In 2009, researchers built a chip with the equivalent of 200,000 neurons and 50 million synapses in an effort to mimic a brain in silicon. The work came from FACETS (Fast Analog Computing with Emergent Transient States), a European project coordinated by physicist Karlheinz Meier at Heidelberg University in Germany, and was reported by MIT Technology Review in "Building a Brain on a Silicon Chip" (March 25, 2009). The chip holds only a small fraction of the neurons and connections in a real brain, but its design was meant to scale up.
What the team built
Most brain research on computers simulates neurons in software: a program steps through equations that describe how each neuron's voltage changes over time. The FACETS team took a different route. They built the neurons and synapses as physical circuits made of transistors and capacitors, so the electronics behave like the biology instead of calculating it. Meier summed up the difference in the Technology Review piece: rather than simulating neurons, the group was building them.
The circuits sit on a standard eight-inch silicon wafer, and the wafer is used whole rather than being cut into separate chips. According to the article, a neuron circuit needs about 100 components and a synapse about 20. Because there are far more synapses than neurons, the synapses take up most of the area on the wafer.
Why build a brain in analog silicon?
A physical circuit can run much faster than the biology it copies. Real neurons work on time scales of milliseconds. Small on-chip capacitors and transistors charge and discharge far faster, so the hardware model runs in accelerated time. The Technology Review article reported that the prototype ran about 100,000 times faster than real time; Meier described it as simulating a day in a second. That speed lets researchers watch learning that would take days or weeks of biological time in a few seconds.
There is also a power argument. A software simulation spends energy moving numbers in and out of memory for every neuron at every time step. An analog circuit simply lets currents flow, so the cost per synaptic event can be much lower. The article did not give power figures for the FACETS chip, so this post does not quote one.
How it learns
The synapse circuits implement spike-timing dependent plasticity (STDP), a learning rule credited in the article to Henry Markram. The idea is simple: if one neuron fires just before the neuron it connects to, the connection gets stronger; if it fires just after, the connection gets weaker. Repeated over many spikes, the network strengthens the paths that carry useful cause-and-effect patterns. Markram also cautioned in the article that the chip is not a brain, and is better seen as a platform for artificial intelligence research than a way to understand biology.
Where it fits
This kind of hardware is now called neuromorphic computing: chips whose structure copies networks of spiking neurons rather than a conventional processor and memory. The FACETS wafer is an early example of the analog, accelerated branch of the field. Other groups have built digital neuromorphic chips that run in real time. For an engineer, the interesting points are the wafer-scale integration, the mixed-signal circuit design, and the way area is dominated by the synapse array, much as memory arrays dominate many digital chips.
Deep dive: the numbers and details
This part holds the figures and a few worked estimates. Skip it if the overview is all you need.
The reported figures
| Item | Value reported in 2009 |
|---|---|
| Neurons on the wafer | 200,000 |
| Synapses on the wafer | 50 million |
| Earlier prototype | 384 neurons, 100,000 synapses |
| Components per neuron circuit | about 100 |
| Components per synapse circuit | about 20 |
| Wafer size | eight inch (200 mm) |
| Speed versus biology | about 100,000 times faster ("a day in a second") |
| Project scale | 15 groups in seven countries, budget of 10.5 million euros |
| Planned "superchip" | one billion neurons and 10^13 (ten trillion) synapses from many connected wafers |
All values are from the Technology Review article. The synapse count for the superchip appears there as "1013", which reads as 10 to the 13th power with the superscript lost.
Why synapses dominate the area
A quick count shows why the article says synapses take most of the wafer:
- Neurons: 200,000 x 100 components = 20 million components.
- Synapses: 50,000,000 x 20 components = 1 billion components.
- Share taken by synapses: 1,000 / (1,000 + 20) = about 98 percent of all components.
The design has 250 synapses per neuron (50 million / 200,000). Any attempt to raise connectivity per neuron makes the synapse array grow even faster, so the size and accuracy of one synapse circuit decides how far the design can scale.
What 100,000 times faster means
One day is 86,400 seconds, so "a day in a second" is an acceleration of about 86,400, consistent with the round figure of 100,000. A biological process that takes one hour (3,600 s) would run in about 36 ms on the hardware at a factor of 100,000. A learning experiment of 1,000 trials of one second each would take 1,000 s of biological time but about 10 ms on the wafer. The catch is that the input and output systems must handle spikes at that accelerated rate, which makes communication bandwidth, not computation, the hard part of such systems.
The STDP rule in numbers
A common textbook form of STDP changes the weight w of a synapse by an amount that depends on the time difference dt = t_post - t_pre between the postsynaptic and presynaptic spikes:
if dt > 0: dw = +A_plus * exp(-dt / tau_plus) (pre before post: strengthen)
if dt < 0: dw = -A_minus * exp( dt / tau_minus) (post before pre: weaken)
With a time constant of 20 ms in biological time, a pair 10 ms apart gives exp(-0.5), about 0.61 of the maximum change. On hardware accelerated 100,000 times, the same 20 ms window becomes 200 ns, which is why the circuits must measure spike timing on a sub-microsecond scale. This form of the rule is general background, not a description of the exact FACETS circuit.
What happened next
- The Heidelberg work continued as the BrainScaleS systems. A 2020 paper from the Heidelberg group, "Accelerated Analog Neuromorphic Computing" (Schemmel et al.), describes the first generation, BrainScaleS-1, as a wafer of about 500 interconnected chips built in a 180 nm process, and its successor BrainScaleS-2 in 65 nm, which adds a digital plasticity processor so that learning rules can be programmed.
- In 2017, a team including Karlheinz Meier trained a deep spiking network on the BrainScaleS wafer-scale system, with an acceleration factor of 10,000 relative to biological time, as reported in "Neuromorphic Hardware In The Loop" (arXiv, 2017).
- BrainScaleS-2 was developed within the neuromorphic computing activity of the European Human Brain Project, which ran from October 1, 2013 to September 30, 2023.
Further reading
- MIT Technology Review: Building a Brain on a Silicon Chip (March 25, 2009)
- Schemmel et al.: Accelerated Analog Neuromorphic Computing (arXiv, 2020)
- Neuromorphic computing: history (Wikipedia)


Do we want such a brain to scale up as much as our brain. It will only evolve in causing credit crisis, foreclosures and loss of jobs.
ReplyDeleteMaybe not being so smart may be to it’s advantage.