AKIDA benchmarking?, page-130

Hi everyone

I thought I would just reactivate this thread about AKIDA benchmarking because it could be that someone reading this has actually managed to get their hands on one of the Development systems that we have been able to order for $4995 or $9995 since the 21st October https://brainchipinc.com/brn-taking-orders-of-akida-ai/

Brainchip's website includes a long list of models with accuracy figures obtained using the Akida simulation system https://doc.brainchipinc.com/zoo_performances.html
That table also provides information about the amount of memory used by each model - ranging from very small (22.8KB for Audio Keyword Spotting) to very large 34825.2KB for one of the ImageNet implementations). Some of these presumably can't be run directly on the chip - which apparently has 8 MB of on-board SRAM - and would need to use external memory. I understand that the PCIe board that is provided with the development systems comes with 3GB of external memory, so that should not be a problem, but it presumably has an impact on throughput.

Unfortunately, BrainChip doesn't provide any information about actual performances levels obtained with the chip - things like the number of images processed per second, for example. So, I'm hoping that someone with access to one of the boards could just provide the information for the rest of us. Any takers? I really don't think that this would be revealing any top secret proprietary information - it's just a few numbers. But it would be very useful.