And this is already possible with @BrainChip_inc #AKD1000, let alone the newer generations. The technology clearly isn’t standing still. So why are we still waiting for commercial traction? @computeontology
#AI #Brainchip #neuromorphicAI #Semiconductors
I'm going to call this DragonFly. Four BrainChip Akida neuromorphic chips on one PCIe switch, forming a neuromorphic RAID.
Most don't know that the AKD1000 has an ARM Cortex-M4 chip. I finally had some time to take advantage of this part of the chip's architecture, something I didn't do with the demos.
I made the fourth chip's onboard M4 a controller for the other three, over the switch, with the host CPUs out of the loop. Lose a chip's model? The pod rebuilds it bit-exact from parity. On silicon.
RAID, but for neuromorphic chips instead of disks, another first for the neuromorphic world.
Three chips hold model programs. The fourth holds XOR parity and becomes a hot spare, its ARM M4 pulls the others' data over the switch, builds parity, scrubs for corruption, and reconstructs a lost chip from the survivors.
The chip's own onboard microcontroller drive peer chips across the PCIe switch. Proven, bit-exact: write to a peer, read from a peer, a mailbox + doorbell protocol between the chips' M4s, all with the server CPU sitting it out, automatic failover and redundancy.
So, the RAID automatically protects what the chip LEARNED.
Akida learns on-chip, in the field, with no labels. I ran a real learning step on one chip and the pod journaled that exact learned change and folded it into parity. Lose the chip, keep the lesson.
So a neuromorphic chip that learns something in the field can have that learning reconstructed onto a replacement from its neighbors. The model's redundancy lives on the card, not somewhere else.
Where and how this would fit in the real world:
* On-site learning where you can't afford to lose it: drones, remote sensors
* Disconnected or contested environments: no phoning home to re-download a model
* Space and radiation: bit-flips and chip loss are expected; on-card scrub + rebuild is the answer, and neuromorphic is already low-power
* Industrial and medical appliances: model integrity survives a chip fault with no service call.
* Large multi-chip neuromorphic fabrics as they scale: host-free resilience between chips becomes a requirement, not a luxury.
2
110
#ultralowpower, #alwayson #AI for #medical implants and #wearable sensors. Detecting anomalies locally, reducing latency, battery consumption and data transmission. The list goes on and on @BrainChip_inc
#BrainChip #EdgeAI #MedTech #Akida
lnkd.in/p/gBggSX6S
3
96
Me retweeted
⭐️ BrainChip - Akida: Schutz gegen Cyberangriffe
Durch die Kombination aus offener Software-Integration und neuromorpher Hardware entsteht ein kosteneffizienter, reproduzierbarer Schutzstandard für kritische Kälte-Infrastrukturen.
lnkd.in/p/eBVzpgdb
@BrainChip_inc #Akida
Me retweeted
⭐️ BrainChip - Neuromorphyx at Swiss Robotics Day in Zurich
"Start with NeuroSight if you already have your own compute platform,BrainBoard1500 to evaluate Akida inference, or NeuroReflex for the combined stack with end-to-end neuromorphic vision pipeline."
@BrainChip_inc #Akida
Me retweeted
Chips that learn on the job have a problem: when the rules change, old habits get in the way. Our @BrainChip_inc AKD1500 relearning a new rule got stuck at 46-51/96.
One 23 KB "forget this part" write and it relearned to 74-76, same as a fresh chip, in both runs. its other skill stayed intact:
This is now documented as a technical paper, you can find it here, citable via the DOI on Zenodo:
kevindjohnson.org/Forty-Test…
This is a PCIe switch, it now holds 4 AKD1000's in one PCIe slot on an Intel i3-3220, 3.30 GHz, Ivy Bridge via an old HP-branded Foxconn 2ABF motherboard. The switch can actually do PCIe Gen3 but here is limited only by the CPU to 2.0. Basically, this means we can use multiple Akida cards in virtually any PCIe slot. Where we can put four, we can put many more.
To the @BrainChip_inc community!
