|
|
|
|
|
|
by sixdimensional
4077 days ago
|
|
|
I'm personally interested in your comment about Postgres, I have been thinking about implementing digital circuit versions of certain database operators to see if FPGAs can be used for distributed, parallel processing of data streams. I think there is a fit for digital versions of analog circuits to substitute for software operators (if that makes any sense - this is pushing my knowledge level to the limit). I've seen some amazing related work at MIT here:
http://people.csail.mit.edu/wjun/papers/fpga2014-wjun.pdf |
|
|
Some people will say that there are already pcie cards with fpgas on them and that this is not a new idea. But the issue is that existing cards like Digilent make are there for building pcie card designs. The FPGA (usually one) on the board has to implement the pcie fabric plus whatever it wants to do. Imagine if when you write an OpenGL/CL program you had to deal with all the DMA work and quirks of the underlying setup. Instead the grid of fpgas should have standard ways to access certain mapped regions of memory through a different chip that handles the PCIe transport layer stuff. If this was available then adding FPGA support into POSTGRES would not require bypassing the HAL of your kernel or worrying about timing details of something fundamental to modern computers (and thus software engineers) like PCIe.
Obviously the FPGA is still something that is hardware specific, but who knows what could be done with that. Even if we had to Synthesize from raw HDL every time, at least loading the chips would WORK. We could start using these chips to solve problems instead of thing slike a one off project by an EE major to build a grid of FPGAs that process bitcoin. The bitcoin work was impressive but it bothers me SO MUCH that they used reconfigurable devices to build a solid card with a fixed function.