For latency-sensitive trading applications, speed is of the essence, but so also is consistency. To get market data feeds in nanoseconds during times when the markets are slow and at microseconds when the markets are fast is unacceptable.

xCelor, a provider of high-performance trading hardware, has launched a new line of deterministic Layer 1 network switches with port-to-port latency of between 2 and 4 nanoseconds. That means the switches can replicate market data with zero jitter, so the application can be certain that it will receive and process the data within the same time interval regardless of how busy the market might be.

“With the xPort series, we’re reducing network latency to the lowest possible so the end user can focus on the content of the data,” said Stefan Gratzl, xCelor’s chief architect.

The switches, called the xPort series, reduce inbound market data latency by cutout out unnecessary processing employed by the full-service switches typically used in data centers today. The xPort switch simply replicates data, as quickly as possible, from one exchange handoff to multiple servers.

“The best full-service switches will have port-to-port latency of 300 nanoseconds,” said Rob Walker, xCelor’s chief technology officer. “The problem is that 300 nanoseconds is what you observe when the market is quiet. But when the market is busy, around the market open or close, which coincidentally is when low latency trading strategies are operating, that 300 nanoseconds on a full service switch can stretch out to 6 microseconds.”

The problem with full-service switches is that they are designed for applications that need to ship moderate sized data packets to millions of network nodes around the world.

“The full service switches that typical trading firms deploy in exchange data centers wouldn’t be out of place in a Facebook data center,” Walker said. “When you deploy it in Facebook, you need to store the routing table of the entire internet, whereas when you install in an exchange data center, you only need to store five or six routes. That versatility of a full service switch comes at a cost. The reason they’re not deterministic is that they examine every packet that comes in, even if all the packets are going to the same place.”

Because the xCelor switches have less processing power than the full-service switches, they can operate at much higher speeds.

“We took that processor power out of the equation to solve the problem of distributing market data quickly and deterministically,” Walker said.

xCelor will soon be announcing a “second-generation” switch that contains an FPGA card, which will enable trading firms to embed many data processing tasks at the source, i.e., within the switch itself.

Although many trading firms already use FPGA cards within data centers, they need to install them on every server, which is costly.

“Market data feeds have huge bandwidth,” Walker said. “The reason you would deploy an FPGA card in a server is to sort through 10,000 symbols worth of data as its arriving and chop it into the 50 or 100 symbols that the server app is interested in. You don’t have that kind of bandwidth between the network card and the CPU. The problem is an FPGA card in every server is very expensive. With our layer one switches, we are putting in an FPGA at the front door.”