Several years ago, at the beginning of this decade, researchers started to look into computer architectures that would mimic the neuronal networks we find in our brain, in particular networks where storage and computation was one and the same. DARPA launched the project SyNAPSE to foster the creation of neuromorphic chips (that’s how they are called) and IBM delivered TrueNorth in 2016, a neuromorphic chip with an equivalent 1,000,000 neurones and 256 million synapses. I consumed 70mW and could perform 46 billion synaptic like operation per second.
Now Intel has announced the creation of a system delivering 8.3 million neurones equivalent by scaling up its Loihi neuromorphic chip to 64 processors. The new configuration, called Pohoiki Beach increases by 1,000 times the performances of a current CPU with a 10,000 times improvement in energy efficiency.
Each Loihi processor contains 128 neuromorphic cores, 3 Lakefield CPU cores and an off-chic network. 8 to 32 Loihi chips can be aggregated onto a Nahuku board and several of these were joined into a single system, the announced Pohoiki Beach. So if you do the math, the Pohoiki Beach contains 8,192 neuromorphic cores. Each Loihi consists of 2 billion transistors fabricated in the 14nm technology. It has an equivalent of 131,072 neurones and 130 million synapses. That means Pohoiki Beach reaches 8,388,608 neurones and 8.384 billion synapses.
A more powerful version, code named Pohoiki Springs is in the making and will contain 100 million neurones equivalent and almost 100 billion synaptic junctions. This version will be 1/1000 the size of our brain (100 billion neurones and 100 trillion synapses). So, in terms of hardware we are not there, yet.
The big question mark is if hardware is what it takes to have intelligence… For sure without any hardware you cannot have any software, hence no intelligence would be possible. Someone is convinced that it is all about software (i.e. it is a matter of understanding the processes giving rise to intelligence and once you get the understanding you can replicate them in software). Others are saying that software is not enough, you need a specific hardware to support a software that could really deliver intelligence. The jury is still out. In the meantime, research is making progress on both hardware and software….