Running #TornadoVM using the Intel ARC on Windows 11 via MinG64 using the OpenCL Backend.

Integrated GPUs, such as Intel HD graphics, share the memory with the main CPU, offering a common view of the main memory. This means that, potentially, we could use buffer pointers allocated on the host side on the GPU, and, therefore, save the data transfer time (e.g., a copy from host memory to device memory, which can be expensive in many cases). However, does share memory really impact performance if we measure end-to-end applications on GPUs? In this post, we try to answer this question.

Level-Zero is a close to bare-metal API for programming heterogeneous architectures, and it is shipped as part of Intel oneAPI. Additionally, it can be used as a standalone API. This article shows the basic architecture, what is used for, and an example for dispatching matrix multiplication on the Intel HD Graphics with SPIR-V.