Kevin D. Johnson | Heterogeneous Compute Ontology

Building systems that discover what they don't know.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

Stats matter. Meaningful signals matter more.
Your storage cluster is chattering constantly. Almost none of what it says is something you need to hear.
I simulated Weebit Nano ReRAM on FPGA and paired it with BrainChip Akida, so nodes speak only when you really need to listen.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

@Horex_13 And, I would add the fuller Peter van der Made design and legacy. Huge untapped potential for the markets and humanity more broadly speaking.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

The heterogeneous compute ontology here is driven by the fact that data centers will be reconfigured in light of sustainable realities, less power, more intelligent AI, less brute force approaches, and a more custom fit with new paradigms at hand to handle the same.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

Karp is right and I would only add, where we are with AI is not where we will be. The frontier labs will fall behind because they’ve adopted the wrong model. People should own their own data, neuromorphic technology with a heterogeneous compute ontology is changing the game.

[Jawwwn @jawwwn_]

Palantir CEO Alex Karp RIPS into Anthropic’s Dario Amodei and OpenAI’s Sam Altman:

“We have people trying to drug addict us to a future they believe they control.”

“I've spent a lot of time with Dario… [Anthropic] wants to tell you—we have to march into a future where we own nothing, where our businesses aren't profitable, where none of us have jobs, and where our adversaries win.”

“Palantir and our allies are guiding into a future where you have growth and jobs.”

Via @CNBC

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

@CKeenan69 Doing what I can, it’s essential that everyone works together on this!

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

@CKeenan69 Paradigm change is hard, it takes time. That said, I think we are seeing a groundswell right now. Things will continue to be more obvious as we go.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

Coming soon, more financial demos with all the usual characters. Stay tuned!

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

@StefanS59896378 @BrainChip_inc Yes, both inside and outside the data center.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

@StefanS59896378 Exactly.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

By the way, the Akida chips were in Pennsylvania, the simulated Weebit Nano ReRAM was in Austin. Don't assume they need to be on the same silicon or even in the same state. The heterogeneous compute ontology makes distance largely irrelevant. Run anywhere. Run everywhere.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

More info on demo #72 - Your storage cluster is talking constantly. Almost none of it is worth hearing.

Every node in a large filesystem, IBM Spectrum Scale or GPFS in our case, produces a firehose: temperature, load, latency, I/O counters, network statistics. Every node, every second, forever. And the overwhelming majority of it says exactly the same thing. Everything is fine.

Today that data is handled one way almost everywhere. Sensors on each node collect metrics, ship them across the network to a collector, which stores and aggregates them, and something central then looks for problems. Underneath it all sits one assumption: to examine a measurement, you first have to move it.

At a thousand nodes you pay for that movement, that storage, and that searching. And nearly all of what you moved said nothing was wrong.

So I tested the assumption. What if the raw signals never left the node?

Here is what I built on a live GPFS cluster.

Each of ten storage nodes watches 35 of its own signals. Those feed a small analog array, the kind of circuit Weebit Nano is building with ReRAM, which spreads them into a much richer pattern using the manufacturing variation between individual devices. The flaw you would normally engineer away is the thing that makes it work.

That pattern goes to a real @BrainChip_inc AKD1000 chip version 1 chip sitting in the node, which learns that node's ordinary behaviour on the chip itself, while it runs.

The result is a model of normal that is 4,157 bytes. The whole eleven node fleet comes to 58 kilobytes. You could email it.

Then we drove heavy I/O at one node at a time and watched what each node's own model said about itself. The loaded node's match to its learned normal fell away, every time. Its nine neighbours did not budge. Thirty trials, twenty nine went the right way, a separation of 11.4 sigma.

One surprise. We ran it twice: the array simulated in full 64 bit floating point, then on real hardware at four bit precision. The four bit version detected better. So the precision a real ReRAM part would impose is not what limits this.

Because every model lives in the shared namespace, a genuinely hard question becomes a directory listing. One node deviating means that node has a problem. All of them deviating in the same window means the fabric does.

And notice what actually crossed the network to make that possible. Not the measurements. The raw signals stayed on the node that made them and were thrown away once they had updated its model. Only the answer moved.

The cost per node is fixed. Ten nodes, 59 KB. A thousand nodes, about 5 MB. Raw telemetry grows forever. A model does not.

Storage systems have always been the place where data sits and waits for someone else to ask it something. There is no good reason they cannot also be the place where data stays where it needs to stay and does not move unless it really needs to.

Weebit Nano ReRAM's design and BrainChip Akida silicon can make that happen.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

https://t.co/QYBDbrCaC6

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

Stats matter. Meaningful signals matter more.
Your storage cluster is chattering constantly. Almost none of what it says is something you need to hear.
I simulated Weebit Nano ReRAM on FPGA and paired it with BrainChip Akida, so nodes speak only when you really need to listen.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

More info on demo #72 - Your storage cluster is talking constantly. Almost none of it is worth hearing.

