both as standalone discrete computing elements and as hybrid CPU-FPGA devices.
That’s why all the news we’ve seen about Altera since 2016 is indicative of massive growth in FPGAs. So at least in the short term, they have little choice but to be a graft for other FPGA makers.
This acquisition is not only a milestone in FPGA development, but also an acknowledgement by Intel of the enormous potential of FPGAs as the powerful computing gas pedals of the future, not only influencing major enterprise decisions and market trends, but also accelerating workloads in the enterprise, facilitating internal searches in hyperscale data centers, and raising the profile of high-performance computing simulations.
As we span 2016, the addition of machine learning and deep learning to FPGAs in the application middle class has put another hammer on the FPGA industry.
Why Everyone Favors FPGAs
First, the software stack for programming FPGAs has evolved, especially with the help of Altera, which added support for the OpenCL development environment. But not everyone is an avid fan of OpenCL.
First there was Nvidia, which created its own CUDA parallel programming environment for its Tesla GPU gas pedal. Then there was SRC Computers, which not only provided hybrid CPU-FPGA systems for defense and intelligence back in 2002, but by mid-2016 further brought its own Carte programming environment, which enables automatic conversion of C and Fortran programs to the FPGA’s hardware description language (HDL), into the commercial market.
Another factor driving the adoption of FPGAs is the increasing struggle to improve the performance of multi-core CPUs as chip manufacturing technology struggles to continue scaling. While CPUs have gained big jumps in performance, they have been used primarily to scale the performance throughput of the CPU rather than the individual performance of a single CPU core. (We know that architectural enhancements are difficult). But FPGAs and GPU gas pedals have made convincing per-watt performance improvements.
The GPUs run hotter and have similar performance per watt, but they also bring more power to the table.
This improved performance per watt explains why the world’s most powerful supercomputers moved to parallel clusters in the late 1990s, and why they are now moving to hybrid machines instead of Intel’s next CPU-GPU hybrid mainstay, the Xeon Phi processor “Knights Landing (KNL). KNL).
With the help of Altera FPGA co-processor and Xeon Phi processor Knights Landing, Intel can not only maintain its competitive advantage at the high end. And continue to lead the competition with Nvidia, IBM and Mellanox in the Open power alliance.
Intel believes in the rapid growth of workloads in the hyperscale computing, cloud and HPC markets. To promote its computing business continues to thrive. This situation can only become a seller of FPGAs, otherwise others will steal this only way out.
But that’s not what Intel is telling everyone. We don’t see this as a defensive war or anything else,” Intel’s CEO Brian Krzanich said at a press conference following the news of the Altera acquisition, they said.
“We think the Internet of Things and data centers are huge. Those are also the products our customers want to build. Thirty percent of our cloud workloads will be on these products, based on our predictions of how we see trends changing and how the market will evolve.
This is used to demonstrate that these workloads can be moved to silicon one way or another. We think the best approach is to use a combination of Xeon processors and FPGAs that have the best performance and cost advantages in the industry. This will bring better products and performance to the industrial space. And in the IoT, this will extend to potential markets against ASICs and ASSPs; and in the data center, it will move workload to silicon, driving rapid cloud growth.
Krzanich explained: “You can think of FPGAs as a bunch of gates and the ability to program them at any time. FPGAs can be used as gas pedals for multiple domains, can perform facial searches while encrypting, and can reprogram FPGAs in essentially microseconds. this is much lower cost and more flexible than large scale individual custom parts.”
Intel sees a bigger opportunity
Intel sees a much bigger opportunity than that.
Intel CEO Brian Krzanich’s announcement after the acquisition closed that up to a third of cloud service providers will use hybrid CPU-FPGA server nodes by 2020 was a shocker. This also gives Altera, which has been targeting data centers since late 2014, an opportunity of about $1 billion in FPGAs. That amount is roughly three times the revenue of Nvidia’s currently popular Tesla computing engine.
In early 2014, Intel demonstrated a prototype Xeon-FPGA chip in the same package and intends to launch this chip in 2017. This was based on an idea for a Xeon with FPGA circuitry presented by Diane Bryant, then GM of the Data Center Group, which was introduced shortly after.
On the conference call announcing the Altera deal, Krzanich did not specify the timing of the launch of this Xeon-FPGA device, but he did say that Intel will create a single die hybrid Atom-FPGA device for the IoT market. Intel is examining whether it needs to do a single package mix for Atom and Altera FPGAs during the hybrid transition phase.
During a conference call with Pacific Crest Securities in early 2016, Jason Waxman, general manager of Intel’s cloud infrastructure group, said FPGAs have become a hot topic when discussing with research analysts about Intel’s data center business.
