Kevin Covington, CEO of Australian low-latency technology solutions provider Metamako, says that while the race to zero for some in the trading community resulted in firms migrating large parts of their infrastructure as close to trading venues as possible—at huge cost—that was only part of the problem: “People began to realize that the latency was in the computer fabric,” he says.

This is where FPGAs—which reduce the response time of the circuits—come into play, as they increase the throughput of systems and decrease data load times, enabling applications to process financial data at a faster pace.  

For the Masses

Robert D’Arco, CEO of Chicago-based Rival Systems, a provider of trading and risk management software, says that in the capital markets, FPGAs have been used to process market data better, but haven’t been democratized. “The reality is that they haven’t really been adopted by the masses because of the complexity and some limitations to the FPGA cards themselves. The complexity is that it’s very difficult to develop logic within a particular card, and [as a result] you’re limited to how much memory and how much work you can do on the card,” he says.

Earlier last year, Rival teamed up with FPGA ultra-low latency trading solutions provider Algo-Logic Systems to develop an integrated offering combining Algo-Logic’s FPGA hardware and Rival’s trading and algorithmic strategy development software. The solution—which is geared toward the futures and options markets—enables traders to capture the sub-microsecond latency and deterministic performance previously enjoyed only by those trading firms with the resources to afford expensive internal infrastructures. 

From a market data standpoint, firms have begun using FPGAs to build end-to-end trading logic, D’Arco says. This involves capturing market data, performing “basic” calculations, and then sending an order out. However, it isn’t as easy as that: A firm must first have an infrastructure appropriate to handle this kind of traffic, and a team of highly skilled developers to do the work. And typically only the firms with the biggest budgets are able to take on such talent. 

Rival aims to bring FPGAs to the masses by employing a hybrid approach. In this instance, the FPGA kicks in when there are latency-sensitive processes to run, but doesn’t interfere with the software for the majority of work, letting those programs run on their own. 

“By having that hybrid approach, you get the best of both worlds,” D’Arco says. “You can get around some of the limitations of the FPGA card by doing the less pertinent work in the software space and really leveraging the card to do what it’s made for.”

An example of this is when orders and quotes are being sent out using software and suddenly a specific event happens and those quotes need to be cancelled as fast as possible. “The logic of detecting that event in the market and then sending the message to the exchange to get it out as soon as possible is all happening on the FPGA card. That sort of really critical, extremely low-latency piece is all happening on the FPGA card, but all the other logic is still happening in the software space. The beauty of that solution is that it’s completely seamless to the end user, so that they don’t have to do anything or write any code. It just works,” D’Arco says. 

Not Just About Speed

While FPGAs have been used by many different market participants, such as high-frequency trading (HFT) shops for hardware acceleration and pure speed, data giant Thomson Reuters uses it for something else—throughput and capacity. 

“We aren’t using this great technology for what it was originally designed for, but we are using it for throughput, cost control, and performance in terms of capacity and throughput,” says Douglas Munn, head of Elektron Real Time at Thomson Reuters.

Last October, Thomson Reuters launched a direct feed for the voluminous feed of US options quote and trade data from the Options Price Reporting Authority (Opra) via its Elektron Real Time consolidated feed of historical and reference data sourced from exchanges and over-the-counter (OTC) markets, which gives clients access to Opra data without needing additional server capacity to handle the high-volume feed.

To do this, Thomson Reuters partnered with UK-based Celoxica, which provides hardware-accelerated products using FPGA-based architectures. Through the pairing, Thomson Reuters leverages the UK-based vendor’s FPGA technology to deliver data on an exchange-by-exchange basis at minimal latency. 

Munn says Opra’s feed carries data on 10 million instruments, and 31 million updates per second that the vendor must normalize and distribute. “The update rate is high, and because of that we’ve had to continually upgrade and change our technology to ensure we are putting together something that is reliable, fast, efficient and cost-effective,” he says.

Prior to its partnership with Celoxica, running the platform required a lot of hardware. “When we moved over to the new hardware, we dropped our hardware usage by 70 percent by using the FPGA technology,” Munn says.

The more efficient firms can be with their hardware footprint, the better it is for their business as a whole, he adds, as they can free up that excess hardware to run other operations, or can retire unused hardware to save power and money.

But there are still more use-cases for FPGAs beyond speed and throughput, adds Rival’s D’Arco—for example, pre-trade risk checks, where FPGAs can be used to quickly perform checks on trade orders and ensure they are within a certain size or limit. If they comply with the firm’s limits, they continue on their way to market with virtually no delay. If they are outside of those parameters, FPGAs can be set to reject those trades.  