What’s your take on the latest developments surrounding renewBRN and yesterday’s board resignations?
Are you backing BRAINCHIP and its current direction, or do you support renewBRN and its proposed changes?
Let’s hear your #thoughts. #Opinion
Ladies and gentlemen… Dr. Steven D. Harbour!
“We don’t have to wait…@BrainChip_inc #Akida is commercial available”
lnkd.in/p/gkPvnRGR
Countless use cases for @BrainChip_inc’s #Akida #Neuromorphic chips already exist. Yet here we are, still debating “if”, “when” and “probably”, @elonmusk, @MercedesBenz, @Apple @Arm etc
The technology is here. Don’t pretend like it’s not!
BREAKING: If a post on X starts with “BREAKING”, there’s a good chance the story isn’t even true. So take a brake
#news #x #clickbait #nothingtondowithneuromorphicAI or @BrainChip_inc .. not even #Semiconductors so…
4
124
Interesting turn of events in the RenewBRN saga at BrainChip! Let’s see who they bring to the table… @computeontology any Idea? I…I don’t know…
asx.api.cmfyapp.com/asx-rese…
Weekly reminder
Somebody sent me an AI-generated comparison that sets my neuromorphic work against Sparxity (formerly SpiNNcloud). It praises the engineering, but calls my work a “hack,” sends Akida back to the edge, and gives the data center to Sparxity.
The argument rests on one presupposition: a chip can only do what a marketing brochure might say. That’s taxonomy passed off as physics. A chip’s capability belongs to the silicon, not the marketing approach.
Take the main objection, that the AKD1500 has only 1 MB of memory. That figure is on-chip memory. The chip works out of host memory over DMA, and we’ve taken that window from 4 MiB to 768 MB per chip. Across 22 chips, that’s nearly 17 GB of addressable model memory, and it grows with every chip you add. Separately, we’ve run a 78 MiB layer bit-exact on a completely stock chip by splitting it. The same software runs on Epyc servers, NVIDIA Jetson, Qualcomm boards, and 131 cloud nodes on six continents.
I’ve never positioned Akida as a large-language-model engine. My papers say the opposite. The argument is that neuromorphic processors are peers to CPUs, GPUs and quantum processors, and each kind of work belongs on the processor whose physics suits it.
Placement runs on workload management capabilities already in production across thousands of sites, data centers included. On that view a Sparxity rack isn’t a rival, it’s one more tier we can work with in advancing Akida’s reach in a heterogeneous compute ontology. Both/And, not Either/Or.
What Akida suits is the standing watch, among other things, and the data center has one. The costliest data movement there isn’t on a chip’s bus. It’s copying petabytes out of storage to wherever the analysis runs. So I’ve built the Engram Storage 1000: data-center storage servers with 16 to 64 Akida processors beside the data. Data is judged as it’s written and archives are re-read where they sit. Only what the fabric flags goes on to GPUs. Only the answer moves.
A “hack” doesn’t come with a chain of trust or IEEE peer-reviewed design proofs. Akida serves models through vLLM, the industry’s standard serving layer, on a secure path. Every model is signed, verified bit-exact on reserved silicon before release, and checked again at use. Processors are compared with each other across the fleet, so correctness is checked continuously rather than assumed. All of it runs under the NSA’s CNSA 2.0 quantum-resistant suite, ahead of the January 2027 requirement for new national security systems.
As for the comparison, read the sources, not the AI summary. SpiNNaker2 was designed for brain simulation and its LLM positioning is recent. Its installations are at universities and research centers. Jülich is a two-rack research system. The T-Systems deal is a letter of intent around a chip slated for 2028. The efficiency multiples are a homepage chart label and a single-board result with the paper still to come.
The software described in its published sources is a low-level programming capability through research toolkits, reaching boards over a mere 1 Gbit Ethernet while AI data-center fabrics run at 400GbE and up.
Those sources describe no signed or verified models, no post-quantum cryptography, no tenant isolation, no vLLM or standard serving endpoint, and no production workload management placing work across a fleet, and no distributed file system.
