Challenge #26 BrainChip - jtristan123/HW-for-AI-ML-ECE-410 GitHub Wiki
Tasks:
- Listen to the EETimes BrainChip podcast (47min) at https://www.eetimes.com/podcasts/brainchips-ip-for-targeting-ai-applications-at-the-edge
- Put what you learned in perspective in a short write-up. In particular, compare the BrainChip approach to GPUs and other neuromorphic chips that we have seen in class.
Write-up
NOTES and quotes from podcast and transcript:
- In AI terms, BrainChip is a relatively old company. It was founded by Peter Van Der Made in 2004 to commercialize neuromorphic devices and dynamic learning.
- BrainChip has developed something called Temporal Event-Based Neural Networks, or TENNs
- And so what’s inside this recurrent neural network? You have this thing which is essential: it’s a state. What is a state? It’s a summary of everything that has come before that the network has been exposed to.
- In the TENNs approach, this is based in the basic architecture—versus maybe other approaches people might use which would be to first detect an object and then put that into some sort of mechanism that maintains continuity in an ad hoc manner. So the TENNs approach is, in my opinion, much more elegant and complete
- Causal basically means that the current prediction only depends on the current input and all the previous input.
- The difference between a recurrent network and a typical, let’s say, convolution network or feed-forward network is that a traditional neural network, or a feed-forward network, does not contain internal states. The internal state can be considered as a memory of the network. So what happens is that if you feed the data to the system, the network will intrinsically memorize what it has seen in the past in a very abstract, compressed state. So each layer of the network contains an internal state, and that is the traditional recurrent network.
- And one advantage of TENNs is that we don’t actually suffer from this sequential sort of blockade because we have a—we admit like a parallel form for our network. So, in some sense, I guess our network is a recurrent network that can also overcome the limitations of traditional nonlinear recurrent networks.
- But I guess the core difference between Mamba and TENNs is that Mamba uses internal state as well, but their internal state is actually expanded—meaning that they have a very large internal state bank. So this is very good for GPU inference, obviously, as proven by their recent popularity—they can do large language modeling very well, especially for hybrid models, due to their large internal state bank. But one of the disadvantages is that this is very poorly supported for edge inference, because you have to maintain all these large states and they have to continuously evolve them, which is not amenable for lightweight compute. What’s the difference between an event and a spike? A spike is usually defined as a single value—either the appearance of something (a one) or non-appearance (a zero). An event could also carry information additional to that zero or one; it can carry the activation level, say, of a previous neuron.
- At BrainChip, we are currently finishing off our Akida 2.0 architecture. And this is an expansion on all of BrainChip’s prior work. It’s essentially a neuromorphic chip, it’s event-based, and it makes use of the intrinsic sparsity in signals to achieve high levels of efficiency.
- In our case, with an Akida system, we have, at some point, a level of parallelisation we can use—it’s always there. But it will also have sparsity or, in other words, only focus on non-zero activation in order to be more efficient.
BTW I asked Chatgpt to make a table to compare the other chips we seen chatgot:
make me a talbe that goes with it
| Feature / System | BrainChip Akida | GPUs (e.g., NVIDIA) | Intel Loihi / IBM TrueNorth |
|---|---|---|---|
| Type | Neuromorphic (event-based) chip | Traditional AI hardware (parallel compute) | Neuromorphic (spiking neural networks) |
| Power Usage | Very low (edge-optimized) | High | Low |
| Processing Style | Continuous, event-based | Batch-based | Event-based |
| Target Applications | Real-time: voice, gesture, eye tracking | Heavy workloads: vision, LLMs, simulation | Research: brain-inspired tasks |
| Memory / State | Maintains internal state (TENNs) | Stateless unless explicitly modeled | Spike-driven memory (neuron-like) |
| Ease of Training | Relatively simple | Requires high compute & large data | Complex (SNNs are harder to train) |
| Biological Inspiration | Moderate (temporal context awareness) | Low | High (models neurons and spikes closely) |
| Best Use Case | Embedded AI at the edge | Cloud/data center AI | Experimental neuromorphic research |
I guess overall compared to other neuromorphic chips like Intel’s Loihi or IBM’s TrueNorth, Akida uses something called Temporal Event-Based Neural Networks (TENNs). These networks can remember information over time, which helps them understand patterns that change. Unlike GPUs, which use lots of power and process large amounts of data in batches, Akida runs continuously and responds to events as they happen more like how the human brain works (I see why it called brainchip). This makes it great for things like voice commands, hand gestures, or eye tracking where quick response and low energy use matter.