4 ms·
Very cool concept to have compute and memory unified. On first glance, this seems to fall in the middle between GPU & CPU based computation, but this is ultima
by holstvoogd 6y ago
Very cool concept to have compute and memory unified.
On first glance, this seems to fall in the middle between GPU & CPU based computation, but this is ultimately a graph processor. So no magic computational-memory performance panacea yet unfortunately ;)
I wonder if graph processing might be more broadly applicable than matrix processing as done on GPU's. Atleast 'sprinkeling' computation between memory seems like a good idea for HPC. Moving data is expensive and slow.
- yvdriess 6y ago> I wonder if graph processing might be more broadly applicable than matrix processing as done on GPU's. Ultimately, yes. Graph processing is slow on traditional architectures because it's a very sparse compute problem, which defeats most performance devices developed to help the memory wall: D cache, TLB, etc are mentioned in the paper. For sparse compute, you just hit that memory wall straight up, and things like caches are more a liability than a help. There are plenty of other spare compute problems that do not fall into the graph processing category. Genomics springs to mind (plenty of graphs there too though). Another factor is that a lot of sparse problems are kind of avoided in research, because it is usually more optimal to reformulate it into a dense problem. With hardware like this, I fully expect certain approaches seeing a comeback. Fully sparse neural nets come to mind.