Hacker News Reader

Comment by JoachimSchipper | Hacker News Reader

FPGAs are hard/expensive to program (including being very painful to debug), at least for most tasks - and few people are good at both FPGA and software development, so you may need different people to do the programming. New FPGA generations can require substantial re-work, at least for best performance. Many tasks would require kernel support (which is another "hard/expensive to program, and annoying to debug", although not nearly as bad).

You need a deployment scenario where FPGAs beat GPUs (and CPUs) by enough to be worth that pain, but not by enough that someone already has a ready-made ASIC for you. I agree that one would expect to see more deployment of FPGA's naively, but not a lot.

View on HN

This seems right, except that Microsoft has clearly developed a higher level language than VHDL / Verilog for writing "virtualized network functions" that compiles down to something they can flash onto an FPGA (they call it Flow Tables in the presentation). So there's a higher level language that's, presumably, more intuitive for "normal" engineers to work with. When a new FPGA comes out the compiler is re-written and high level code stays the same. Basically the same thing we do with CPUs. So they're not _inherently_ hard to program.

SDN is common, and definitely a scenario where FPGAs outperform GPUs. Same for machine learning. (FWIW, Google uses ASICs for both of these use cases.) What we need is a high level language and access to FPGAs. I've been interested in business opportunities in this area, but it's hard to monetize. What's somewhat surprising is that the language bit hasn't been solved in the open source world either. My suspicion is that the reason an open source language like Flow Tables doesn't exist (AFAIK) is that there's not widespread access to FPGAs.

One of the great benefits of FPGA compared to ASIC is that when you have some obscure low level bug appear you can just patch it on already sold devices, just like in software. You can also implement new features the same way, limited of course.

Yep, this right here. You really need a good reason to choose an FPGA. The inherent throughput and a lot of stuff is incredibly powerful, but the development is much more involved. To me anything sufficiently nontrivial always feels building a house with Lego.

Reply on news.ycombinator.com