First, while he didn’t name any vendor or any device specifications, Waxman determined that Intel has already provided samples of its Xeon plus FPGA hybrid computing engine to some customers.
During the conference, Waxman even talked about what drove Intel to acquire Altera and plug into programmable computing devices. Intel clearly wants to make FPGAs mainstream, even if that might cannibalize some of Xeon’s business in the data center. (We believe that because Intel believes this self-inflicted cannibalization is inevitable, the best way to control it is to make FPGAs part of the Xeon lineup.)
Waxman said, “I think this acquisition could involve a number of things, and some of them are beyond the scope of the data center group.”
First, a potential core business is often driven by manufacturing leadership. We have a good handle on that, and there are good synergies in doing so.
Again, there are IoT “groups” that have a strong interest in this.
We know that certain large-scale workload extensions (e.g., machine learning, certain networking capabilities) are attracting more and more attention. We’re just realizing that we may be able to make some performance breakthroughs, and this would be a good opportunity to port FPGAs from data center applications to more suitable, broad development areas.
But the collaboration in the data center group, FPGAs are nothing more than a companion to CPUs to help solve the problems of cloud service providers and other types of large-scale applications.
Key applications that Intel sees as having priority and high demand for FPGA acceleration include machine learning, search engine indexing, encryption and data compression. As Waxman points out, these tend to be very targeted and have no uniform use cases. This is the basis for Krzanich’s categorical statement that one-third of cloud-based service providers will use FPGA acceleration within five years.
Crossing the FPGA Barrier
While everyone complains about how hard it is to program FPGAs, Intel doesn’t back down from that. While not revealing much about the program, Waxman suggests ways to make FPGAs easier to use and understand.
What we have is unique,” Waxman said, “and that’s something that no one else can give. That’s our ability to understand these workloads and the ability to drive acceleration.
“We see a shortcut to facilitating machine learning, accelerating storage encryption, accelerating network functions,” Waxman emphasizes. That’s based on our deep understanding of these workloads, and that’s what allows us to see opportunities like this.
But there are still some difficulties that FPGAs have to face right now because people are writing RTL right now. We are a company that writes RTL, so we can solve that problem. First we make it work, and then we can lower the barrier to entry. The third step is real economics of scale, and that’s all on the strength of integration and manufacturing.
To address these barriers, we offer a range of approaches.
X86+FPGA?
For those speculating that Intel intends to replace Xeons with FPGAs, Waxman says that’s a bunch of hogwash.
For algorithms that have a strong need for high speed and repeatability, Waxman says, FPGAs with their inherent advantages are their best bet. And for those data operations and conversions that have extremely high latency requirements, FPGAs are also candidates.
Considering that Altera has already integrated an ARM processor and FPGA on a SoC, it’s natural to think that Intel would try to fully replace the ARM core with an X86 core to make a similar device. But it doesn’t look like this will happen.
First, at Intel’s Q2 2016 financial statement, Krzanich promised that Intel would strengthen its support for customers currently using Altera’s ARM-FPGA chips.
Waxman further clarified, “Our view is that there will be some form of integration of FPGAs into Xeon. We have publicly announced that we will build the first generation of devices using this single package, but we will adjust our direction as we progress, and may even implement on the same die. We will learn what the right combination is based on customer feedback.
By the way, I still expect to see systems that are not integrated, keeping in mind that they will do system level synergy. We will not integrate Xeon with FPGAs in multiple combinations, instead we will find the right target and balance in the market.”
Programming issues bear the brunt
While Altera’s toolset leverages the OpenCL programming model to obtain application code and convert it to RTL (the native language of FPGAs), interestingly, Intel does not see the future success of FPGAs in the data center as being based on improvements in OpenCL integration with RTL tools or broader adoption of OpenCL.
Waxman also says emphatically, “It’s not based on OpenCL.” While we do see OpenCL as a way to further expand the scope of FPGA adoption, initial cloud deployments of FPGAs may currently be done by more capable companies, but they are not asking us to provide OpenCL, Waxman added.
Waxman hinted that Intel has plans to make FPGAs easier to program, without being “free” to talk about it. He said Intel will provide RTL libraries for programmers to call routines deployed on FPGAs and drive the formation of gates that execute applications on them, enabling gates for application routines rather than letting them create their own routines. This makes some sense, in the same way that Convey (now a division of Micron Technology) used FPGAs to accelerate system processing a few years ago.
