Jason J. Gullickson

Jason J. Gullickson

Thoughts on an FPGA Cloud

We were talking about applying FPGA technology to a specific engineering problem at Murfie when Preston asked “has anyone built a cloud of FPGA’s?”.

It was one of those questions that as soon as you hear it, it seems like an obvious thing to do, but I’d never heard of it being done, and it’s certainly in the center of the Venn Diagram of my interests.  Some casual research doesn’t reveal the existence of such a thing, but it does show why it might not yet exist.  FPGA products lean heavily in the direction of embedded electronics, aimed at being incorporated into a piece of hardware and interfaced with various other physical bits (sensors, actuators, etc.).  This application doesn’t really lend itself to the “… As A Service” model, but it’s certainly not the only application for FPGA’s.

Pure software systems can benefit from FPGA implementations if they contain the right kind of problems.  An easy example of this is the Bitcoin miner , an extremely compute-bound problem which can be executed much faster (and with dramatically lower power consumption) using an FPGA implementation of its core algorithm.

While it’s true that most software lack such a clear need for hardware-based acceleration (and the associated engineering & design), I think that a lot of modern software has “pockets” of this kind of processing, and especially as we build systems that rely more and more on hashing algorithms, encryption and also become more savvy about using asynchronous and parallel-processing architectures.

So to me the missing piece of a valuable “FPGA PAAS “ is an efficient means of expressing existing or new software in ways that can leverage the performance, efficiency and flexibility of FPGA hardware.  To some degree this is a self-solving problem, history shows that making tools available to developers grows the base of developers who know how to use the tools.  In addition to expanding the knowledge of leveraging FPGA tech in software, I think that a properly designed platform could automate the process as well (runtimes and compilers capable of identifying candidates for optimization, potentially generating the optimized implementation on-the-fly).  A “ soft core “ ARM processor, implemented in an commodity FPGA will run at about 23 the speed of its traditional silicon implementation, but the soft processor could also monitor the code running on it and adapt itself to run that specific code much more efficiently than a general-purpose processor.

Having given this some thought, I think there is potential for a cloud of FPGA’s, even if nobody knows it yet.  Projects like HP’s “Moonshot” and the broader push toward modular, configurable computing point this direction as well, and it’s hard to imagine a more dynamic platform than one which can change it’s fundamental hardware implementation at will.

- Jason