Apparently Gemma 4 (by deepmind) are very energy-efficient for their capabilities.
The post itself seems to recommend a company (neuralwatt.com) but it’s unrelated to project wagtail (an open source CMS made with django) so I consider it legit, they’re just vouching for this new approach that shows users their ecological footprint basically.
(Of course, running these at a homelab is possible. hence the relevance of this info for the community)
If you’re running on AMD hardware, drop the power cap to save energy. (e.g. by poking values into /sys/class/drm/card1/device/hwmon/.../power1_cap or similar; check the min and max files adjacent to it for the ranges.) I can lower the power usage by 100W on a discrete AMD GPU and only lose about ~3% tok/s generation (inference is limited more by memory bandwidth than by compute typically on modern hardware). Test different settings on your actual hardware to find the sweet spot balancing performance vs energy usage!
Yeah I run the Gemma 4 12B QAT on my homeserver. It has a 12GB GPU (Nvidia A2000) and that model gives the most bang for the buck being both fast and multimodal.
