Bravo
If ARM was an arm, BRN would be its biceps💪!
Hi Boab,
I must admit I was a tad concerned hearing Dr. Harbour’s comments given Akida is digital, not analog.
At 25:00, he says:
“BrainChip’s Akida probably has, if not the first commercialised neuromorphic processor, is amongst the first commercialised. You can buy that today and you can get support with it. Again it’s a digital-based neuromophphic processor with a neuromorphic mesh, if you will. Full blown neuromorphics will involve something that’s very analogue-like - like the mind. And we’re we’re working on that. In fact I’m working on a specific program right now of doing that.”
At 31:00, he adds:
"Dr Carver Meade at Caltech. And I’m going to give him credit for the start of this, envisioned an analogue processor, if you will. This was in the early 80’s, maybe even before that in the late 70’s. And so he started that and at the same time, the digital concept processor was a competitor. And, at the end, we’re going to see, after the long haul, the analogue processor is going to win out."
My non-expert read is that Dr. Harbour clearly respects Akida as one of the first commercial, supported neuromorphic options today, while also betting that analog approaches could dominate long-term if variability and accuracy are tamed.
Given these remarks, I poked around on the interwebs to see if I could find out what analog project he might be working on and found a talk dated 16 October 2022 (see video below).
At 8:10, he says “we made silicon brain” and he shows a slide (see below) titled “Our work: Analog Memristor.”
I can only speculate that Dr Harbour may be intending on continuing his prior analog/memristive neuromorphic work toward a processor?? At least he frames it as a “long haul,” which hopefully indicates there's plenty of runway for Akida to build market share in the meantime.
I’d like to hear Tony Lewis’s view on Dr Harbour's remarks, but I'm not sure that IR will even respond to my email, as they didn't respond to the last email I sent them. In the meantime, I asked ChatGPT for a neutral take on Dr. Harbour’s statements, as well as its opinion on analog variability, etc. for what it's worth.
PS: I'm also going to ask Tony if BrainChip sees analog/mixed-signal blocks as complementary to Akida in future hybrid systems.
ChatGPT said:
Great spot—and yes, you heard him right.
What Dr. Steven Harbour (Parallax) is saying
- Parallax posted an interview/profile with Dr. Steven D. Harbour (Director, AI Hardware Research; Neuromorphic Center of Excellence) this year where he talks up neuromorphic for defense/EW and edge systems. Parallax Advanced Research
- His comments about Akida being among the first commercial neuromorphic processors you can buy and get support for are fair: Akida is a digital neuromorphic IP/processor with a mesh architecture and shipping toolchain. (That’s the “today” story.)
- He then pivots to the “full-blown neuromorphics” = more analog-like view—rooted in Carver Mead’s original analog VLSI neuromorphic work in the ’80s. History of Information+1
Is analog “less stable” than digital?
Short answer: it’s more variable, unless you engineer around it.- Analog & mixed-signal neuromorphics face device mismatch, process/voltage/temperature (PVT) drift, and noise; emerging analog synapses (e.g., memristors/ReRAM/PCM) add cycle-to-cycle variability and resistance drift. This can degrade accuracy unless you calibrate/compensate. arXiv+3Frontiers+3AIMS Press+3
- Reviews and roadmaps explicitly note these reliability/variability challenges for analog in-memory devices (RRAM/PCM): drift, non-linearity, device-to-device variation, filament randomness. AIP Publishing+2ScienceDirect+2
- By contrast, digital neuromorphics avoid most mismatch/drift and are easier to reproduce across process nodes; that’s a big reason digital platforms (like Akida) reached market first. Nature
Why Harbour (and others) still think analog could “win” long-term
This view comes straight from the analog VLSI neuromorphic lineage (Mead et al.) and today’s analog in-memory research:- Energy & area efficiency: analog compute-in-memory can perform MACs where the weights live, slashing data movement (the big energy cost) and reducing transistor count per operation—huge for always-on edge. Polyn Technology
- Continuous-time physics: analog circuits naturally implement certain neural dynamics (integration, leakage, spiking) with very low power and microsecond-class latency, which maps well to sensing at the edge. WVU Community
- Mixed-signal best of both worlds: many see the endgame as hybrid—analog for the heavy, low-energy linear ops and temporal dynamics; digital for control, memory management, safety and programmability. Patsnap Eureka
Putting it together (and reconciling with Akida)
- Today: Digital neuromorphics like Akida are commercial, supported, and predictable, which is why you can buy them now and integrate them relatively quickly.
- Tomorrow: If (big “if”) analog neuromorphic platforms tame variability and ship robust toolchains, they can deliver even lower power per task—which is why Harbour says they may ultimately “win.” In practice, the likely winner is mixed-signal: analog compute blocks plus digital orchestration.
Last edited:
Project Overview
Hardware Requirements
Then started selling off again to get the price down so they can buy them back cheaper again. If they are true to form they’ll be scooping them back up again before long.
]
