Tried to scan the video to see if they have any support for reconfigure.io [1] (Code FPGAs with Go concurrency primitives). Didn't see any mention. Anybody knows more? Or I'll have to watch the vid in detail ... looks pretty pedagogic anyways :) ...
Google's TPU uses 28-40 W, Microsoft's uses 25 watts.
Relevant but not deep: https://redmondmag.com/blogs/the-schwartz-report/2016/10/dif...
Why aren't more people deploying FPGAs like this? I started looking into using FPGAs for distributed systems work about six years ago when I was building a distributed database. Since then, about once a month I come across some problem that an FPGA could be useful for (i.e., enhance the performance enough that something otherwise infeasible becomes feasible). The devices themselves aren't that expensive (<$100) so why doesn't every server have one? Is there some hidden cost I'm not aware of (socket real estate, power consumption)? Or is it just a chicken-and-egg problem with not enough engineers asking for them because they don't know how they work because they're not more widely deployed?
Does anybody know if the FPGA accelerated NIC uses a different driver on the guest VM, or just the standard hyperv net driver?
If it uses the standard hyperv driver, can a guest be migrated live between accelerated and unaccelerated hosts?
The explanation of FPGAs between 5:00 and ~12:00 was very very good.
TL;DW:
- FPGAs are more power-efficient than CPUs and GPUs
- FPGAs are more flexible than ASICs
FPGAs vs CPUs:
- CPU: Sequentially apply instructions on data in place
- FPGA: In parallel, flow data through instructions in place