Hi manny,
The M2 card has 20 NPUs (equivalent to 4 nodes in later versions, but the geography is a bit different). I think a full sized Akida 1 could have 1024 NPU, and this could be ganged together with other Akida 1s, so the M2 is simply a demonstrator. For a mass commercial market, (> 1 million units), it could all be accommodated on a single chip, taking advantage of Akida's multitasking abilities. How good is that ? - multitasking on a massively parallel neuromorphic processor.
https://shop.brainchipinc.com/products/akida-m-2-card
| Akida M.2 Card, powered by AKD1000 IC with an ARM M.4 CPU plus 20 Akida 1.0 event-based processing NPUs in a mesh. Akida M.2 AKD1000 Card accelerates CNN-based neural network models using BrainChip’s ultra energy-efficient, and purely digital, event-based processing architecture. |
- Form factor: M.2 2260, B+M Key and E Key available
- Host interface: PCIe PHY 2-lane
- Memory interface: LPDDR4 via DMA
- CPU: 32bit ARM M.4
- NPU: 20 x Akida 1 Neuron mesh
- Peak INT8 GOPs: 1.5 TOPs
- On-chip memory: 8MB high-speed near compute SRAM
- Clock frequency: 300MHz
- Operating temperature: 0 – 70°C
- Thermal solution: no fan or heatsink required
- Typical application power: 1W
As you say, the advantage for Kevin's project is that the M2 is fully working COTs with ARM processor and comms interface on chip - plug/train/configure and play, and training is a breeze. I think the ease of training is something which will be of interest to olde worlde CNN AI engineers/programmers.
What this illustrates is that there is a potential market for the clunky old Akida 1 IP, because there will be a price differential for Akida 2 IP, and a much greater price differential for Akida 3. Those later models have higher precision and lower latency, but take comparatively larger slices of wafer real estate per chip.
The other thing is that our very good friends at MF have always cited the big budgets of the tech giants for developing AI as a downer for BRN.
