Me (@justforfunrun) profile picture
Joined July 2026
Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)Photo by Me (@justforfunrun)
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
3
22
287
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
2
19
244
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:
2
13
351
This is now documented as a technical paper, you can find it here, citable via the DOI on Zenodo: kevindjohnson.org/Forty-Test…
When I say the use cases for Akida are endless, I mean it. Run Akida anywhere. Run Akida everywhere.
10
34
581
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.
5
30
363
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.
1
7
119
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.
1
10
37
963
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.
1
3
8
336
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.
4
9
130
⭐️ 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
1
14
36
437
⭐️ BrainChip - Neuromorphyx + fortiss Neuromorphyx: "Axel von Arnim is leading amazing neuromorphic work at fortiss and we stay committed to supporting his team. Hats off and thank you for the wonderful conversations!" @BrainChip_inc @neuromorphyx
11
17
139
⭐️ BrainChip - Swiss Robotics Day Axel von Arnim: "Robots that must track, grasp and navigate need fast perception on very little power. That is exactly what event cameras and neuromorphic computing are built for, and it is what my team at fortiss works on." @BrainChip_inc
1
14
38
479
Adding AI to an edge device usually means replacing the processor you already ship, or adding an accelerator with its own power and thermal budget. There’s a third option: keep your host, add a dedicated AI co-processor. AKD1500’s two differentiators: 1. Sparsity-driven low power: skips inactive values instead of computing them 2. On-chip learning: adapts to users after shipping Read the Blog: na2.hubs.ly/H08_jn00
5
28
92
1,761
Replying to @BrainChip_inc
@BrainChip_inc 's documented Akida Edge Learning interface trains only the final layer. On one AKD1500, we ran four connected learning layers joined by three fixed links. All four changed connections on every one of 64 inputs. After each input we read all four learning layers' connection memory back from the chip. Every recorded state matched our reference exactly. Before learning, all 4,352 reference output values matched too. In the freeze controls, disabled layers stayed unchanged while enabled layers kept learning. Changing only the fixed link feeding the last layer changed its responses and which neuron it rewired, exactly as predicted. It's a step toward networks that keep learning through their whole depth on the device. Limits: every layer gets the same label from us, these are local updates rather than backpropagation, and the inputs are synthetic. Each layer learned on every input, but over the run it rewired 37 to 53 of its 64 neurons, not all of them. This does not yet show better performance on a real task.
6
27
383
⭐️ BrainChip - Strategic Research Systems "I've successfully built Secure Akida, a first in the neuromorphic world in terms of known published work, built to be FIPS 140-3 capable..." Secure neuromorphic computing is now available! @BrainChip_inc @computeontology #Akida
⭐️ BrainChip - Secure Akida/RunAkida "BrainChip provides energy-efficient neuromorphic technology. Kevin D. Johnson is building the deployment, security and coordination around it that helps businesses put that technology to work." @BrainChip_inc @computeontology
12
1
44
904
⭐️ BrainChip - Secure Akida/RunAkida "BrainChip provides energy-efficient neuromorphic technology. Kevin D. Johnson is building the deployment, security and coordination around it that helps businesses put that technology to work." @BrainChip_inc @computeontology
11
1
29
655
We had only ever timed BrainChip's AKD1500 through the host: upload an input, run, check for completion, fetch the result. So we loaded a chain of up to 254 steps into the chip's own SRAM, each step's output feeding the next step's input. The host starts the chain once and then only polls for the finish. No PCIe round trip between steps. Time vs chain length gives the time per step on the chip. On counter8, a tiny 3-layer graph that adds 1 to an 8-bit number: 20.8 µs per step. One ordinary host-fed call of the same graph takes 142.5 µs median. The difference is host-side service around the chip, not a clean measure of PCIe alone. 25,830 step invocations across 330 timed trials. Every output byte and all 400 program words matched, and an independent recount from the raw timestamps reproduced the slope. Limits: one small graph; the step includes on-chip memory and descriptor work, not just arithmetic; no energy measured; a tuned CPU still wins end to end on our matched tasks. The film opens on the actual card under the Nano, then replays the recorded trials: chip outputs, stored connections and the host's real completion checks. Unit activity is computed from the stored program, and the demo chains are shown slowed down (labeled).
1
8
38
567
Got BrainChip's AKD1500 to learn a rule change in the middle of one queued run, with the host sending it nothing new. Setup: an 8-stop map stored in the chip's synapses. We preload one queue of 254 steps and release it once. Of those steps, 161 read a prior chip output; the rest use inputs staged before release. Early in the run two routes swap, and the lessons for the new routes are already in the queue as labels we wrote before release. Across 6 test maps: after the lessons, the learning copy follows the new routes, a frozen copy keeps the old ones, and questions about unchanged routes stay right. In one map the first step after the swap, before any lesson, was wrong. Our laptop model predicted that miss too, and it stays in the record. During each released resident queue, the host supplies no new inputs or labels and fetches no intermediate predictions; it waits and checks for completion. About 18 µs per step on the chip for this mixed predict-and-teach workload, vs a 158 µs median when the host feeds each step. That's a per-step service time, not the time of a weight update. 1,330 chip runs. Every output and the final synapse memory matched the laptop prediction bit for bit. Pre-registered, independently recounted. Limits: synthetic task, lessons staged by us, finite queue, no energy measured, not a CPU comparison. An audit found a transport fault that the first attempt's guard missed. That attempt remains unqualified and preserved. This is a new run with a changed memory-read path and corrected guard.
8
31
707
It’s not always pretty, but it is progress Last time, I showed our current model fitting 64 training scenes. Now its successor can use its predictions to choose a command program on unseen scenes in the same two disc simulator. It sees three frames and predicts eight steps ahead for each of four supplied command programs. A separate selector compares those predictions with a goal image, known to be the outcome of one program, and picks the closest in raw pixels. Two separately initialized learners chose correctly on 76.6% and 78.4% of trials on their respective fresh panels. Each panel had 512 scenes and all four goals per scene, for 2,048 trials. Random choice is 25%. On the second panel, the 64 scene model from our last post scored 24.7%; the earlier design trained on the same stream scored 34.6%. The new learner's mean goal position error was 1.25 pixels, versus 4.65 for random choice, averaged over both discs. That measures the true simulated outcome of the chosen program, not the accuracy of the predicted image. Both learners passed the registered tests, independently recounted from the raw forecasts and simulator truth. In a recorded run on a Jetson Orin Nano, the same seed 43 model chose correctly in 26 of 32 fixed scenes, with all 32 choices matching Ascent. The limits matter. This is offline choice among four given programs, not arbitrary goal planning or closed loop control. Both learners saw the same training stream in the same order. The fresh scenes follow the training distribution and fixed physics. We have not demonstrated reliably readable forecasts. The comparison with our last post changes data, training regime and architecture together, so it does not isolate the cause of the gain.
2
3
10
424
1/8 We’re tracing an AKD1500 learning layer: 128 logical neurons, 64 selected inputs each. These clips pair recorded captures with labelled reconstructions. This is our current reading of bounded experiments, open to correction, not a complete hardware model.
1
7
19
431