Talk:TrueNorth

Talk
"TrueNorth circumvents the von-Neumann-architecture bottlenecks and is very energy-efficient, boasting a power consumption of only 1/10,000th of conventional chips, 63 milliwatts". There is something wrong with this bit of information. 63 thousandths of a watt times 10,000 yields 630 watts. What single electronic chip of any kind uses this much power? Is it possible someone connected with the authoring of this article confused milliwatts (thousandths of a watt) with microwatts (millionths of a watt)?. This would make sense to me then as both power figures would be far more plausible. 63 millionths of a watt for the neuromorphic chip versus 630 thousandths of watt (6.3 Watts) for a conventional chip.

98.206.127.121 (talk) 01:25, 1 February 2015 (UTC)


 * Agreed. IBM's own article here says 70mW. An Intel Xeon E7-8880L v3 has a similar transistor count (5.7bn) and has a 115W TDP, so this directly contradicts the claim in this article. I'm going to remove it. 82.0.26.183 (talk) 22:27, 18 February 2019 (UTC)


 * Actually on reflection, what the sentence actually says is that the power density is 10,000 times better than conventional microprocessors. The example I gave above is an Intel Xeon E7-8880L v3, which has a 115W TDP over a 6.62 cm2 area - a density of 17371 mW/cm2. IBM claim a density of 20 mW/cm2 which is only 870x as good as my randomly chosen example. Perhaps there are some specifics of the type of load that they are factoring in (e.g. the inefficiencies of representing a neuron or synapse in a traditional processor architecture) but I find it hard to believe that other technologies such as nVidia Tensor don't compete in this market. I'll leave the claim for now but change the wording to make it clear that it isn't a hard fact. 82.0.26.183 (talk) 22:41, 18 February 2019 (UTC)

syNAPSE University
No evidence of such has been found on the Internet and no known programs at any universities show up in an EDU search. The claim of a totally new architecture is not claimed by IBM and relearch papers all reference many preexisting architectures. The Cognimem was capable of being taught and was actually usable. we should not make claims on behalf of IBM that IBM is not willing to make them selves. Even IBM quotes should be avoided if they are from earlier documents that IBM has walked back or no longer asserts. Scottprovost (talk) 08:39, 28 May 2016 (UTC)

Programmeable Synapses
The article should more clearly describe the principle limitation of True North: the synaptic weights cannot be learned in situ, in real tasks they must be learned offline by some training procedure (e.g. in a conventional computer) and then programmed onto the chip. This training/importation procedure may consume much more energy than the chip itself so the energy efficiency of the chip might be illusory Paulhummerman (talk) 11:57, 29 October 2016 (UTC)