In a previous post, we described applications of graphic processing units (GPUs) to neuromorphic computing. GPUs are a good fit for simulating neurons, and recent industry trends will most likely increase the appeal of this computing substrate. However, GPUs may not be the only player in attracting the attention of neural modelers. A Lexington (MA) startup company, Adapteva, has recently introduced a chip that looks even more appealing than GPUs.
Epiphany, the Adapteva multicore processor, has been designed with the primary goal of accelerating applications in low-power devices such as smartphones and tablets, but also servers (in particular, to reduce monstrous power demands of large servers). The chip designed by Andreas Olofsson (CEO of Adapteva) is scalable to thousands of cores on a single chip, and can be places alongside CPUs to provide real-time execution of diverse applications (currently, 64 cores in smartphones and up to 4,000 cores in servers).
Adapteva aims to provide a flexible accelerator with multiple cores that can perform a number of functions: basically, a floating point processor specifically designed to be power efficient. En example: a chip running at 1GHz with 16 cores can consume less than 1 watt of power. Below the sketch of the architecture.
It turns out that the architecture Adapteva is working on seems to be very compatible with the demands of neuromorphic algorithms. If you remember the December 2010 IEEE Spectrum article, it's a very similar architecture to what we describe in the article being the "ideal case".
This chip could allow to put much more computation on mobile robots without having to deal with FPGAs or ASICs, and it looks like it will support a much simpler multicore programming paradigm than GPUs.