Swiss biocomputing startup FinalSpark recently launched the world's first online biological computing platform "Neuroplatform" based on in vitro biological neurons, capable of learning and processing information with power consumption reduced by 1 million times compared to traditional digital processors.
Lower-energy Biological Computing Technologies
FinalSpark has stated that in today's society, the significant energy costs required to process digital information have become one of the important prices of modern scientific and technological progress.
Especially with the rise of Generative Artificial Intelligence (AI), the trained AI large model parameters have also become increasingly large, with more and more users using large models, the need for silicon-based AI chips is also increasing, leading to a growing energy consumption.
For instance, training a single large language model like GPT-3 requires about 10 GWh of energy, which is roughly 6,000 times the energy used by a European citizen in a year.
In addition to the substantial energy requirements associated with training LLMs, the cost of inference is also an equally pressing issue.
Data recently disclosed by OpenAI CEO Sam Altman shows that platforms like OpenAI generate over 100 billion words daily through services like ChatGPT, resulting in a considerable energy consumption.
Preliminary calculations using the LLaMA 65B model as a reference point indicate that the energy consumption solely for word generation ranges from 45 to 600 billion joules per day.
While providing AI-driven insights and interactions for millions of users worldwide is essential, the high energy consumption underscores the urgency for a more energy-efficient computing paradigm.
According to a report by market research firm Factorial Funds, OpenAI's text-generating video model Sora uses between 4,200 and 10,500 pieces of H100 GPUs in a month, with a single H100 capable of generating a one-minute video in about 12 minutes or approximately 5 one-minute videos per hour. During peak times, Sora requires 720,000 H100 GPUs.
With a peak power consumption of 700 watts per H100 GPU, factoring in the power requirements of the GPUs themselves, servers, other components, and cooling, an AI data center with 720,000 H100 GPUs would consume over 7.2 billion watt-hours per hour.
A recent research report by Goldman Sachs also indicates that with the rapid advancement of artificial intelligence, the demand for computing power is increasing. It is projected that by 2030, global AI data centers will see a 160% increase in their electricity demand.
FinalSpark points out that the development of artificial intelligence should not be restricted solely due to its high energy consumption by implementing regulations to limit the development and use of existing AI models, as losing the lead in technological advancement could incur significant strategic costs.
Therefore, alongside increasing the supply of green energy, adopting more energy-efficient unconventional computing technologies is crucial as the best way to reduce carbon emissions.
Biological computing is one such more energy-efficient unconventional computing technology, involving building computers using living neurons. One of its major advantages is that the energy required for neuron computation is significantly lower than that of digital computers.
It is estimated that living neurons consume over 1 million times less energy than the digital processors we currently use. For example, the human brain, with approximately 86 billion neurons, consumes only around 20 watts.
FinalSpark suggests that given the significant power consumption issues faced by existing artificial neural networks (ANNs) running on silicon-based chips, the emerging biological neural network (BNN)-based biological computers hold promising prospects.
FinalSpark's biological computing platform, "Neuroplatform," aims to leverage the plasticity of living neurons triggered by electrical stimulation to quantify, store, and process information using the natural abilities of living neurons.
The goal of FinalSpark is to modify living neurons in a predictable and controllable manner through electrical stimulation, allowing for computation.
"We believe this is the future of the emerging field of biological computing, where biological elements are used as hardware. After all, what biological material is better suited for computation than living neurons?" FinalSpark wrote in its official blog post.
To study biological computing technology, FinalSpark's laboratory has used human neurons reprogrammed from adult somatic cells (such as skin cells) into induced pluripotent stem cells (iPSCs) to construct a neural sphere, which is a living "brain-like organ" (FO) consisting of around 10,000 active neurons, with a diameter of about 0.5 millimeters.
Typically, such neural spheres are used for biomedical research, studying brain diseases, and gaining a better understanding of how the human brain functions.
However, FinalSpark is pioneering the use of these spheres for biological computing, aiming to build a new type of computer processor.
The online biological computing platform "Neuroplatform" introduced by FinalSpark consists of 16 of the aforementioned "brain-like organs," with every 4 "brain-like organs" connected using a multi-electrode array (MEA) to form a 3D cell cluster resembling brain tissue.
In other words, each MEA can accommodate four "brain-like organs," totaling 8 electrode connections (as shown below), utilizing a digital-to-analog converter (DAC) to send electrical signals and collecting signals from neurons through an analog-to-digital converter (ADC).
The MEA device utilizes the air-liquid interface (ALI) method, where the brain-like organoids are directly placed on electrodes located above a transparent membrane, with the culture medium flowing beneath the membrane in a 170 μL chamber.
A thin layer of culture medium, formed by surface tension, separates the upper side of the organoids from the humidified air in the culture chamber. The cover partially covering the MEA further protects this arrangement (as shown in the diagram).
Compared to immersion culture methods, this ALI approach achieves higher throughput and stability as it eliminates the need for specialized coatings, and the brain-like organoids are less likely to detach from the electrodes.
