I don’t agree with your point on CUDA. GPUs have been the largest upset in HPC in the last decade, and are actually incredibly good at certain tasks. E.g. in my thesis, I ported the gravitational solver for cosmological structure formation simulations to the GPU using CUDA/HIP and it was actually an order of magnitude faster.
So basically, some time after the Big Bang, the universe was very dense and fairly homogeneous. The question of structure formation is how one gets from this state to today’s universe. This can be used to check whether our understanding of our universe is correct by basically looking at whether the resulting simulated universe looks like our universe.
The way this is actually done is by placing massive (pseudo)particles (and with massive, I’m talking about like 10^10 sun masses, depending on the simulation scale) according to this distribution and looking at how they move under the influence of gravity (Newtonian gravity is enough as the gravitational fields are fairly weak). So in this simple model, the numerical challenge boils down to the gravitational interaction between a fuck ton of particles.
The problem is that for N particles, there are N(N-1) calculations that need to be calculated. This is not feasible for large N, so instead, one needs more sophisticated algorithms. And this is basically the part I ported to the GPU.
I know for a fact it’s used in implant development (joint is scanned producing a point cloud which CUDA is then used to turn into an actual model). You could I guess still call that “research” but it’s an actual application used by a company to develop a product.
I don’t agree with your point on CUDA. GPUs have been the largest upset in HPC in the last decade, and are actually incredibly good at certain tasks. E.g. in my thesis, I ported the gravitational solver for cosmological structure formation simulations to the GPU using CUDA/HIP and it was actually an order of magnitude faster.
Gravitational solver for cosmological structure formation…
So basically, some time after the Big Bang, the universe was very dense and fairly homogeneous. The question of structure formation is how one gets from this state to today’s universe. This can be used to check whether our understanding of our universe is correct by basically looking at whether the resulting simulated universe looks like our universe.
The way this is actually done is by placing massive (pseudo)particles (and with massive, I’m talking about like 10^10 sun masses, depending on the simulation scale) according to this distribution and looking at how they move under the influence of gravity (Newtonian gravity is enough as the gravitational fields are fairly weak). So in this simple model, the numerical challenge boils down to the gravitational interaction between a fuck ton of particles.
The problem is that for N particles, there are N(N-1) calculations that need to be calculated. This is not feasible for large N, so instead, one needs more sophisticated algorithms. And this is basically the part I ported to the GPU.
Oh sure, for research is great… and this is also why I learned CUDA.
Outside of research?
I know for a fact it’s used in implant development (joint is scanned producing a point cloud which CUDA is then used to turn into an actual model). You could I guess still call that “research” but it’s an actual application used by a company to develop a product.