“From a pure trading perspective, the easiest case is a future spread. If you’re looking at a small number of instruments, you’re looking at, let’s say, two futures. If I get billed on this one order I automatically want to shoot out the second leg of that order. FPGAs are very effective at doing that basic logic,” he says.

Apart from specific trading applications, Metamako’s Covington says the biggest growth area for FPGAs is in being able to timestamp data with high precision. This is useful when it comes to complying with new regulations such as the revised Markets in Financial Instruments Directive (Mifid II), which imposes strict timekeeping and time synchronization requirements on trading and reporting activities, as well as for post-trade analysis that requires granular review of trading activity. 

Fuelling AI 

Due to their ability to process large datasets quickly, another potential use-case for FPGAs is accelerating artificial intelligence (AI) and machine learning capabilities. For example, last August, Microsoft unveiled Brainwave, an FPGA-based system for ultra-low latency deep learning in the cloud, developed in collaboration with microprocessor manufacturer Intel. The system is designed for real-time AI, which allows it to process requests with ultra-low latency responses. Microsoft sees real-time AI becoming more important as cloud infrastructures process live data streams.

Metamako’s Covington says this represents one of the more intriguing future use cases for FPGA technology. “An interesting use-case is where people are doing work around AI and machine learning to stream analytics. FPGAs can help an immense amount here. With FPGAs, you can cope with the volume of data.” 

Meanwhile, Rival’s D’Arco adds that people need to be smarter about how they analyze data to come up with better trading ideas. “It’s hard to take all that data, process it, calculate it very quickly, and react in real time,” he says. “FPGAs will help in that, but you have to think about what problem you’re trying to solve. Can the FPGA solve it, or does it overcomplicate it?” 

Buy vs. Build

D’Arco says some of Rival’s clients have previously tried to build their own FPGA solutions, but found it took a lot of time and effort. This is due to the complexity of FPGA cards, which require developers to hard-code tasks into hardware, rather than writing code in software. And while software developers are plentiful, many firms overlook the cost of talent required to build FPGAs, he adds.

“There are definitely firms that are trying to do end-to-end trading systems taking market data, logic and sending orders all on FPGA cards. I think what they realize is that it’s a big effort from a cost perspective. Hiring an FPGA developer costs two to three times more than a developer doing C# or C++ development. It’s just pure labor cost of getting into that space,” he says.

On top of that, how a solution is built and designed is complex, as one might expect when dealing with such low latencies. One of the limitations of taking the build route is firms can lose a lot of flexibility. 

“If you can build a strategy that does one very specific thing and if you can make enough money to do that one very specific thing that justifies the cost of FPGAs, then great. But if not, that is where the model falls apart. You realize after taking a year to get it up and running doing one basic thing, you then spend another six months doing the next thing. The cost-benefit analysis becomes a challenge because you are taking six months to get strategies up and running,” D’Arco adds. 

Munn says Thomson Reuters definitely considered building its solution in-house, but then thought twice. “There are times to partner and there are times to build it yourself,” he says. “In this case, [partnering] looked like a better solution. Our partner, Celoxica, had the technology, capabilities, and expertise there, and what we did was adapt their technology to then publish out our Thomson Reuters APIs and the whole point there is they bring some skillset to the team. By publishing out a Thomson Reuters API with all our symbology, it makes it easy for our customers to upgrade. That’s our overall goal.”

Looking Ahead

Technology is constantly evolving, though whether a new technology will evolve that can replace the use of FPGAs is, at this point in time, anyone’s guess. “The bottom line is that it is [a piece of] technology. It’s going to change. That’s the one thing we know,” Munn says.

For example, Covington says new technologies such as quantum computing could impact the FPGA market. “People are paying attention to quantum encryption, and FPGAs will play a part in that happening,” he says.

Energias Market Research predicts that the value of the global FPGA market will rise to almost $13 billion in 2023, up from $7.1 billion in 2016, driven by increasing demand for smartphones and hand-held devices, bandwidth in wireless networks, and continuous demand for electronics components. Most of those will be evolutions outside of trading. But D’Arco says he believes FPGAs will evolve to be more integrated into trading systems. 

“There’s stuff happening that’s going to make development of FPGA cards easier. As that progresses and becomes more mature, then I think the complexity of FPGA cards is going to be reduced. And if you do that, there will be more adoption,” he says. 

But this is not a silver bullet, either. It is important to determine what problem needs to be solved, and how and where technology like FPGAs can help. 

“People don’t really understand what their actual latency is, and they assume if they use FPGA cards they’ll be better off, when sometimes just using pure software may actually solve the problem,” he says. “The key is to figure out where the real issue is, and how you can solve it.”