Virtualising the FPGA

FPGA in the datacentre was first deployed by Microsoft for Bing acceleration. A user-programmable FPGA fabric as part of the Cloud offering was made available by Amazon at the start of 2017.

Where customer designs are loadable into the FPGA, security issues arise. Overheating, denial of service and data theft must be avoided and a means to virtualise (share) the FPGA resources is needed.

The FPGA has much more user state to context switch than a conventional CPU, so a sub-Hertz context switch rate is likely. This is at least 3-orders coarser than typical CPU time sharing. All recent FPGA families support dynamic loading of selected parts of the die while the remainder continues operating. The reconfiguration time is typically 10 to 100 ms per zone. This perhaps opens the way to finer-grained sharing, if needed.

One approach to virtualisation is to use dynamically-loaded modules under the `Shell' and `Role' approach to FPGA real-estate where the customer only has access to the part of the die via partial-reconfiguration. The service shim or shell provides a protected service layer.

Disadvantages of this technique are that quite a lot of real estate (and especially the multiplier resource therein) is consumed with overhead. Also there are design fragmentation inefficiencies. And many of the FPGA resources, such as most of the I/O pins and hardened peripherals go unused (not a problem for a Dark Silicon future!).

The reconfigurable logic resources of a conventional FPGA can be manually partitioned into user reconfigurable areas and a shell/shim for service access.