Computers

Intel's new neuron-based computer matches brain of a small mammal

Intel's new neuron-based compu...
An array of Intel Loihi neuromorphic research chips.
An array of Intel Loihi neuromorphic research chips.
View 4 Images
An array of Intel Loihi neuromorphic research chips.
1/4
An array of Intel Loihi neuromorphic research chips.
One of the rows of Loihi chips that make up Intel’s latest neuromorphic research system, Pohoiki Springs.
2/4
One of the rows of Loihi chips that make up Intel’s latest neuromorphic research system, Pohoiki Springs.
The full rack system of Pohoiki Springs, Intel's latest and most advanced neuromorphic computing system
3/4
The full rack system of Pohoiki Springs, Intel's latest and most advanced neuromorphic computing system
Pohoiki Springs is made up of 768 Loihi chips, arranged in a series of rows in a five-server box
4/4
Pohoiki Springs is made up of 768 Loihi chips, arranged in a series of rows in a five-server box
View gallery - 4 images

As impressive as computers are becoming, they still pale in comparison to nature’s version – the brain. As such scientists have started designing computer chips that work in a similar way to the brain, using artificial neurons and synapses. Now Intel has unveiled its most powerful “neuromorphic” computing system to date. Named Pohoiki Springs, this system packs in 100 million neurons, putting it on par with the brain of a small mammal.

Traditional computer chips are excellent at quickly crunching huge numbers that would make a human’s eyes water. But they aren’t as adept at abstract problems like spotting the difference between dogs and cats, which we can do seemingly without thinking.

Machine learning is an emerging form of artificial intelligence that’s aiming to improve this. By training a system on thousands or millions of examples of what it needs to know, it can learn patterns and become very good at that type of task.

One of the rows of Loihi chips that make up Intel’s latest neuromorphic research system, Pohoiki Springs.
One of the rows of Loihi chips that make up Intel’s latest neuromorphic research system, Pohoiki Springs.

And that’s the type of computing system that Intel is now experimenting with. The Loihi neuromorphic processor packs 130,000 artificial neurons and 130 million synapses, and functions like a human brain. Traditional computers process information in one area then pass it on to another for storage. But the Loihi – like the brain – performs both functions in the same spot, saving time and energy. Plus, the chip rewires its connections over time, boosting that efficiency further.

Intel says that Loihi is as much as 1,000 times faster and 10,000 times more efficient at certain tasks than conventional processors. Just recently, for example, it was put to work identifying and categorizing smells in a new form of electronic nose.

Pohoiki Springs is the latest and most advanced system to use this processor. It’s made up of 768 Loihi chips in one box the size of five standard servers. That boosts the number of neurons to an astounding 100 million, putting it in the ballpark of the brain of a small mammal. And best of all, it runs on a pretty small amount of power, unlike other energy-guzzling servers.

Pohoiki Springs is made up of 768 Loihi chips, arranged in a series of rows in a five-server box
Pohoiki Springs is made up of 768 Loihi chips, arranged in a series of rows in a five-server box

This new system will be accessible via the cloud, allowing members of the Intel Neuromorphic Research Community (INRC) to tap into it to tackle more complex problems.

“Pohoiki Springs scales up our Loihi neuromorphic research chip by more than 750 times, while operating at a power level of under 500 watts,” says Mike Davies, INRC director. “The system enables our research partners to explore ways to accelerate workloads that run slowly today on conventional architectures, including high-performance computing (HPC) systems.”

Intel says Pohoiki Springs could be put to work in many different areas of computing, such as identifying optimal paths for driving directions or deliveries, minimizing risk in stock portfolio returns, or airline scheduling.

Source: Intel

View gallery - 4 images
6 comments
Jose Gros
A great step ahead, but if we consider that Bipolar Transitors simulate Neurons, Dendrites are the Collector, Axons are the Emitter, and Axons from other Neurons regulate the Base as to modulate the Output in Axon/ Emitter, this does not differ very much to a Triode Bulb, where the grill voltage regulates electron flow from the glowing source to the element at other side. The concept of mimicking Neuron network [it was Santiago Ramón y Cajal (Kajal exists today as name in India, watch: 'APU Trilogy') who, using the stains and work by Golgi proposed Neurons are independent from each other, individualized by a membran, not forming a Syncithium as previously thougt, even if it are inside a functional union], appears in the description of the schyzophrenic Computer HAL in the Novel: '2001, an Space Odissey', by Arthur C Clarke, basis for the Stanley Kubrick movie. 'Nothing new under the Sun', but again, the proverb: 'What a man can imagine, another can turn it into reality' proven right. Thanks! Salut +
Riaanh
I don't think we are going to see intelligence soon emerging from these systems. Even 1 000 000 Million logic gates will not be enough. There are far more going on in the brain. For one it is not binary. Fuzzy logic to the extreme, with a measure of analogue in it. The other feature which I feel is greatly under-appreciated by most researchers is the quantum effects. At the quantum level there is a lot of activity between the cells of biological systems. It is not only Logic circuits, macro biology and Newton mechanics which plays a role in life. I believe even more so in our brains. On an esoteric level, quantum entanglement could perhaps explain the bond between twins. ......what we cannot understand we cannot build!
sidmehta
"minimizing risk in stock portfolios" - how useful is that right now, or for future stock market collapses?
paul314
If that could control the muscles of a robot with the coordination most small mammals exhibit, that would be pretty cool.
bwana4swahili
Only a matter of time before: https://flic.kr/p/2ijyqX9 !
James Midkiff
Oh guys...I know that was probably a major feat of human engineering. But I have a feeling that the hardware was the easy part. Let me know when you get the software worked out. My request is that you give whatever mammal you mimic (you copyright infringers you) a small trademark tic. I'll leave it up to the community to come up with a funny one. My youngest cat spontaneously generated the venerated butt wiggle pounce..so I mean, I'm sure it's possible...and probably been done before.