First mover advantage, page-70

On page 8:

"

Our model has resolutely been developed in the philosophy
of spiking neural networks. As a result, it is essentially
implementable on dedicated neuromorphic hardware,
such as Intel Loihi [3], IBM TrueNorth [1]. These chips
can leverage the binarity and sparsity of spike tensors
navigating through the network. In addition, we believe
that our model being feedforward and requiring a reset
on all of its neurons at each timestep is not a problem,
because resetting membrane potentials is actually less
costly than applying a leak. Therefore, statelesness can be
seen as an advantage over recurrence in spiking models
with similar performances. However, we are aware that
current neuromorphic chips are initially designed for the
implementation of stateful units, and acknowledge that we
do not leverage this feature. Consequently, we believe that
it rather fits to dedicated hardware for stateless models with
sparse activations quantized on 1 bit. We therefore consider
that Brainchip’s Akida chip [35] is a good fit.

"

Disclaimer: one of the authors Timothee Masquelier is the "T" in JAST which Brainchip has exclusive licence to.

For a good understanding of JAST refer UIUX's thread "All roads lead to JAST"
https://hotcopper.com.au/threads/all-roads-lead-to-jast.5683524/