Whilst this is from Sep 23, I don't recall seeing or reading it.
Worth a read through imo.
Comprehending a revolutionary concept is not easy. Too often we get intimidated by jargon or struggle to see the core concept. That happened to me when I became curious about BrainChip (NASDAQ: BRCHF, ASX: BRN).
disruptiveinvestmentideas.substack.com
BrainChip: The Cloud-Free Future is Here
28 SEPT 2023
Introduction
Comprehending a revolutionary concept is not easy. Too often we get intimidated by jargon, struggle to see the core concept, or simply pretend that we are too busy.
That happened to me when I became curious about BrainChip (NASDAQ: BRCHF, ASX: BRN). It doesn’t help that Elon Musk’s Neuralink is inserting chips into the brains of monkeys or pigs. Such images come to mind when we hear of
BrainChip. But no, here you won’t learn about such brain-computer interfaces. BrainChip is a company that commercialized a new semiconductor architecture. So far, this might cause some readers to yawn, stop, and search for other distractions. However, I will explain to you in plain language why this matters for you, as a long-term investor.
Let's start with something you have experienced. Recall the last time you purchased a computer. You were asked to choose between different processor brands and versions, then there was the decision to select a graphic card and finally the question on RAM. All these are chips; semiconductors, right? Apart from the sleek marketing message and promises toward the capabilities that come with selecting one over the other, did you really care about comprehending what was in front of you?
What Is the Problem that BrainChip Has Solved?
In its most simple version, here is the elevator pitch: BrainChip invented and patented a semiconductor architecture that lowers their energy consumption by magnitudes.
At this stage, I will not try to cover all the disclaimers and conditions that will come naturally with such a revolutionary statement. Just let it sink in. Does it matter to me if the energy consumption of my PC gets lowered significantly? Probably not. But when we leave our home turf and think of other concepts like the cloud or data centers, we arrive where the action is. Did you know that these boiler rooms of the Internet consume more than 1% of all energy worldwide1? And finally mentioning the elephant in the room, AI: it costs a company about 100x more to answer a question you type into ChatGPT or Bard in comparison to a Google search.
Now, lifting off and getting the helicopter view: there are huge efficiency and performance gains in case we could find a way to realize what BrainChip’s new chip architecture is promising.
The Details
And here’s where most will get lost: technical jargon. One needs to expand their vocabulary to understand where the investment thesis is. These are concepts not visible or under-appreciated in our daily lives. So, let's dive in!
This above-mentioned boiler room of the internet is not running on steam anymore. We are electrified. Electricity is needed to allow semiconductors to come up with endless sequences of zeros and ones. The plumbing, to stay with the analogy, is what makes electrons take different pathways through the circuitry.
Chip giants like Intel, AMD, and NVIDIA have optimized their designs to achieve a truly astonishing number of computations. Progress has been relentless and it seems no law of physics can stop them. The most advanced chips used in today's desktops or data centers have one thing in common: they must conduct an incomprehensible number of calculations. Electrons race through their circuitry and generate so much heat that data centers spend more on air-conditioning than on the actual semiconductor hardware.1
Here you might say that over the last years, your phone has seldom grown warm and can often do similar calculations on a smaller scale. This is the achievement of a small UK company, now famous and known on the NASDAQ by the ticker ARM. More than 30 years ago, they came up with a novel energy-efficient chip architecture. Today, virtually all phones use this arm design.2
Vocabulary
At this stage, let's get accustomed to a few technical terms required to understand the investment case I am preparing for BrainChip:
- CPU:
Central
Processing
Unit, used for most day-to-day calculations
- GPU:
Graphic
Processing
Unit, used to render graphics and A.I.-type calculations
- Instructions Set: Commands (vocabulary) the processor understands
- Architecture: Circuitry depending on the complexity of this set of instructions
- X86: Complex instruction set, practically unchanged since the 1970s
- Sequential Processing: A CPU starts and finishes a calculation in sequential order
- Parallel Processing: Runs calculations in parallel
- Core: A unit that conducts the individual calculation
Transition
Why is this transition still ongoing, considering the obvious advantages coming with a more energy-efficient architecture? Processors require a set of instructions to do their calculations. Software and hardware need to speak the same language to execute these instructions. Based on a specific task, like running a spreadsheet (CPU) or generating a 3D animation (GPU), different chips and sets of instructions give the best results.
