Hi Manny,
Perfection is a lofty goal in such a rapidly changing technological environment.
The initial demonstration Akida chip had 1-bit weights and activations. It was ultra-fast and ultra-power efficient. The EAP feedback led to the change to add 2-bit and 4-bit weights/activations and CNN2SNN to improve accuracy and to provide backwards compatibility with pre-existing CNN system models.
Thus the original 1-bit weights/activations format is still available in the updated Akida 1000 for applications where this provides sufficient accuracy, and still provides the same power efficiency and speed.
So the original Akida demonstration chip and Akida 1000 were both technical successes in that they performed as they were deigned. Commercially, the Akida 1000 was never really given the chance to fully test the market as the marketing effort was switched to IP licensing, a strategy which substantially changed the target market from end users to chip makers and major chip consumers. The original marketing plan was to supply chips for those who required less than 1 million chips, and to licence the IP to those who required more. Thus the general market was not educated to the value of the chip. Potential customers would need to have known about Akida's capabilities so they could determine if they had a use case.
I'm sure the potential profit margin for the IP far outweighs the profit from the chips, while, on the other hand, the sale of chips has a much shorter lead time than IP licensing.
So, as Sir Humphery would have said "That is a very courageous move ..."
Before the rug was pulled, Akida 1000 had some limited commercial success, but further potential success was curtailed, or rather stalled, by the announcement of Akida 2.
"Stalled" is more appropriate, as Akida 1000 is not dead, it's just lying doggo.
Akida 1500 is simply Akida 1000 without the ARM Cortx processor, and, in its current embodiment, in 22nm FDSoI, though obviously it could be made in any CMOS process.
After Akida 1000 was commercialized, customers requested LSTM and the recently-developed Transformer capabilities, so, to some extent, Akida 1000 was superseded for applications which needed these capabilities. But Akida 1000 is still the best option for applications which do not require these capabilities.
Akida 2, with TeNNs, ViT, and Long Skip, provides a solution which, in some respects, surpasses these capabilities.
but bad news for Qualcomm I guess, considering they're working heavily with RISC-V chips.
