@thelinuxexperiment

The 'AI accelerator' module @mechasystems are testing for the Mecha Comet is made by the militaristic Israeli AI firm Hailo [1,2].
The chairperson of Hailo was formerly president and CEO of the zionist Bank Leumi which contribute significantly to the decades-long colonisation and occupation of Palestinian land, and contributes significantly to Israels ongoing genocide of palestinians [3-5].
Bank Leumi is on the divestment shortlist of the BDS movement [6,7].

Supporting and legitimising genocide is not cool of anybody. That includes producers of an otherwise very cool little Linux handheld device.

[1] https://hailo.ai/products/ai-accelerators/hailo-8l-m-2-ai-acceleration-module-for-ai-light-applications/

[2] https://en.wikipedia.org/wiki/Hailo_Technologies

[3] https://en.wikipedia.org/wiki/Rakefet_Russak-Aminoach

[4] https://en.wikipedia.org/wiki/Bank_Leumi

[5] https://wearedissenters.org/bank-leumi-an-instrument-of-settler-colonialism-and-genocide/

[6] https://investigate.afsc.org/company/bank-leumi-le-israel

[7] https://investigate.info/divest

#mechaComet #hailo #Hailo8L #hailo8 #boycottIsrael #bds #linux #StopGenocide

@kbob @drwho

On Halio's site for the chip (https://hailo.ai/files/hailo-8l-m-2-et-product-brief-en/) they don’t talk about FLOPs, they talk about “operations” (and claim 13 teraops).

A quick search turns up nothing but marketing fluff, though this one is pretty interesting: https://hailo.ai/products/technology/

It’s a dataflow processor array with a flexible internal network. They also talk about the cleverness of their compiler technology and how it takes advantage of their dataflow architecture.

While they have ported Tensorflow and Pytorch (plus a couple of other neural-net frameworks), there’s no indication what size floating point operations they’re doing. Typically, “AI processors” save energy by using fewer bits — 16, 8, or even fewer (I’ve saw a paper the other day that talked about a mixture of 1 and 2 bit operations).

So, can your DSP application be built using Tensorflow or Pytorch? Does it not depend on great floating point accuracy?

#raspberrypi #hailo8