The article covers software approaches (more energy-efficient algorithms) and mentions GPUs but not TPUs or ASICs.
Specialized chips (built with dynamic fabrication capacities) are far more energy efficient for specific types of workloads. We see this with mining ASICs, SSL accelerators, and also with Tensor Processing Units (for deep learning).
The externalities of energy production are the ultimate concern. If you're using cheap, clean energy with minimized external costs ("sustainable energy"), the energy-efficiency of the algorithm and the chips is of much less concern.
Could we recognize products, services, and data centers that were produced with and/or run on directly sourced clean energy as "200% Green"; with a logo on the box and/or the footer? 100% offset by PPAs is certainly progress.
"Unlearning" is one algorithmic approach that may yield substantial energy consumption gains.
With many deep learning models, it's not possible to determine when or from what source something was learned: it's not possible to "back out" a change to the network and so the whole model has to be re-trained from scratch; which is O(n) instead of O(1.x).
2 comments
[ 2.9 ms ] story [ 12.2 ms ] threadSpecialized chips (built with dynamic fabrication capacities) are far more energy efficient for specific types of workloads. We see this with mining ASICs, SSL accelerators, and also with Tensor Processing Units (for deep learning).
The externalities of energy production are the ultimate concern. If you're using cheap, clean energy with minimized external costs ("sustainable energy"), the energy-efficiency of the algorithm and the chips is of much less concern.
Could we recognize products, services, and data centers that were produced with and/or run on directly sourced clean energy as "200% Green"; with a logo on the box and/or the footer? 100% offset by PPAs is certainly progress.
With many deep learning models, it's not possible to determine when or from what source something was learned: it's not possible to "back out" a change to the network and so the whole model has to be re-trained from scratch; which is O(n) instead of O(1.x).