Meanderful: FPGAs and AI Processors: DNN and CNN for All(meanderful.blogspot.com)
meanderful.blogspot.com
Meanderful: FPGAs and AI Processors: DNN and CNN for All
https://meanderful.blogspot.com/2017/06/fpgas-and-ai-processors-dnn-and-cnn-for.html?m=1
3 comments
Better than 50% discounts for reasonable FPGA volume, true. Still expensive when compared to a V100.
Microsoft is seeing ~90Tops for their MS-FP8 datatypes: https://www.top500.org/news/microsoft-takes-fpga-powered-dee...
It is certainly hard to compete against an NVidia V100 with its tensor ops.
It can get interesting though. There are quite a few graphs that tensorflow cannot place on a GPU, especially with RNNs. An FPGA does give you more flexibility in architecture which was, and would remain one of the motivations for MS to choose them as a platform.
Microsoft is seeing ~90Tops for their MS-FP8 datatypes: https://www.top500.org/news/microsoft-takes-fpga-powered-dee...
It is certainly hard to compete against an NVidia V100 with its tensor ops.
It can get interesting though. There are quite a few graphs that tensorflow cannot place on a GPU, especially with RNNs. An FPGA does give you more flexibility in architecture which was, and would remain one of the motivations for MS to choose them as a platform.
Been very curious to see a comparison between standard computing hardware and neural specific hardware. Nice article!
The price is a fair complaint but it's a bit misleading - actually boards aren't $50k for people who are serious about building an application with them.
The bigger concern for me is that all the benchmarks I've seen show janky FPGA solutions that cut a lot of corners barely compete with fairly mature GPU solutions. So you're spending 20x the development time for 90% of the performance with 95% of the algorithm. The big promise from Intel was that the Stratix 10 chips would be 5x as big with 2x the Clock speed and your CNN algorithm would just magically kick arse. That doesn't seem to have happened - Stratix 10s are almost impossible to get hold off, the promises about performance are getting watered down and of course, by the time you can actually get an application running on it GPU technology has moved on and the performance still isn't competitive. At this point I don't see FPGA being a player in the AI market except for extremely embedded solutions.