Success and Failure Stories
The slides can be downloaded from here.
Some wrap-up of the most common reasoning for and against GPGPU computation (on a very high level):
- Hundredfold performance gain for some (easily parallelizable) applications
- Hard to program
- Restrictive architecture, memory access
- Different execution paths cause idle time on other threads
- Performance gains hard to get for other applications than the ones that properly fit the model
- Host-device memory transfer is slow ¿ latency
- Not usable for smaller problem instances
I've only listed a single "pro", but it's a fairly major one...