Remember me

Register  |   Lost password?

The Trading Mesh

Adopting the Right Approach to FPGA Implementation

Mon, 11 Feb 2013 04:33:22 GMT           

An Interview with Arnaud Derasse & Laurent de Barry, Enyx

(11th February 2013)

  

Sponsored by Enyx (www.enyx.fr)

 

FPGA technology is becoming increasingly prevalent in the financial trading ecosystem, as firms look to do more and more on hardware, running applications at “wire level” to reduce latency.

 

One supplier that is helping firms to achieve ultra-fast, deterministic throughput at various points in the trading chain, is Paris-based Enyx, specialists in FPGA hardware acceleration solutions for financial trading.

 

We spoke to Enyx Co-Founders Arnaud Derasse, CEO and Laurent de Barry,

Head of Customer Support, to find out more about the challenges they help their customers to overcome.

 

HFT Review: Arnaud, Laurent, welcome to HFT Review. Can we start with you giving some background on the formation of your company? How did it all come about?

 

Arnaud.tiffArnaud Derasse: Enyx was created three years ago, by six engineers working in Paris (what we call the “young team”) along with some industry veterans including Emmanuel Pelle and Yann L’Huillier (who we call the “grey hairs”).

 

The six of us in Paris all had backgrounds studying micro-electronics at “les grandes écoles” in France, we’d all worked together at the same company (an FPGA-based market data specialist), and we saw an opportunity to push FPGA in other areas of financial trading, not just in market data acquisition.

 

HFTR: Such as?

 

Laurent de Barry: Areas such as pre-trade risk; TCP acceleration for order entry; wireless links.

 

HFTR: Wireless is a hot topic in the HFT space at the moment, so let’s start there. Can you expand a little on what you guys are doing in that area?

 

AD: We provide appliances for bandwidth management for radio links. As your readers are probably aware, all the major telecoms operators in US, UK and Europe have been setting up radio links between the main trading venues, but what they may not know is that Enyx has been working on the London-Frankfurt link, operating appliances and equipment for bandwidth sharing, compression, dealing with water refraction problems, redundancy management, etc.

 

HFTR: Where do FPGAs fit in?

 

LdB: The reason why we use FPGAs to maintain the low latency of the solution from end to end is because the legacy wireless equipment was really slow, so we’re using FPGAs to speed up the process.

 

Text Box: “It would be strange to pay for expensive, ultra-low latency microwave equipment and then to use conventional software for all the other steps”AD: The telcos gain speed using wireless versus fibre optic cables, but they still need some equipment to share bandwidth between customers, to compress data, to do filtering, etc. But they can’t do that using software because it would be very strange to ask their customers to pay for very expensive equipment to upgrade to ultra-low latency with microwave and then to use conventional software to do all the other steps.

 

So we have projects working on replicators for towers (there can be a lot of towers between two cities) and we have projects to cancel negative effects over water, for example when crossing the Channel, going from London to Frankfurt the signal reflects on the water and sometimes cancel itself. To avoid this problem, the physics tells us that using two reception antenna and arbitrating the two signals received will avoid any data corruption or loss.

 

LdB: Despite the importance of this data recovery, the most important feature of our product is the fair bandwidth management over this low bandwidth wireless link. Several customers will share this link and it is important that they get the bandwidth they paid for. To ensure this, our product performs transparent Ethernet segmentation and reassembly and as result, prevents any user monopolizing the link.

 

AD: And all that equipment has to be operating at low latency to guarantee to the customer that the whole chain is low latency.

 

HFTR: Looking at the various different types of firm that make up the financial trading ecosystem, exchanges; market data vendors; banks; broker/dealers; proprietary trading firms; hedge funds; buy side asset managers; etc., can you talk us through how FPGA technology can benefit each of those categories of firm?

 

AD: Certainly. Exchanges and trading venues can use FPGAs for broadcasting market data to their customers in a high-performance manner.

 

LdB: They can also use FPGAs for inline pre-trade risk. By having the FPGA integrated on a smart NIC card, the exchange can do some risk checks against incoming order flow at the wire level.

 

AD: Looking at market data vendors, they generally already have some kind of technology in place, so we help them update their feed handling and translation technology to FPGA so they can move to the high frequency and high performance space, by providing very low latency data acquisition, feed handling, and normalisation of data to their customers.

 

HFTR: Are you seeing more and more market data vendors wanting to migrate to FPGA-based solutions?

 

AD: Yes. Some of them try to develop those solutions internally and some prefer to work with specialist vendors like Enyx.

 

Text Box: “Because some market data vendors have to service both legacy traders and high frequency traders, there’s a need to both maintain the functionality and lower the latency at the same time”LdB: Because some market data vendors have to service both legacy traders and high frequency traders, there’s a need to both maintain the functionality and lower the latency at the same time. So we typically need to integrate our technology into their existing infrastructure to enable that. We are, for example, working closely with ORC to provide an ultra low latency version of their Liquidator software. ORC customers will then be able to take advantage of the FPGA technology without needing to make any changes in the deployment of their ORC Trading environment, including algorithms or screen set ups.

 

HFTR: What about the sell-side, banks, broker/dealers, etc? Do you find that they’re increasingly becoming interested in running pre-trade risk checks on FPGA, for example?

 

AD: Yes, for sure. Also, where a bank is a market maker, they can use FPGA technology for feed translation and data acquisition too, as well as acceleration of order entry using fast TCP bypass. One of our products integrates a TCP stack to enable clients to have ultra-low latency order pass. We can help them in all three of those areas: pre-trade risk; feed handling; and acceleration of order flow.

