3 ms·
What Is Heterogeneous System Architecture (HSA)?
- valarauca1 11y agoThis needs a (2012) label. HSA already happened, it didn't live up to the hype.
- acaloiar 11y agoIt's worth mentioning that while HSA specifically may or may not have lived up to its "hype", it did succeed in being on of the first technologies (to my knowledge) to provide hardware support for cache-coherent memory between processing elements. This is an important concept to parallel computing because there is significant latency involved in transferring data between CPU and GPU RAM. HSA backs OpenCL 2.0's SVM (Shared Virtual Memory) and CUDA 6's Unified Memory with actual hardware, making these APIs more than convenient software abstractions. Hardware unified memory opens up a whole new set of algorithms to GPU-parallelization by removing the memory transfer latency barrier.
- fulafel 11y agoWe're still waiting for the software side to shape up. Operating systems and legacy driver stacks are slow to change, and OpenCL's decline (due to Apple's disengagement) doesn't really help. There's stuff happening though, here's a recent article about GCC 6 & HSA now that drivers are starting to arrive in Linux: https://www.phoronix.com/scan.php?page=news_item&px=hsa-offloading-gcc6-hopes https://www.phoronix.com/scan.php?page=news_item&px=hsa-offl... Real world GPGPU applications are currently almost exclusively on CUDA, and GPGPU programming language development seems stagnant. A return to the roots (=games & consoles) for GPUs?