Unfortunately this is not true in numerics. Lots of stupid heavy cfd/fea type workloads parellize well but aren't gpu accelerated. The reasons aren't clear to me, but a lot of the popular solvers are cpu only and involve mostly fp calcs. There are a few solvers that use gpus but they tend to be less accurate in exchange.
Reasons : there is a significant amount of work needed to get codes to work on a distributed hybrid or gpu-only fashion. It's a completely different coding paradigm that needs significant studies before commercial entities adopt gpu use at scale. All-gpu solvers are starting to be developed, such as fun3d GPU[0], but features are very limited. GPU development is starting to catch up in the community, so it won't be long before a significant portion can operate heterogeneously or in gpu-only mode.
'this is not true in numerics' - shows no evidence...
GPUs are gaining traction in FP workloads, it can be seen clearly with CPU/GPU data-center market share
Moore's law is pretty much over, we can't simply print more performance these days, we are going to see major shift to accelerators which would require some rewrites, otherwise you're going to be stuck
eyegor|2 years ago
uguuo_o|2 years ago
[0] https://fun3d.larc.nasa.gov/GPU_March_2021.pdf
CyberRage|2 years ago
GPUs are gaining traction in FP workloads, it can be seen clearly with CPU/GPU data-center market share
Moore's law is pretty much over, we can't simply print more performance these days, we are going to see major shift to accelerators which would require some rewrites, otherwise you're going to be stuck
_hypx|2 years ago