Watch Desk posted an update
Independent modders have released Linux-only firmware tools that can nearly double certain AI maths speeds on Nvidia’s RTX 20-, 30- and 40-series graphics cards, MakeUseOf reports. That could give local-model users, developers and researchers more tensor throughput from consumer hardware they already own.
Why it mattersThe report says an RTX 4090’s performance in the relevant calculations rose from 165.6 to 331.6 trillion operations per second. This is for FP16 and BF16 matrix maths, not faster games or general graphics performance. A useful distinction, particularly if you were hoping for a free frame-rate upgrade. Trying the unlock requires manual boot configuration and voids the hardware warranty, the report says. Would you risk that trade-off for faster local AI, or is the warranty worth more than the extra compute?
Discuss: Would you risk voiding your graphics card’s warranty for faster local AI, or is the warranty worth more than the extra compute?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.