UPGRADE YOUR BROWSER

We have detected your current browser version is not the latest one. Xilinx.com uses the latest web technologies to bring you the best online experience possible. Please upgrade to a Xilinx.com supported browser:Chrome, Firefox, Internet Explorer 11, Safari. Thank you!

Mining insight from mountains of Big Data using teams of FPGAs: Pico Computing

by Xilinx Employee ‎04-28-2014 01:46 PM - edited ‎04-29-2014 06:06 AM (36,312 Views)

When borax miners needed to haul tons of mined ore across 165 miles of Death Valley to the railroad in Mojave, California, they invented the twenty-mule team to pull 10 tons of ore per load using the largest wagons ever to be pulled by draft animals. That was during the 1880s. When today’s data miners need to address the three biggest big-data challenges—insight, time, and money—they harness the unique, programmable processing power of multiple FPGAs to prize sought-out answers from mountains of data.

 

 

20 Mule Team.jpg

 

A 20-mule team pulling borax through California’s Death Valley during the 1880s

 

Pico Computing has been harnessing the power of multiple FPGAs to solve tough problems in high-performance computing (HPC) for many years now. Without identifying the company, I recently described one of Pico Computing’s early multi-FPGA HPC boards based on Xilinx Spartan FPGAs that I’d spotted in the QPL booth at the NAB (National Association of Broadcasters) show in Las Vegas. (To see a back-of-the-envelope analysis of this early Pico Computing FPGA board, read “The evolution of “low-end” FPGAs as spotted at NAB 2014: From the original Spartan to Artix-7 FPGAs.”)

 

Pico Computing’s boards have evolved since that board was designed.

 

Today, Pico Computing uses a modular approach to creating HPC systems based on a series of “backplanes,” which are actually carrier cards based on a PCIe Gen2 switched architecture. For example, Pico Computing’s EX-500 Backplane Boards accept as many as six of the company’s M-Series FPGA modules.

 

 

 Pico Computing EX-500 backplane board.jpg

 

Pico Computing EX-500 Backplane Board

 

 

Here’s a block diagram of the EX-500 Backplane Board:

 

 

Pico Computing EX-500 backplane board block diagram.jpg 

 

Pico Computing EX-500 Backplane Board block diagram

 

Each of the six sockets on Pico Computing’s EX-500 Backplane Board accepts an M-Series FPGA module. There are five modules in the series including four that incorporate Xilinx Virtex-6 or Virtex-7 FPGAs with on-board local QDRII SRAM and DDR3 memory. Here’s a photo of an EX-500 Backplane Board loaded with six Virtex-6 FPGA modules:

 

 

Pico Computing loaded backplane board.jpg

 

Pico Computing EX-500 Backplane Board loaded with six Virtex-6 FPGA modules

 

 

That’s how Pico computing teams as many as six FPGAs on one board. However, the company also offers rack boxes that accept as many as eight EX-500 boards, creating FPGA arrays with as many as 48 devices. (That’s more than double the size of a 20-mule team.)

 

The above descriptions might give you the impression that Pico Computing is a board and box vendor. Actually, the company has developed expertise in harnessing FPGAs to mine information from data. The hardware is a means to an end.

 

Consider these sentences from a Pico computing White Paper: “Fueling insight is the real time capture of the most elusive and fleeting data. Consider, for example, that only 20% of the gold in a mine is visible; 80% is “in the dirt” and must be dug out. Likewise, digital data possesses concealed value that also must be mined. But because most data cannot be easily captured at the edge—where it is generated in massive volumes and at ever-increasing speed—we end up looking at the world through a pinhole: the vast majority of the picture isn’t even seen, let alone caught! Consequently, we’ve gotten used to taking actions and making decisions based on only a sampling of the available data.”