I think there’s a continuum of acceleration,” Waxman says. At the beginning, you may not know what you’re trying to accelerate and just do some experimentation, so at that stage of acceleration, you want a more general purpose. When you start to really want to accelerate, you want to be more efficient, lower power and less space, and that’s when you move the focus to the FPGA.”
Waxman also cited Microsoft’s use of FPGA acceleration on its “Catapult” system as an example.
The system uses its Open Cloud Server and adds an FPGA mezzanine card as a gas pedal. We studied this project in March, applying these gas pedals to the same image recognition training algorithm on Google, and the results showed that the 25-watt FPGA device improved performance/watt compared to the server using the Nvidia Tesla K20 GPU gas pedal (235 watts).
As we said, we have no doubt about the performance numbers posted by Microsoft and Google. But it’s not fair to discrete GPUs or FPGAs to execute application performance and to measure their own thermal profiles. You have to see this at the server node level.
If you realize this, Microsoft servers that get FPGA-assisted are only slightly ahead of Google servers with Tesla K20s at the system level. (These are just our estimates based on image processing performance per watt per second). In this comparison, Microsoft should not consider cost. And frankly, unlike the Tesla GPUs that come with everything, Microsoft Open Cloud Server doesn’t use Juice or Cooling. a real review would use GPU mezzanine cards anyhow, while still considering factors like heat, performance, and price.
But the focus of Waxman’s discussion remains on that. “At some point, you really want that solution that surprises you and can do it with lower power consumption. And that’s where our FPGA solutions excel.”
Cloud Business
The last thing to consider is Intel’s cloud business. These customers now account for 25 percent of their data center group revenue.
Overall, their purchases are growing by about 25% per year. The overall data center group business is expected to grow by 15% for the next few years starting in 2016. Let’s do some math.
If Intel’s plan goes as planned, his data center group will have $16.6 billion in revenue in 2016. Cloud service providers (which include cloud builders and hyperscale computers that use our language on The Next Platform) account for about $4.1 billion, with the rest going to Intel Data Center, with sales figures of about $12.5 billion. As a result, Intel Data Center business is growing at about 12 percent (in addition to the cloud), which is half the rate of the cloud. Intel needs to meet the growth and apparent FPGA demand in the cloud any way it can, even if it only takes up a little bit of Xeon capacity. For Intel it is this option that is better than the option of allowing GPU acceleration to continue to grow.
The programming aspect is probably one of the main factors preventing widespread adoption of FPGAs (unlike other gas pedals with rich development ecosystems such as CUDA for Nvidia GPUs). This allows developers to go to C-based languages to do extended designs, or use OpenCL, rather than the low-level models that have plagued FPGA development in the past. But even with all the milestones in the application process, FPGAs are still not favored by the mainstream. We will be exploring ways and opportunities to solve the programming problem.
While we’ve talked to many vendors in this relatively small ecosystem (including Altera and Xilinx, two major vendors), according to longtime FPGA researcher Russell Tessier, the days of FPGAs making a splash in the broader market are ahead, and new developments mean broader adoption.
Having studied FPGAs for more than two decades at the University of Massachusetts (where he also worked at Altera and was the founder of Virtual Machine Engineering, which Mentor Graphics acquired), he sees a formal slow change in the landscape of FPGAs from scientific projects to enterprise applications. He believes that the key to this comes from the improvement of design tools, as designers continue to improve their designs to a high level. In addition to this, the tool vendor can better guide the chip development. He added that the large amount of logic within the device means that users are able to achieve more functionality, which makes FPGAs more widely appealing to more areas.
One of the clear trends in FPGAs over the past few years is that these devices are easier to “program,” Tessier said.
Xilinx is now encouraging design in C when using its Vivado products, and Altera has an OpenCL environment that has been developed. The key is that both companies are trying to create an environment that allows users to use more familiar programming (such as C and OpenCL) without having to use Verilog or VHDL, which is what RTL design experts are known for. while there have been good successes over the past few years, this is still in the advancement phase, but it will help move more things into the mainstream.
One of the really beneficial factors for FPGAs is that if they are used with chips to create a fast internal interconnect, it can solve the limitations in memory and data movement. This advantage is the main incentive to attract Intel to acquire Altera. In addition, if large companies like Intel and IBM can actively promote the construction of the software ecosystem of FPGAs, its application market will expand rapidly. mainstreaming of FPGAs (at least now not as important as GPUs,) may emerge more quickly.
