IBM neurocomputer detailed

Technology News | April 8, 2016

By Julien Happich

Digital Signal Processing

Its chip architecture, array of evaluation boards, reference systems and software ecosystem were described by their architect at the International Symposium on Physical Design 2016. ISPD 2016 is an Association of Computing Machinery (ACM) conference on next-generation chips sponsored by Intel, IBM, Cadence, Global Foundries, IMEC, Oracle, Synopsys, TSMC, Altera, Xilinx and other stellar chip makers worldwide.

From the first microprocessor to today’s supercomputers-on-a-chip, clock speed and power consumption have steadily increased, whereas the IBM TrueNorth designers say we should have been going in the opposite direction – down the curve toward lower-clock speeds and less power consumption, ultimately getting down to the 10-Hz, 20-watt metrics of the human brain.

At ISPD, IBM expressed it aspirations for its brain-like computers, hoping they will become a household name for applications from ultra-smart Internet of Things (IoT) to ultra-smart cars to ultra-smart cameras, ultra-smart drones, ultra-smart medical devices and of course ultra-smart supercomputers.

Fig.1: IBM’s brain-like architecture started in upper left by studying the cortical neuronal structure of the human brain, moving right to structural, functional and finally a physical implementation. Likewise more complex neuromorphic inspirations move from other upper left downward from the core concept, to single chips to multi-chip modules. (Source: EE Times)

In his invited paper “Design and Tool Flow of IBM’s TrueNorth: An Ultra-Low Power Programmable Neurosynaptic Chip with 1-Million Neurons” IBM’s Low-Power Neuromorphic Circuit Designer, Filipp Akopyan described the company’s hardware, software and growing ecosystem of support.

Akopyan said IBM’s target for its TrueNorth chips was Edge-of-the-Net and Big Data solutions, where massive amounts of real-time data need to be processed by ultra-low-power devices—namely its low-cost 5.4 billion transistor neurosynaptic chips that nevertheless consume a mere 700 milliWatts.

Fig. 2: The biggest difference with TrueNorth chips is that they have both traditional synchronous parts (for interfaces and timing) and also asynchronous parts (for the brain-like structures) the flow for which has to be developed separately but simultaneously, as illustrated here. (Source: EE Times)

“Today we are producing massive amounts data with our mobile devices which today must be handled by cloud computers, but TrueNorth can be located on the edge of the network where the data is streaming in to intelligently process that data, only sending up to the cloud important items and summaries,” said Akopyan. “But computing resources are going the wrong direction to higher power, whereas they should be going to lower power preferably scaling down to biological levels.”

IBM starts with a neuromorphic core—a number of neurons connected with dendrites (inputs) and axons (outputs). Any neuron can send a message—called a spike—to any other neuron as a one or a zero (voltage pulse or “spike”). On-chip neurons can message from-any to-any other neuron within each chip—the preferred method of communications since sparse local communications are not only most like those in the brain, but introduce the least amount of signal latency. Any neuron can send a message to others off-chip or even off-board, but obviously with higher latency.

Fig. 3: All the specific components per TrueNorth chip are detailed on this diagram regarding its implementation of brain-like functions. (Source: EE Times)

The key to the 70-milliWatt power consumption of TrueNorth is its use of asynchronous logic that consumes zero power for the entire chip until a particular neuron is used at which point it is turned on only for its period of use communicating with another neuron. Structurally, the 5.4 billion transistor would consume 50-to-100 watts if constructed synchronously, but its asynchronous construction cuts power to the bone.

To enable the any-to-any neuron connection structurally, a gigantic on-chip crossbar switch is used to connect the neurons on the 5.4 billion transistor chip “one of the biggest in world” according to Akopyan.

Fig. 4: IBM’s first board level solution uses a single TrueNorth chip which DARPA and the Lawrence Livermore National Laboratories started with for their Predator drone and nuclear stewardship programs, respectively. (Source: EE Times)

IBM uses tools from many vendors including Cadence, Synopsys and Spice, but it also had to create some of its own EDA tools to support co-design of its partly synchronous and partly asynchronous architecture. The tool created by IBM was dubbed CoSim which, as its name implies, allow it to co-simulate different parts of the chip on different tools simultaneously.

“We used different simulators to design different parts of the TrueNorth chip,” said Akopyan. “Our CoSim tool allows a hybrid tool flow to do custom co-simulations of TrueNorth.”

Fig. 5: IBM’s first sale of its 16-chip board was acquired by Lawrence Livermore National Laboratories, which is currently using the board to simulate explosions of the U.S. aging nuclear arsenal (now that underground tests are banned by international treaty for every nuclear power except North Korea). (Source: EE Times)

The result is a central array of 64-by-64 neurosynaptic cores for 4096 total each with 256 neurons and 64k of memory synapses for a total of one million neurons and 256 million synapses using 5.4 billion transistors that consume just 70 milliWatts of power. Also, the cores can be tiled without bounds, allowing future neuromorphic processors to be expanded as bigger chips become available. The current chip uses Samsung’s 28 nanometer low-power process.

Fig. 6: IBM’s long-term plan is to come out
with larger and larger models of brain-like neuromorphic
computers, the biggest of which – with the current chip –
will be one percent of a human brain. (Source: EE Times)

To demonstrate its TrueNorth chip to potential customers—from IoT makers to military drone makers to supercomputer manufacturers, IBM currently has single-chip stand-alone boards, a chassis that holds 16 boards, one holding 48 boards, a single board containing 16 TrueNorth chips with plans for chassis holding 64-, 256-, 1024- and 4096-chips, the latter of which is 1 percent as big as the human brain.