4 ms·
Sure, you can't run a 1990s-era HPC application unmodified on a GPU, but you can write applications that are functionally equivalent (take the same inputs, prod
by jonas21 6y ago
Sure, you can't run a 1990s-era HPC application unmodified on a GPU, but you can write applications that are functionally equivalent (take the same inputs, produce the same outputs) that do the heavy lifting on the GPU.
The latest DoE supercomputers are mostly GPUs. Summit has around 10 PFLOPS on CPUs and 215 PFLOPS on GPUs.
https://en.wikichip.org/wiki/supercomputers/summit https://en.wikichip.org/wiki/supercomputers/summit
- philipkglass 6y agoYes, I agree with that. But it also takes a lot of software work. Getting decent performance on a GPU for an existing scientific application requires something closer to a full rewrite than the usual porting effort to a novel CPU architecture. That is one of the reasons that people are writing new scientific applications to target GPUs from the beginning rather than just adapting mature older programs to GPUs. (Though there is some of both approaches.)