Please enable JavaScript.
Coggle requires JavaScript to display documents.
Week 8 Distributed (GPGPU (Limitations (Very low level programming. Thus…
Week 8 Distributed
GPGPU
-
Differences from GPU
- Programmable pipelines
- Higher precision integer arithmetic units
- Higher precision floating point arithmetic units
-
-
-
Problems in integration
Should not demand more cores than what the GPGPUs can provide. Then we lose the speed of the private memory
If many GPGPUs connected, need to have high bandwidth main bus to handle traffic.
-
Vector processing
-
-
-
-
Loop parallelising
If there is a loop that doesn't depend on previous iterations, then just put each iteration as elements in a vector and do vector computation for improved performance
Multi Processing
Multiple CPUs sharing same main bus, memory and IO
-
-
-
Issues
-
Cache coherency
-
-
Snooping technique uses intelligent caches.
- Broadcast any writes to other caches
- Each cache updates themselves
Data dependency
-
-
Levels
-
-
-
Depending on which, we can use different techniques to make it more parallel.
-
Grid app performance
Gaming, military & parametric processing since messaging is short & infrequent
Parametric and mesh based computations need to be looked out for dependencies & need a good way to divide problem
Parallelism Problem
Exploit opportunities for concurrency. Basis for superscalar, parallel, grid