 

And as for the prop trading HFT firms, they have similar requirements to the prop trading desks at the banks.

 

HFTR: Where might firms use FPGA-enable appliances versus smart NICs or maybe even FPGA-enabled switches? How and where is each of these approaches being implemented?

 

AD: Firms that want to get the benefits of FPGA but don’t have in-house FPGA expertise will generally use appliances. They just want something turnkey. FPGA-enabled NICS are more common where firms have engineering teams and where the FPGA can be integrated into their existing infrastructure.

 

LdB: At Enyx, we’ve taken a card-based approach because it’s modular. Smart NIC, card-based solutions can be easily integrated into existing servers and improve existing infrastructure.

 

AD: If the customer needs an appliance, that’s also very easy. Buy a server, push the card into the server and you’ve got an appliance. For instance, with some of the telecoms firms where we’re operating radio links, we’re selling servers containing the FPGA cards and a browser-based graphical user interface. They are basically appliances.

 

LdB: And it’s a turnkey solution, you just have to plug it in and it’s up and running.

 

AD: Our Solutions branch doesn’t generally work with firms that have their own FPGA expertise in-house, since those firms would generally build their own FPGA solutions using their own expertise. They would be more likely to work with our Technology branch, which provides custom IP cores for specific purposes.

 

HFTR: What’s the balance between customers who are fairly new to FPGA and who deal with your Solutions branch, versus those who have FPGA expertise in-house, who can program in Verilog or VHDL and who deal more with your Technology division?

 

Text Box: “A lot of firms have gone down the route of buying the hardware and the IP cores to build something in-house, but have come back to buying off-the-shelf solutions because they found it all too complex”AD: Most of the Tier One banks now have FPGA teams and try to do things themselves. But hedge funds and buy-sides generally prefer to buy “off-the-shelf” because they understand that it’s very expensive to have an FPGA team and that if you’re trying to do everything yourself, the time-to-market to get something working is very long. It’s interesting that a lot of these firms have gone down the route of trying to buy the hardware and the IP cores to build something themselves in-house, but have come back to buying off-the-shelf solutions because they found it all too complex.

 

And to do very generic things like market data acquisition, why re-invent the wheel?

 

LdB: The FPGA teams at the big banks generally work on custom IP blocks for specific purposes rather than market data acquisition or pre-trade risk.

 

HFTR: What kind of specific purposes?

 

AD: Mostly trading algorithms and other things specific to their own business and trading strategies.

 

HFTR: Are you seeing a trend of more and more firms trying to program their algorithms onto FPGA using custom IP blocks?

 

LdB: Well, most firms don’t really communicate about this, but for those who do talk about it, it’s more a dream that they want to achieve rather than reality. Usually the firms that talk to us about putting their algorithms on an FPGA don’t really understand the complexity of FPGAs.

 

HFTR: Does that mean that educating the market a large part of what you do?

 

LdB: Yes, a part of our role at Enyx is to guide them along, understand their needs, educate them on the complexity of the project they want to start and take them back to the reality so we can continue step-by-step through the project.

 

AD: The first step is usually to get the market data and order entry accelerated and maybe after a year or so, they’ll also have their algorithms in FPGA.

 

At a big bank, they may have five people working for two years to get their custom IP blocks live on FPGA, so we tell these smaller firms, “yes, it is possible, we will help you, but we need to start with the first step, then the second”, and so on.

 

Text Box: “Our advice would always be to take a step-by-step approach to projects, with little iterations. Start with something that will work really quickly and then evolve from that”LdB: A lot of firms really don’t understand the time-to-market of custom FPGA solutions. So our advice would always be to take a step-by-step approach to projects, with little iterations. Start with something that will work really quickly and then evolve from that.

 

AD: Firms also need to concentrate on what they are good at and off load the rest to specialists or off-the shelf solutions.

 

HFTR: Do you have any particular success stories you can share?

 

AD: Well, as you know this market is very secretive so we can’t name names, but there was one particular investment bank in the UK we worked with recently. They were running an internal project to put the full trading chain for their prop business onto FPGA, i.e. market data, trading engine and order flow. The project was a little stuck, they’d been working on it for over a year and they were finding it very difficult to get various things working. So we came in and helped them by providing some IP cores and doing some consulting. They concentrated on their trading engine and we took care of the all the problems they had around PCIe, connectivity, communication, all that stuff. And within two months they went into production. We helped them to concentrate on their business and we took care of the rest. That was a very interesting project.

 

HFTR: Looking ahead, what does the future hold for Enyx?

Text Box: “In the future, more and more market data vendors will integrate FPGAs and create very low latency offerings”

AD: In the future, more and more market data vendors will integrate FPGAs and create very low latency offerings. Also, it will become a commodity for everyone to have pre-trade risk in FPGA. But FPGAs won’t be the answer to everything.

 

From an Enyx perspective, we need to ensure we’re listening to the market to be sure that we have an appropriate offering for all the bottlenecks and all the specific points where FPGA is needed.

 

We also think that OpenCL is the next trend in 2013, and we are preparing products leveraging this technology to offer flexibility and to break down time to market of our client’s internal projects.

 

We’ve also just announced the opening of an Enyx office in Asia (read the press release at http://www.thetradingmesh.com/pg/newsfeeds/enyx/item/71067)

 

HFTR: And are you completely focused on the financial trading space or do you operate in other areas too?

 

LdB: On the TCP and UDP side, we’re working with a partner, PLDA, which could open up new markets for us, like defence and telecom. But for now we’re completely focused on finance.

 

HFTR: Gentlemen, thank you.

 

Click here to download this article as a pdf

 

, , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,