A simulation that runs faster on a commercial graphics card than on some supercomputers could drastically cut the cost of studying how our brains work.
Researchers have long used digital models to better understand our brains in the hope of developing cures for diseases such as Alzheimer鈥檚 or Parkinson鈥檚, but simulating the number of neurons and synapses of even the simplest creature is enormously computationally intensive, meaning even supercomputers struggle.
Before running a simulation of the brain鈥檚 neurons and vast number of synaptic connections the model must be transferred into the computer鈥檚 working memory, complete with the starting state of every synapse. As the simulation progresses, the computer must keep referring to this set of data to retrieve or update the state of each synaptic connection. This memory access is far slower than the actual calculations involved in the simulation, so it becomes a bottleneck.
Advertisement
Commercial graphics cards, known as GPUs, are designed to render 3D scenes by rapidly carrying out many arithmetic calculations in parallel, an ability that also makes them particularly speedy at other tasks, including simulating synaptic connections.
James Knight at the University of Sussex, UK, and his colleagues created a simulation that uses a random number generator as part of the process of creating a synaptic state. Although this random element means the simulation can鈥檛 refer to the exact starting state of the model each time it needs to create a new connection, the team found that this approach produced results comparable to traditional simulations.
It also makes things much faster, as the computer only needs to handle data about the state of the synapses that it is currently modelling, just as a video game only renders the part of the world the player is looking at.
The team used an existing model of a macaque monkey鈥檚 visual cortex, consisting of more than 4 million neurons, as a benchmark. In 2018, one second of brain activity inside the model was simulated on an IBM Blue Gene/Q supercomputer in 12 minutes. Using a commercially available graphics card, Knight鈥檚 team was able to carry out the same task in just under 8 minutes.
The IBM supercomputer was already several years old when that experiment was conducted. A newer JURECA supercomputer has been able to run the same simulation in just 31 seconds, but these can cost tens of millions of pounds and require a team of staff to maintain. By contrast, Knight says the Nvidia Titan RTX hardware used in his tests costs just a few thousand pounds.
鈥淭his potentially means that researchers whose primary focus isn鈥檛 dealing with supercomputers could explore things with this model,鈥� he says.
But there is a flaw. When we learn, our brains are constantly weakening or strengthening the connections between synapses, an ability known as synaptic plasticity. The GPU simulation can鈥檛 do this, because it always has to recalculate the connections from scratch, reverting back to the model鈥檚 original state.
Knight believes a hybrid approach using his new technique and a traditional model where the state of synapses is stored in memory and can be updated would allow plasticity where needed and high speed where it isn鈥檛, but the team has yet to try this.
鈥淭here is a massive benefit to be gained from working out tricks that let us scale up neural simulations on GPUs, given their wide availability,鈥� says Simon Schultz at Imperial College London. 鈥淏ut getting synaptic plasticity working efficiently is the bane of scaling up neural simulations, and a problem which this algorithm unfortunately doesn鈥檛 solve.鈥�
Nature Computational Science
