Post #4146371
2026-07-27 20:33 UTC
@anthropy@mastodon.derg.nz @Em0nM4stodon@infosec.exchange
Only recent versions of the tensor cores do support the low precision formats. OG tensor cores (and TPUs) for that matter, would have best performance with fp16 but also support "pseudo" fp32 (same range but less precision) and even fp64 albeit with much reduced performance. These are the useful number formats for scientific and simulation stuff like weather and climate modeling.
The low precision formats are a rather new addition and pretty useless for DSP & sim.
Replies (1)
-
@anthropy@mastodon.derg.nz 2026-07-27 20:35
@datenwolf@chaos.social @Em0nM4stodon@infosec.exchange I mean.. things like modern weather prediction and protein folding runs on neuralnets, which in turn use low precision tensors. While using way less power and compute too. honestly I don't hate neuralnets, or even LLMs (even things like Nvidia DLSS are LLMs so it doesn't say much). I don't think TPUs (or other 'low precision' FPU/ALU hardware) will really become obsolete any time soon.