It looks like you're new here. If you want to get involved, click one of these buttons!Sign In
It looks like you're new here. If you want to get involved, click one of these buttons!
Unfortunately, the only information I've been able to find online regarding this stick and the internal hardware is a two page product brief of the chip and some very oblique YouTube videos. While this seems like a very interesting product and could promise quite a bit, at the moment, I struggle with justifying the time and costs associated with bringing this into my R&D and dev environment without better information. The last thing any of us want to do is spend a great deal of time and energy trying to make a device do something it may not be able to do. From the perspective of an engineer, a device with lack of technical data might as well be a Christmas tree ornament.
Given this is a new technology with a new chip in a new field, what I would like to see, and I'm sure others would as well, is a white paper that outlines the operational specifications of the chip. Items of interest to me (and I'm sure others):
* Performance across multiple NN architectures, sizes, activation functions, etc.
**** A generalist specification of "teraflops within a 1 watt envelope" doesn't exactly mean much anymore given the high variability of ALU, VALU, and MALU architecture and design.
**** The USB interface also changes things quite a bit. The size of the data being sent to and from the chip over USB is foreseeably a bottleneck for how fast and how well the stick will work for different tasks.
Performance comparisons (ie "benchmarks") between other common low power processors running the same NN (ex Atom processors, smaller i3 processors, and various ARM processors like the Broadcom BCM2837)
**** If the goal and the selling point of the device is to provide the ability to solve NNs quickly and at low power, then we need to see how much more efficient it is. Granted, I'm fairly certain a SoC that is optimized for matrice operations will do far better than the GPU of an ARM SoC meant for phones or DVD players. That said, however, if there is only a modest performance boost for certain NN architectures/sizes, it would be wise (and responsible) to tell developers, if only to foster good relationships with them.
**** It would be also beneficial to see pass/fail rates of the stick verses (presumably) smaller ANN that execute at roughly the same speed on common mobile processors.
Power Usage characteristics across various NN sizes/architectures as well as across any different hardware power profiles
Full suite of mechanical, dust, moisture, electromagnetic, and temperature tolerances (aka environmental tolerances).
**** If these are going in quadcopters, rovers, robots, etc, it is incredibly important to understand the environmental tolerances of the stick. In a nutshell, if the stick is only usable in indoor environments, it's applications are limited.