For financial firms obsessed with speed – getting data, parsing it, executing on it, getting trades into the market – the solution is increasingly found in more flexible hardware.
While artificial intelligence remains the most prominent, sexy end of financial innovation, the ability to put it into play in low-latency strategies relies on the physical circuitry.
And it’s not just the most niche set of high-frequency traders that are looking to more flexible, agile versions of hardware. So are prime brokers, broader quant investors, crypto-currency miners, and banks’ trading floors.
What is FPGA?
At the heart of this is a hardware technology called field-programmable gate array, or FPGA. For anyone not in I.T., that’s a mouthful. The clue is in the first word: field. As in, out in the wild. FPGA is an integrated circuit – hardware – that a customer can configure after they buy it from a vendor.
Traditionally, hardware is delivered in the form of CPUs or GPUs, central processing units or graphic processing units, which are generic machines to carry out programs. They’re too general and too off-the-shelf to meet the needs of, say, high-frequency traders or other high-usage businesses.
On the extreme end to meet such demand, hardware has become very specialized, as is the case with application-specific integrated circuits, or ASICs, which are processors designed to do just one thing. ASICs are popular with bitcoin miners, for example, as the computers need to focus solely on solving the mathematical puzzles required to generate a new block of data.
The volume and variety of data keeps increasing
Mutema Pittman, Intel
But ASICs can’t be used for anything else than what they’re built for. And trading algorithms need to be adapted all the time.
That’s where FPGA comes in: it’s hardware that you can adjust. FPGA is not a new technology, but it’s new to financial services. It’s been adopted by tech-heavy players in the U.S. but has been slow to make its way to Asia.
Adaptable – and fast
It was the focus of discussion, however, during a conference last week for technology officers in Hong Kong organized by STAC, a U.S.-based association that tests and benchmarks finance-oriented hardware.
“FPGAs give you the power of hardware-dedicated architecture as well as flexibility because you can reprogram them,” said Mutema Pittman, who heads the enterprise business division for network and custom logic at Intel. “The volume and variety of data keeps increasing, and firms need to react to this in a timely manner even as market requirements continue to change.”
There are other aspects of FPGA that make it more adaptable. It can talk to other devices with ease, making it a useful server for communicating with, say, a stock exchange’s API. It can also access computer memory directly, without intervening layers, which makes it faster than traditional CPUs.
The first use case for FGPA is market data
Miguel Ortega
“The first use case for FPGA is market data,” said Miguel Ortega, who heads market data for a global bank in Tokyo. “It makes it easier to convert data feeds from exchanges into applications to make decisions,” such as trade signals, placing orders, and ensuring trades comply with mandates or rules.
“FGPA are used for strategies that require the lowest latency,” said James Morris, Sydney-based technology leader at Optiver Asia Pacific, an electronic market maker. “FPGAs enable firms to access data from exchanges, make decisions, and execute – at the speed of nanoseconds.”
Although firms like Optiver build their own hardware, not all firms need to be at the bleeding edge of speed, depending on the trading strategy. But some vendors are producing technology to allow their FPGA boards to operate as fast as possible. One such company, Exablaze, enables trades at latencies as low as 31 nanoseconds, says Matthew Grosvenor, Sydney-based senior vice president of technology: “That’s easily a 10-times speed improvement by moving software from CPUs to FPGA firmware.”
A nanosecond is a thousand-millionth of a second, or 1/1,000,000,000 of a second.
Other use cases
Other vendors cater to quant shops and investment banks’ trading floors that don’t need to operate quite at that speed, but still need to be fast.
While FPGA’s most obvious use case is ultra-low latency, some functions require computing to go along with the processing, such as smart-order routing or sentiment analysis. These don’t need to be done at the pace of nanoseconds, and flexible hardware could be a substitute for, say, co-locating servers at the exchange’s data center.
FPGA are used for strategies that require the lowest latency
James Morris, Optiver
FPGA may also be a good hardware choice for even less time-sensitive functions such as pricing options or running Monte Carlo risk scenarios, because they can draw the data directly from sources. But they don’t make sense for heavy computing needs, such as anything requiring complex mathematics.
So why is FPGA still an obscure technology in Asia?
Finicky FPGA
First, it comes with DevOps challenges. Not a lot of programmers know how to use it, and even fewer in finance. Second, there’s no open-source tools for FPGA (unlike for software), so firms must either buy from vendors, or build it from scratch themselves.
Third, most stock exchanges in the region haven’t invested in FPGA either, preferring to hand off data through software solutions like APIs. Some exchanges, unfamiliar with the tech, are concerned it will invite a barrage of customized orders they won’t be able to handle.
This view might change as they compete to ensure their matching engines remain able to handle ultra-fast trades, but it’s mainly venues in developed markets like Japan and Australia that cater to delivering data to FPGA hardware.
Fourth, Asia’s market is fragmented. Yes FPGA is flexible, but it still needs to be adapted for each market’s particular environment but vendor solutions (primarily developed for the U.S.) don’t cater to such nuances.
For banks trying to keep up with their lucrative HFT clients, they also need to invest in bespoke FPGA connectivity. It’s an expensive proposition.
“But for someone sending enough orders, it’s worth it,” said Ortega.
