http://developer.amd.com/resources/heterogeneous-computing/what-is-heterogeneous-system-architecture-hsa/ So the challenges of homogeneity and heterogeneity in system architecture are radically different. From my understanding, CPU's are better suited to running the operating system and dealing with memory, but GPU's are better suited towards parallelism and large data sets/computation. This is caused mostly by the differences in architectures of the two different units. From http://superuser.com/questions/308771/why-are-we-still-using-cpus-instead-of-gpus, GPU's are Turing complete and can thus run any computer algorithm; thus, one would be inclined to believe that GPU's can run algorithms that aren't particularly targeted towards GPU usage.
I imagine that if a specific set of the silicon dye is devoted to something, say rasterization, then even if we end up with not enough computing power of that particular unit, we can give the rest to other units which are not necessarily as suited to run those algorithms (such as the GPU or the CPU if it is idle). In this sense, we can still maintain the high amount of benefit from specialization, with a little bit of lost time from the other units trying to run the algorithms (but still getting it done in parallel).
This comment was marked helpful 0 times.
benchoi
The PetaBricks language helps with the automatic generation of architecture-specific implementations, which reduces the portability nightmare while giving good performance. Programmers offer the compiler choices in how a program might be implemented, allowing it to optimize for the appropriate architecture.
This comment was marked helpful 1 times.
bwasti
@pwei: GPUs are good with large datasets, up to a point. CPUs of today's era have far more memory available (extremely high-end GPUs can have up to 6 GB of memory, whereas a common laptop has around 8 GB). Although pulling from disk is a nightmare, CPUs can still do it, whereas GPUs are limited, in many cases, to no more than the memory available. The bandwidth is obviously far superior on a GPU, but if you have an extremely large dataset, it becomes infeasible to use them.
http://developer.amd.com/resources/heterogeneous-computing/what-is-heterogeneous-system-architecture-hsa/ So the challenges of homogeneity and heterogeneity in system architecture are radically different. From my understanding, CPU's are better suited to running the operating system and dealing with memory, but GPU's are better suited towards parallelism and large data sets/computation. This is caused mostly by the differences in architectures of the two different units. From http://superuser.com/questions/308771/why-are-we-still-using-cpus-instead-of-gpus, GPU's are Turing complete and can thus run any computer algorithm; thus, one would be inclined to believe that GPU's can run algorithms that aren't particularly targeted towards GPU usage.
I imagine that if a specific set of the silicon dye is devoted to something, say rasterization, then even if we end up with not enough computing power of that particular unit, we can give the rest to other units which are not necessarily as suited to run those algorithms (such as the GPU or the CPU if it is idle). In this sense, we can still maintain the high amount of benefit from specialization, with a little bit of lost time from the other units trying to run the algorithms (but still getting it done in parallel).
This comment was marked helpful 0 times.
The PetaBricks language helps with the automatic generation of architecture-specific implementations, which reduces the portability nightmare while giving good performance. Programmers offer the compiler choices in how a program might be implemented, allowing it to optimize for the appropriate architecture.
This comment was marked helpful 1 times.
@pwei: GPUs are good with large datasets, up to a point. CPUs of today's era have far more memory available (extremely high-end GPUs can have up to 6 GB of memory, whereas a common laptop has around 8 GB). Although pulling from disk is a nightmare, CPUs can still do it, whereas GPUs are limited, in many cases, to no more than the memory available. The bandwidth is obviously far superior on a GPU, but if you have an extremely large dataset, it becomes infeasible to use them.
This comment was marked helpful 0 times.