Jeff, fpga developer, Nallatech Releases FPGA Boards for High Frequency Trading, here.
I was looking around for FPGA based PCIe boards when I came across something interesting from Nallatech. They’ve created two OpenCL compatible PCIe boards designed especially for the finance market. Named rather creatively “Nallatech 385” and “Nallatech 395“, they’re both based on the Stratix V from Altera which is the best of Altera’s high-end FPGAs. Both boards have a PCIe Gen3 x 8 lane host interface that should allow you to transfer data between the FPGA and the host machine at about 7.88 gigabytes per second (985MB/s per lane x 8). For connection to the market, the boards have SFP+ cages into which you can plug modules for 1Gb Ethernet, 10Gb Ethernet, SONET/SDH & OTN. The main differences between the boards seems to be the amount and speed of the on-board memory and the number of network interfaces.
I think the main selling point for these boards is the OpenCL compatibility. OpenCL is a programming language/framework that allows you to write C code and have the execution distributed across an array of hardware devices such as GPUs and FPGA boards. This is important to firms in the finance industry because they don’t typically employ FPGA programmers, but they’ve been employing C programmers for years. Most of their existing trading algorithms are implemented in C code because that was the most efficient way to implement it, unless you were willing to write assembly code. Today, the most efficient way to do it isn’t in C code and it isn’t even by using a computer. The fastest algorithms today are running on specialized hardware. In my opinion, OpenCL is just an intermediate solution created for a market that is not ready to throw their existing technologies away and start from scratch. But of course I would say that, I’m a hardware guy…
Lawrence Latif, The Inquirer, AMD thinks most programmers will not use CUDA or OpenCL, here.
CHIP DESIGNER AMD believes that most software developers won’t use CUDA or OpenCL to create code that runs on the GPU.
AMD has spent a lot of effort promoting OpenCL in the hope that developers will make use of the GPGPU in the firm’s accelerated processing units (APUs). However the firm thinks most developers will shun GPGPU specific languages such as CUDA and OpenCL and stick with what they already know.
Margaret Lewis, director of software for AMD’s server business unit, said that while OpenCL abstracted the GPU architecture to a degree, developers still require knowledge of how the GPU works.