Move Heterogenous Required with airSlate SignNow

Move heterogenous required

Welcome back to the oneAPI channel.​ In previous video we’ve seen a SYCL simple code that can run on a CPU and a GPU​ Why is this so important and will only become mode important in the future?​ What do we need accelerators for?​ Why do we see so many silicon options and new accelerators? Take a CPU for example..​ I can run many applications and record the performance for each of them..​ A CPU can usually run everything, any piece of code, But when the workload we are trying to run is very specialized, for example rendering 3D graphics, ​ our CPU, as powerful as it is, ​ cannot compete with an architecture that was built specifically to execute such workload, a GPU for example​ You can see in this diagram, that a GPU can run a smaller set of applications, a smaller set of instructions​ , but with much higher performance…​ And we might have an AI accelerator like Intel VPU or Habana Gaudi processor, ​ that might look like that. And can accelerate AI.. Deep learning training or inference​ Or you might have an FPGA that theoretically can be re-programmed to the range of applications we want it to support​ So you might have a bitstream that put the FPGA here on the left side of the graph, supporting specific set of applications Or another bitstream that supports a different set of applications.. And this is why we need accelerators​ And the truth is that, in order to cover the full range of applications, we usually need two devices​ Usually a CPU and an accelerator… what we call a heterogeneous system​ ​ Each of these technologies offers breakthroughs in power/performance​ But with that comes a series of problems​ Each accelerator architecture carries its unique toolchain ​ It make sense, they support different instruction set, different memory maps and so on..​ But the result is that all of these pieces of code, are different​ Which means longer time of development, longer development process​ It makes it more expensive to move from one accelerator to the other ​ from one vendor to the other​ It becomes more expensive to find programmers with the right expertise​ and software becomes the bottleneck​ ​ What oneAPI is offering is a unified software environment for all of these devices​ So you can have one code base for a CPU and an accelerator, a GPU for example​ We will explain the SW Stack underneath Data-Parallel-C++ in another video and there is lots of materials online​ ​ But our target is simple, and this is what we will explain during this series of short videos.​ And it is How to program heterogenous and parallel systems using SYCL, with the DPC++ compiler​ In the next video, ​ How SYCL can help you program a heterogeneous system​ See you there..​