cross pond high tech

At its re:Invent conference, AWS today announced the launch of its Inferentia chips, which it initially announced last year. These new chips promise to make inferencing, that is, using the machine learning models you pre-trained earlier, significantly faster and cost effective.

As AWS CEO Andy Jassy noted, a lot of companies are focusing on custom chips that let you train models (though Google and others would surely disagree there). Inferencing tends to work well on regular CPUs, but custom chips are obviously going to be faster. With Inferentia, AWS offers lower latency and three times the throughput at 40% lower cost per inference compared to a regular G4 instance on EC4.

The new Inf1 instances promise up to 2,000 TOPS and feature integrations with TensorFlow, PyTorch and MXNet, as well as the ONNX format for moving models between frameworks. For now, it’s only available in the EC2 compute service, but it will come to AWS’s container services and its SageMaker machine learning service soon, too.

Apple emprunte à son tour la "Verticale du Fou" et abandonne Intel. 3 fois plus performant par Watt dissipé, le processeur "maison" #M1 — annoncé le 10 Novembre dernier — semble confirmer à quel point l'architecture #ARM a réussi à rattraper puis à dépasser l'architecture #X86 du géant de Santa Clara.

Les 16 milliards de transistors du M1 sont en effet gravés en 5 nanomètres, là où Intel ne parvient toujours pas à maîtriser ni 10 ni 7 nanos.

Un excellent article de Tom's #Hardware montre l'étendue de la menace qui pèse désormais sur Intel ; ce n'est pas tant la part de marché du Mac (à peine 9%) que celle de ses développeurs (30%) qui vont de plus en plus exclusivement basculer dans l'architecture ARM.

Neural networks are the crown jewel of the AI boom. They gorge on data and do things like transcribe speech or describe images with near-perfect accuracy (see “10 breakthrough technologies 2013: Deep learning”). The catch is that neural nets, which are modeled loosely on the structure of the human brain, are typically constructed in software rather than hardware, and the software runs on conventional computer chips. That slows things down. IBM has now shown that building key features of a neural net directly in silicon can make it 100 times more efficient. Chips built this way might turbocharge machine learning in coming years. The IBM chip, like a neural net written in software, mimics the synapses that connect individual neurons in a brain. The strength of these synaptic connections needs to be tuned in order for the network to learn. In a living brain, this happens in the form of connections growing or withering over time. That is easy to reproduce in software but has proved infuriatingly difficult to