Hasn't the Akida 1500 been (more quickly) developed at the insistance of a client? Who knows which one.
Hi Dhm,
Akida 1500 has 8 nodes compared with Akida1000's 20 nodes. 8 nodes is 32 NPUs so a reasonable NN can be configured.
It also does not include the ARM Cortex processor.
Some other peripherals, (possibly comms interfaces?) have been left off.
It also includes CNN2SNN input and probably an SNN2CNN output so it can slot into existing CNN applications.
While we do not know what, if any, new features are added, this means it is a much smaller chip than Akida 1000, which in turn means more chips per wafer, so a much cheaper chip without the impost of an ARM licence/royalty fee.
It does meet the requirements of a recent NASA SBIR for 22nm FD-SOI without a processor.
The compact size would enhance the 1500's compatibility with near-sensor applications.
I don't know what the maximum throughput (fps or equivalent) of the 1500 would be, but given that Prophesee was not entirely satisfied with the Synsense (analog?) SNN's ability to match Prophesee's speed, would a reduced Akida still meet Prophesee's performance?
Whatever the application is, there will need to be an associated processor to configure the 1500. With the PCIe demonstrator board, that may be, eg, a Raspberry Pi processor. For the 1500 IP, Akida is processor agnostic.
Providing th SNN2CNN output is interesting in that it indicates one potential use id with legacy CNN software. Thousands of hours can be invested in developing software, so customers may be reluctant to discard that investment.
Any CNN software will have an associated 8-bit or more model library (eg, millions of hand-classified images) which also represents a very substantial investment of time and expertise.
FD-SoI is low leakage due to the (I)nsulator, and thus low power. It is also radiation hard, so it can withstand hostile environments.
Such extremely low power means that it may be able to be installed in applications where cooling is difficult to provide, and acting as an accelerator, 1500 will also relieve the main processor of much of the heavy lifting, again reducing the cooling requirement for the main processor.
So we are looking for a customer who has a need for near-sensor classification/inference (and ML?), has a significant investment in (legacy) CNN-dependent software, and a need to conserve power. Such legacy software would have the virtue of being extensively field tested and de-bugged, so the owner would be reluctant to build new software from the ground up.
Automotive, for one, comes to mind. There is massive investment in ADAS/AV software, all of which would have been designed for CNN.
DVS does have sparse sensor data, but Prophesee generates data at enormous speed. BrainChip are fully aware of Phophesee's capabilities, and, no doubt, the opportunity for near sensor inference/classification/ML.
While 1500 seems to meet the NASA SBIR, it is not a high volume consumer, so may not be the primary target, unless USAF and DoD generally are involved on the quiet.
Who knows?