We will always remain in a certain flux: one technology finds more adoptions, and the other gets scaled down. Initially, we fine-tuned the X86 architecture and added more and more cores. This kept up with the demands to a degree. Data centers expanded with parallel computing designs. This became more and more unsustainable and expensive. With the widespread introduction of generative AI (text, images, music, and code), we are experiencing a watershed moment right now.
For home computers, Apple is leading the field to bring arm architecture into our homes. Data centers can’t change their hardware as nimbly as we consumers can. They are stuck for a while with an expensive set of legacy X86-style hardware for the near future.
Data
Don’t you sometimes wonder where all this data resides? Sure, we can see a memory stick or a hard drive. But the bulk of the world's data resides in the cloud, aka data centers.
The amount of data generated by humans writing a message, filling out a form or saving a picture is less than the data volume generated when machines have exchanges with each other. The gap is growing exponentially.3 You might wonder why machines decide to generate data. No, they are not sentient – not so far at least. This data volume is generated by sensors, as well as from simulations, machine learning, and blockchain. Obvious sensors like temperature probes or traffic cameras might come to mind. But we’re getting side-tracked by attempting to understand each of these autonomous data sources: our world is awash in non-human-generated synthetic data. All this data is backed up to central servers where programs run operations to make sense of it.
The Edge
This non-human data is generated to a large degree “on the edge”. Devices/sensors generate data, flowing to the cloud for interpretation.4 This is what is causing the increase in data, inundating our internet with traffic and increasing the size of data centers.
View attachment 58660
BrainChip
Here comes our company. A micro-cap. A company with just a hundred employees and no turnover to speak of. Investor circles call these companies “story stocks.” As they have no turnover to show for it, they convince via their story. That should give the context. Nothing is certain in this domain. Risk is abundant. Success can take a generation.
Arm proved its design & functionality 20 – 30 years ago without instant success. It required the widespread adoption of mobile phone-computer hybrids, commonly referred to now just as phones. I want to provide the reasoning why this will repeat with BrainChip: we are currently experiencing an equivalent catalytic event, as evidenced by the exponential generation of non-human data volumes.
Their Secret Sauce
Understanding how this start-up-like company can find a solution to the global data dilemma requires additional vocabulary.
- Neuromorphic: Brain-like semiconductor architecture
- IoT:
Internet
of
Things, the non-human data tsunami minions
- Event-Based: Running a calculation when needed, when an event occurs
- Latency: The time it takes between exchanges
- Neural Network: A machine-learning model simulating how our brain works
- Spike: Data burst that occurs when an event is recorded
- Convolution: mathematical operation to extract features from images or signals
You could say that they patented the operation model of our brain. This is a lot to take in. We are talking about a set of instructions. Brainchip is not selling a synthetic brain, with neurons and synapses. Their founder was an early proponent of what is now called neuromorphic engineering. The concept was too abstract for many years to warrant commercial attention.
Intriguing? Here is the catch: at this stage, most potential investors will get lost further researching what has been achieved. To a large degree, this is the reason why the community surrounding BrainChip is regarded with suspicion. It is simply difficult to wrap your head/brain around this concept.
In Simple Terms
Our brain operates on a meager 20 watts, a light bulb's worth of energy consumption. This is the case when playing chess or daydreaming. Evolutionarily, we have achieved something that has not been replicated by any commercial chip architecture. It’s because our brain works in an event-based manner. Example needed?
Imagine looking at a blank sheet of A4 paper that has a dot in its center. Humans have no difficulty identifying this dot. In comparison, a camera combined with image analysis will analyze each pixel, line by line, to determine that a certain location has a higher density of contrast. Calculations will then determine the event (localization of the dot). The software can’t differentiate between the data (the white part of the sheet of paper) and the significance of the dot (the black part). It is just stoically analyzing the data from top to bottom.
The neurons of our brain will only pass on information when an event occurs. This way, it can remain in a certain way dormant and only consume energy during an event!
This concept has fascinated scientists and culminated in a neuromorphic design logic. The founder of BrainChip noticed the commercial value and patented these advances. They run what’s called a
spiking
neural
network (SNN) on their chip.
Why Now?