The increase in standard core processor integration is definitely the key,” Tessier explains. The barriers used to be language and tooling, and as these become less and less of a barrier, it opens the door to new collaboration opportunities for chip vendors. As these and other “mainstreaming” trends emerge, the application areas for FPGAs that continue to make a difference will continue to grow. For example, financial services stores were the first users of FPGAs for financial trend and stock selection analysis, but the use cases are expanding. Now there are more powerful devices that can solve bigger problems
A wider range of applications
In addition, FPGAs are finding new uses through other new areas, including DNA sequencing, security, encryption and some critical machine learning tasks.
Of course, we want FPGAs to be powerful and “in” the world’s largest cloud and hyperscale data centers, strongly agreed Hamant Dhulla, vice president of Xilnix’s data center division. He told The Next Platform in early 2016 that “heterogeneous computing is no longer a trend, it’s a reality,” and that’s when Microsoft launched the Catapult case using FPGAs (there are many now or will be many later), Intel acquired Altera as well as seeing more announcements that FPGAs would be widely used in data centers.
FPGAs are emerging in more diverse application areas from machine learning, high performance computing, data analytics, etc. These are related to the increasing availability of on-chip memory embedded in FPGAs, which are expected by FPGA manufacturers and potential end users.
Dhulla said the market potential is large enough to allow Xilinx to adapt the way it does business. Storage and networking have dominated the FPGA user base for the past few years. But demand on the compute side will far outpace storage and networking over the next five years, and both will continue along a steady line of growth.
In other popular areas of FPGAs (including machine learning), they are more like a “collaborative” gas pedal with a GPU. There is no doubt that for the training part of many machine learning workloads, the GPU is the primary one. So a lot of compute power is needed here, just like HPC, where the power envelope tradeoff is worthwhile. But these customers are buying tens or hundreds of GPUs, not hundreds of thousands, and the huge number of gas pedals being used in the inference part of the machine learning pipeline is where the market is.
As we pointed out, Nvidia is using two separate GPUs (with M4 for training and lower power M4 inserts to cut down on servers) to offset this, but Dhulla believes that FPGAs can still reduce power consumption by adopting a PCIe approach that can also be embedded in hyperscale data centers.
Their SDAccel programming environment makes it more practical by providing high-level interfaces to C, C ++ and OpenCL, but the real way to drive hyperscale and HPC adoption is through end-user examples.
When it comes to these early adopters, it’s like setting the stage for the next generation of FPGA applications, and Dhulla points to companies like Edico Genome. Xilinx is also currently working with customers in other areas, including the historical computing side of oil and gas and finance. Early customers are using Xilinx’s FPGAs for machine learning, image recognition and analysis, and security, which can be seen as the first step in the development of their compute acceleration business.
The real opportunity for large-scale adoption of FPGAs lies in the cloud, despite double-precision performance and poor overall price. This is because FPGAs can provide advantages that GPUs cannot. If FPGA vendors can convince their end users that their gas pedals can provide considerable performance gains (and in some cases they will) to critical workloads. Providing a programming environment that advances OpenCL development through completeness-wise with other gas pedals (e.g. CUDA) solves the price problem by making FPGAs available in the cloud. This could be a new hope.
Of course, this hope comes from deploying FPGAs within a cloud-based architecture with ultra-dense servers, rather than on a stand-alone sale. This model is already happening in financial services for FPGAs.
Just as their GPU gas pedal “partners” are pulling around deep learning to get more users quickly, FPGA devices are exploring a real opportunity to invade the market by solving the problems of neural networks and deep learning.
New application hosts mean new markets, and with the rollout of cloud-based applications eliminating some of the management overhead, it could mean wider adoption. FPGA vendors are pushing hard for its use in some key machine learning, neural networking and search applications. FPGAs are becoming more prevalent in hyperscale contexts in areas such as natural language processing, medical imaging, deep data detection and more.
Over the past year, multiple applications of FPGAs have gained exposure, particularly in areas such as deep learning and neural networks, as well as image recognition and natural language processing. For example, Microsoft is using FPGAs to deliver 2x the search service on 1,632 nodes with innovative high-throughput networks to support Altera FPGA-driven work. Chinese search engine giant Baidu (also a user of many deep learning and neural network task GPUs) is using FPGAs to perform storage control with a daily data throughput of between 100 TB and 1 petabyte.
Large-scale data center and other applications in the field using FPGAs are drawing more attention to the single-precision floating-point performance of FPGAs.
While some case studies use (including the Baidu example), GPUs as compute gas pedals and FPGAs for the storage side, researchers at Altera, Xilnix, Nallatech, and IBM in the OpenPower Consortium are demonstrating the bright future of FPGAs in cloud-based deep learning.
It can be said that now belongs to a golden age of FPGAs.