The electrodes in this bio-computing system can both stimulate (input) and record (output).
The corresponding analog-to-digital and digital-to-analog conversions are performed by the Intan RHS 32 probe.
Stimulation is carried out using a current controller with a range from 10 nA to 2.5 mA, while recording is achieved by measuring the voltage on each electrode, with a sampling frequency of 30 kHz, a resolution of 16 bits, and an accuracy of 0.15 μV.
The probe is connected to the Intan RHS controller, which is then linked to the computer via a USB port.
The electrical activity of each of the 32 electrodes is measured in microvolts per second. Each group of 8 electrodes records activity from different brain-like assemblies.
The graph displays the electrical activity recorded by each of the 32 electrodes. It is evident that the activity recorded by each electrode is distinct.
Since each group of 8 electrodes records activity from different brain-like organs, and for a given organ, each electrode records from a different location.
The data being displayed is continuously updated online, and researchers can view it around the clock via the FinalSpark website (https://finalspark.com/live/).
FinalSpark states, "We compared the recording characteristics of this ALI device with monitoring of submerged brain-like organs using the MCS MEA (60MEA200/30iR-Ti), employing the same Intan system for voltage conversion."
The following graph illustrates superimposed action potentials recorded using the ALI and submerged devices, demonstrating similar signal characteristics.
The Neuroplatform system relies on a laptop for control, providing access to three resources:
A database storing all information about the Neuroplatform system;
Intan software running on a dedicated PC to record the number of detected spikes within a 200-millisecond window and set stimulation parameters;
A Raspberry Pi development board triggering current stimulation based on the stimulation parameters.
The lifespan of the brain-like organoids has exceeded 100 days.
To sustain the life of the brain-like organoids, they need to be kept in a sterile environment at around 37°C and continuously supplied with neural medium (NM).
To address this, FinalSpark has designed a closed-loop microfluidic system that enables continuous medium supply, reducing physical interventions in the incubator and ensuring stable environmental conditions.
Reportedly, the medium circulates at a rate of 15uL/min, with the medium flow controlled by the BT-100 2J peristaltic pump and adjusted continuously as needed (e.g., during experimental runs). The peristaltic pump is connected to the PC control software via an RS485 interface for programming (e.g., Python) or manual operation.
The microfluidic circuit system is made of 0.8 millimeter (inner diameter, ID) tubing. Continuous monitoring of the microfluidic circuit and flow rates is achieved using Fluigent flow sensors, which are connected via USB to the Neuroplatform control center. Data related to the flow rates are stored in a database for future access.
Additionally, each MEA is equipped with a 12.3-megapixel camera that can be controlled interactively or programmatically (e.g., via Raspberry Pi) to capture static images or record videos. This allows for the identification of issues such as cell death, organoid displacement possibly caused by microfluidics, changes in medium acidity (using color analysis as our medium contains phenol red), contamination, generation of neural melanin (potentially occurring during dopamine release), overflow (medium inadvertently filling the chamber above the membrane), or bubbles in the medium.
Regarding the lifespan of organoids, FinalSpark indicates that initially their lifespan was only a few hours, but through various improvements, especially those related to microfluidic systems, their lifespan has been extended to over 100 days in the best cases.
32 research groups have already applied for collaboration.
In order to facilitate collaborative research on bio-computing based on artificial biological neural networks (BNN) and to develop novel approaches for neural networks using biological neurons, a system capable of conducting numerous experiments is required. Therefore, FinalSpark has developed the Neuroplatform system, supporting 24/7 electrical stimulation and action potential monitoring, allowing researchers worldwide to conduct electrophysiological experiments on an unprecedented scale.
FinalSpark states, "Building the next generation of bio-processors using live neurons is no easy task. Despite many advantages such as energy efficiency, scalability, and proven information processing capabilities, bio-processors from live neurons are challenging to develop. We still do not know how to program them. Unlike digital computers, bio-computers are true black boxes. For this reason, we need a large number of experiments to make them functional. However, if we find a way to control these black boxes, they can become truly powerful IT tools."
Over the past three years, the Neuroplatform system has collected data from over 1,000 organoids, accumulating more than 18 TB of data.
As of 2024, the system is now open for research purposes.
Despite 32 research groups requesting access to the Neuroplatform, the current infrastructure can only accommodate 7 research groups due to their own research needs. Therefore, FinalSpark is expanding the scale of the AC/DC hardware system to support more users simultaneously.
Fred Jordan, co-founder of FinalSpark, said, "We firmly believe that only through international cooperation can such an ambitious goal be achieved."
It should be noted that FinalSpark is currently limited to running closed-loop algorithms for neural plasticity on a single organoid, as these algorithms require sending real-time adaptive analog signals to each organoid.
To address this, FinalSpark's software is also being updated to support parallel closed-loop operations on up to 32 organoids in the future.