Investors wait until there is a need to fund new innovations. This inflection point is now. We are surrounded by sensors in cars, outdoors, homes, and a multitude of smart devices. Examples needed?
- Driver-Alertness: Detects if a driver is losing attention
- Crowd Management: Build-up of city traffic, or crowds during an event
- Biometric Recognition: Border control and traveling
- Alexa/Siri: Low latency keyword detection to complex questions
- Hearing Aids: Discern and amplify sounds selectively to understand what was said
- Vital Signs: Wearables support monitoring and preventive medicine
- Industrial Predictive Maintenance: Alerts get sent before equipment or infrastructure breaks down
These advances have already been implemented. And we will soon have fully autonomous, self-driving cars/taxis. The amount of additional data that will be transferred with these applications will inundate the internet. Unless we can stop these sensors from communicating their data to the cloud.
Neural Networks
This data (consisting of video/images, sound, and other measurable sensor results) needs to be classified and converted into a neural network. Currently, these data streams are analyzed in a process called convolution, resulting in a
Convolutional
Neural
Network (CNN). This is achieved in a central location utilizing top-of-the-range GPUs. The process can take a year, cost millions, and is reliant on high-quality human-screened data. Once completed the neural network will be installed on the final device (car, hearing aid or sensor). But things stop here. This neural network is a one-trick pony. It can’t learn from its observations. In case something changes, all needs to be recalibrated at HQ.
BrainChip generates as well a neural network, but it's a
Spiking
Neural
Network (SNN). It can be quickly trained on a much smaller number of lower quality, non-human validated samples. On top of it, once the SNN has been established, it keeps learning and continuously updates its model. How is this possible - SNNs are not living, right? Apart from CNNs, BrainChips SNN model parameters (called weights & biases) are not fixed. These values get changed on the chip's memory when the SNN “learns.”
How Can Brainchip Make Money?
They are the only company with a commercial neuromorphic chip architecture and corresponding patent-protected intellectual property (IP). Like arm’s revenue model, BrainChip sells its IP to anyone in the business of chip design. They are already a:
- member of arm’s A.I. partnership program
- partner with Intel’s foundry services
- selected company by PROPHESEE, the global leader in IoT machine vision
- selected company by SiFive: open-source AI chip design
Additionally, they work with enablement partners to provide a vertically complete solution to simplify evaluation and implementation. Completing the offering are integration partners offering ready-to-use
system-on-chip (SoC) products.
As their business is essentially licensing software, they have very high margins. Currently, arm is valued at about 50 billion dollars. I see it entirely feasible that BrainChip will reach a comparable valuation when its design architecture gets adopted. They are currently valued at about 200 million dollars. Where else can you get such a 250x potential?
View attachment 58661
The Solution
With BrainChip’s processor, called Akida, all computations are performed on the chip. It requires no internet connection at all.
- It operates at a fraction of the 20 Watts that our brain requires. So, no cooling or main power sources are required; a simple battery will do!
- All data collected remains where it was observed/generated. No more privacy worries about hacked cloud servers. The data does not leave the sensor chip.
In Summary
Data processing will remain in constant flux and hardware updates are costly. Even so, we presently experience the confluence of multiple evolutions:
- X86 architecture will be phased out as arm’s processors have proven faster and more energy efficient. Soon, every computer will run on less complex arm-like instruction architecture
- Chronological data from IoT devices, demand forecasting or image processing will be enabled by BrainChip’s Akida in real-time and cloud-independent
- Governments won’t allow Internet giants to amass data or cross-border transfer of data to central server farms. Regulation will benefit technologies that can function without the need for data transfers
www-hackster-io.cdn.ampproject.org
circle8.io
![Professor Hoejun Yoo of the Department of Electrical and Electronic Engineering at KAIST is explaining complementary-transformer technology at the Sejong Government Complex on the 6th. [ⓒ Digital Daily]](/proxy.php?image=https%3A%2F%2Fwww.ddaily.co.kr%2Fphotos%2F2024%2F03%2F06%2F2024030613365326846_l.jpg&hash=8d6cd00f83d601d9f1573589a18b85b9 "Professor Hoejun Yoo of the Department of Electrical and Electronic Engineering at KAIST is explaining complementary-transformer technology at the Sejong Government Complex on the 6th. [ⓒ Digital Daily]")
www.linkedin.com
