Hi JK,
Akida's forte is identifying (classifying) input signals from sensors. In simple applications this may be sufficient to trigger a direct response or action.
However, in some cases the Akida output is used as an input to another CPU/GPU (von Neumann processor) to form part of that computer's program variables.
In the first case, the entire process gets the full power saving/speed improvement from Akida.
In the second case, the benefit is the reduction in power/time which Akida brings to the classification task while the CPU performs the remaining processes under the control of its software program. This is important because the classification task carried out on a software controlled CPU uses very large amounts of power and takes a relatively long time.
Classification of an image on a CPU uses CNN (convolutional neural network) processes which involve multiplying multi-bit (8, 16, 32, 63) bytes representing each pixel on the sensor. Multiplication involves the number of computer operations determined by the square of the number of bits in the byte, so an 8-bit byte multiplication would involve, 64 computer operations. For 32-bit bytes, 1024 operations are required to process the output from a single pixel, whether it's value has changed or not.
On the other hand, Akida ignores pixels whose output value does not change, and only performs a computer operation for the pixels whose output changes (an event). This is "sparsity". In addition, in 1-bit mode there is only a singe computer operation for each pixel event.
For example, the sparsity may reduce the number of events by, say, 40%.
Even in 4-bit mode, Akida only needs 16 computer operations, and that only for pixels whose output has changed.
Hence there are large savings in power and time in using Akida to do the classification task compared to using, eg, a 32-bit ARM Cortex microprocessor.
While the rest of the program may be carried out on the microprocessor, this uses comparatively little power compared to the power the microprocessor would have used performing the CNN task. So there are still large power savings to be made by using Akida in "accelerator" mode as an input device for a von Neumann processor.
The other point is that Akida performs its classification independent of any processor with which it is associated. For example, Akida 1000 includes an ARM Cortex processor, but this is only used for configuration of the arrangement of the NPU nodes to optimize performance for the particular task, but the ARM Cortex plays no part in actual classification task. The ARM Cortex does not form part of the Akida IP. Akida is "processor-agnostic" and can operate with any CPU.