Every node in a large filesystem, IBM Spectrum Scale or GPFS in our case, produces a firehose: temperature, load, latency, I/O counters, network statistics. Every node, every second, forever. And the overwhelming majority of it says exactly the same thing. Everything is fine.

Today that data is handled one way almost everywhere. Sensors on each node collect metrics, ship them across the network to a collector, which stores and aggregates them, and something central then looks for problems. Underneath it all sits one assumption: to examine a measurement, you first have to move it.

At a thousand nodes you pay for that movement, that storage, and that searching. And nearly all of what you moved said nothing was wrong.

So I tested the assumption. What if the raw signals never left the node?

Here is what I built on a live GPFS cluster.

Each of ten storage nodes watches 35 of its own signals. Those feed a small analog array, the kind of circuit Weebit Nano is building with ReRAM, which spreads them into a much richer pattern using the manufacturing variation between individual devices. The flaw you would normally engineer away is the thing that makes it work.

That pattern goes to a real @BrainChip_inc AKD1000 chip version 1 chip sitting in the node, which learns that node's ordinary behaviour on the chip itself, while it runs.

The result is a model of normal that is 4,157 bytes. The whole eleven node fleet comes to 58 kilobytes. You could email it.

Then we drove heavy I/O at one node at a time and watched what each node's own model said about itself. The loaded node's match to its learned normal fell away, every time. Its nine neighbours did not budge. Thirty trials, twenty nine went the right way, a separation of 11.4 sigma.

One surprise. We ran it twice: the array simulated in full 64 bit floating point, then on real hardware at four bit precision. The four bit version detected better. So the precision a real ReRAM part would impose is not what limits this.

Because every model lives in the shared namespace, a genuinely hard question becomes a directory listing. One node deviating means that node has a problem. All of them deviating in the same window means the fabric does.

And notice what actually crossed the network to make that possible. Not the measurements. The raw signals stayed on the node that made them and were thrown away once they had updated its model. Only the answer moved.

The cost per node is fixed. Ten nodes, 59 KB. A thousand nodes, about 5 MB. Raw telemetry grows forever. A model does not.

Storage systems have always been the place where data sits and waits for someone else to ask it something. There is no good reason they cannot also be the place where data stays where it needs to stay and does not move unless it really needs to.

Weebit Nano ReRAM's design and BrainChip Akida silicon can make that happen.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

Stats matter. Meaningful signals matter more.
Your storage cluster is chattering constantly. Almost none of what it says is something you need to hear.
I simulated Weebit Nano ReRAM on FPGA and paired it with BrainChip Akida, so nodes speak only when you really need to listen.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

Talking to the FPGA again today...nodes chattering in PA to the FPGA here in Austin.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

@Alacritic_Super I don't have any code for you, but BrainChip has an SDK and sample models on their developer website. https://t.co/bBQSbYtKOo

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

Remember this? I think it's high time to produce another track... Music via @BrainChip_inc Akida
https://t.co/fe2PYs7n2f

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

@geerlingguy Demo #71, also found here: https://t.co/dghcNW9Yd1

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

I'm dedicating this build to @geerlingguy because I love watching his recent videos on time.

What I built and measured. 13 hours of real GPS and crystal recording, locked to 12 satellites. The ReRAM array on a Tang Primer 20K FPGA, from published Weebit Nano measurements. The readout on a real AKD1000, hardware only, bit identical to v1 simulation, 800 bytes on chip.

A seamless takeover, caught about 1 hour 54 minutes after the drag began, at one false alarm per day. The GPS receiver never detected it at all (simulated of course, because broadcasting fake signals is a crime).

The wider point here is the heterogeneous compute ontology at work again, because it treats Akida as the peer it is to CPU/GPU/QPU, and in this case, Weebit Nano ReRAM.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

I built this demo by modeling Weebit Nano ReRAM together with BrainChip's Akida. The truth only lasts so long when you're spoofing GPS signals. With this system, sooner or later you're going to be found out.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

I built this demo by modeling Weebit Nano ReRAM together with BrainChip's Akida. The truth only lasts so long when you're spoofing GPS signals. With this system, sooner or later you're going to be found out.

[Kevin D. Johnson | Heterogeneous Compute Ontology @computeontology]

Demo incoming today. By the way, the market often uses terms like accelerator imprecisely. BrainChip's Akida is not an accelerator for the CPU, it's a neuromorphic chip in a class all its own, a peer in the heterogeneous compute ontology to CPU/GPU/QPU and more.

@BrainChip_inc