That’s still a research platform, not an enterprise one. None of that is a knock on serious people doing innovative work with real hardware. It just isn’t evidence that SpiNNaker2 is somehow superior to Akida, alone or orchestrated through a heterogeneous compute ontology.
“Edge” is a perspective, and so is “data center.” Neither really speaks to what’s actually needed for any given workload. The point here isn’t about hardware differences and what’s better. The point is that being different makes more work possible where it needs to run.
That’s why we can say, “Run Akida anywhere. Run Akida everywhere.” And, I will keep saying it.
teapawt.net/computeontology/status…
5
178
Me retweeted
We plotted the AKD1500's predictions back onto a mouse brain atlas.
It learned from 8,192 unlabelled, encoded gene profiles from one mouse. The host used 60 labelled examples to name its memory records, then classified 9,511 cells from another mouse using the chip's scores.
Agreement with atlas labels: 64–69%, versus 18–23% for untrained memory, balancing classes and sections.
Second image: averaged across three repetitions, label-free memory scored higher than the supervised rule with 1 or 4 labels per type; supervised scored higher at 16. Only the ringed AKD points at 4 labels use chip scores. All other points are software results.
The chip never saw positions; the host added them for this map. K-means, a conventional software method, had the highest average at every label budget.
Me retweeted
Labels are expensive. Here, the chip learned without cell-type labels; the host assigned names afterward using 60 labelled examples.
Watch the AKD1500 learn from 8,192 unlabelled mouse-cell gene profiles, then score 9,511 cells from a second mouse. This is a continuous live screen capture of the chip run, using saved data.
The host encodes the profiles, uses 60 labelled examples (four for each of 15 cell types) to name the learned memory records, and puts the predictions back on the brain map. The chip never sees positions.
We repeated one earlier run: every score and measured memory state matched exactly. Agreement with atlas labels was 69.0%, balancing cell types and the three sections. This demonstrates learning and scoring on the device; it is not a new accuracy test. Conventional software methods scored higher in the earlier comparison.
Dr. Steven d. Harbour knows what’s possible with #neuromorphiccomputing, and he’s had first hand exposure to @BrainChip_inc ’s #technology. Perhaps we’re getting closer to the point where the market starts recognising what has been possible for years.
lnkd.in/p/gYQH4-qB
Me retweeted
A box fragment kept beating the whole box. We kept the same scorer and regions and added a patch map. Mac replay of two MechDog walks got both answers right on 133/137 frames, up from 116. False calls fell from 9 to 0 on 24 archive requests. Kept 234/239 older answers.
Me retweeted
The patch scorer ran on the AKD1500. All 618 answers on the same 309 saved frames matched the Mac, misses included. Scoring only patches that could affect the box choice used 76% fewer chip queries than a full-frame scan. The host encoded the images and chose the regions.
What? 😳
This chip can change its learned memory right where it runs.
We fed a @BrainChip_inc AKD1500 patterns from mouse brain cells, plus their known cell types. genes are instructions cells use, and different cell types use them in different amounts. that's the pattern we gave the chip.
Five memory banks, trained one at a time. 6,417 lessons. every recorded response and final memory matched our software prediction exactly.
We also tested a surprise from an earlier failed run: with learning enabled, unlabelled inputs can still change memory. All 80 recorded responses and 25 memory snapshots matched the prediction we filed.
The video includes a live repeat using 16 saved cell profiles with no cell-type labels supplied; all 16 score sets and all 16 memory readbacks matched the earlier run exactly.
A step toward devices that adapt locally. The computer still prepares inputs, provides labels and combines the banks. Accuracy was measured in software, not on chip.
Me retweeted
⭐️ BrainChip - Frontgrade Gaisler / GR801 at Space Tech Expo Europe
What if satellites could make critical decisions in orbit without waiting for instructions from Earth?
Frontgrade Gaisler is working to make that possible through neuromorphic processing➡️#Akida @BrainChip_inc






















