Dr. Moritz Lehmann
@ProjectPhysX@mast.hpc.social
Summa cum laude Physics #PhD 🖖🧐🎓 | Graduate at EliteNet Bavaria 🧬 & DLR 🛰 | Developer of #FluidX3D #CFD 🌊 | Khronos #OpenCL Advisor 💻 | #GPU Wizard at #Intel 🟦 https://github.com/ProjectPhysX/FluidX3D
mast.hpc.social
Replying to
@azonenberg@ioc.exchange
@azonenberg@ioc.exchange @Methylzero@mast.hpc.social @niconiconi@mk.absturztau.be for #HPC workloads the low arithmetic intensity is just perfect. High memory bandwidth (1.5 TB/s) and capacity are the star of the show. Even with half the die size and compute, you would still get the same performance in lattice Boltzmann for example, so memory bound these codes are.
Similar story for some AI inference workloads, and that's what drives the market price of pretty much all GPUs nowadays.