The Pros and Cons of SpiNNaker Hardware
- The current agents run on both Nest and SpiNNaker.
- Load time on SpiNNaker is longer than Nest.
-
Vision Net Size | Nest Load | SpiNNaker Load | Nest Run | SpiNNaker Run |
20x20 | 20 seconds | 1 minutes | 1 to 10 | 1 to 1 |
40x40 | 3 minutes | 5 minutes | 1 to 100 | 1 to 1 |
100x100 | 3.5 hours | 5 hours
| 1 to 1000
| 1 to 1 |
- Note that the Nest times are on my laptop.
- There's also time to get the spike results back after the simulation,
but that's not generally a problem.
- When you actually get to running simulations that take a few minutes of
real time, the benefits of the hardware are huge.
- I want to note that I'm running on just one 48 chip board.
- There are some downsides to the hardware.
- The board crashes. This is not often a problem with the agent, but it
does seem
to crash relatively frequently in my automated tests.
- Bursty input: I have to take pictures of the environment, pixelate
them and send them to the board. There is an image translation problem
(that takes up processing that makes it longer to send), but there is
also a problem that spikes don't get sent regularly. In Nest
it really is like clockwork.
- This might be improved by using another channel than the TCP/IP cable.
- Also, in Nest, if you'd like, you can cheat and write your own learning
rules in python.