Leveraging Hardware Accelerators (HWA) is paramount in the area of high-performance computing and machine learning. Our solution incorporates robust support for hardware accelerators through the Direct Memory Connector (DMC). This offers a seamless bridge between externally allocated memory and the ETAS DMS platform.
DMC caters to diverse HWAs from various vendors, including general-purpose GPUs and specialized HWAs for Vision programming, Convolutional Neural Networks (CNNs), and Tensor processing.
Integration with NVIDIA's middleware framework, NvSci, ensures efficient memory allocation and synchronization for devices like NVIDIA Xavier and Orin. This framework unifies host and device memory, incorporating CUDA, TensorRT, and NvMedia for optimal performance.
CARMA, our middleware platform, provides a dedicated wrapper to facilitate direct access to NvSci-allocated memory for CPU processing. This eliminates the need for redundant data copies, enhancing overall efficiency.
Specific HWA business logic, essential for high-performance tasks, is crafted manually by software developers. This tailored approach ensures optimal utilization of the underlying hardware capabilities.
The Gateway, comprising handwritten NvSci code, facilitates zero-copy data transfer from HWA to CPU activities. This seamless integration allows for efficient processing without the overhead of data duplication.
The MWALA plugin implements the MWALA API for NvSci, enabling CPU activities to leverage custom NvSci code seamlessly. This integration, coupled with the NVSci Wrapper, ensures a smooth transition of HWA-acquired data into CPU processing workflows.
The CARMA platform, with its NvSci Wrapper, eradicates the need for copying extensive payload data to CPU-accessible memory, enhancing overall processing speed and efficiency.
The manual crafting of HWA business logic allows for tailored integration, ensuring that the unique capabilities of each HWA type and model are fully utilized.
Leveraging the MWALA 2.0 abstraction on the ETAS DMS side ensures seamless middleware integration, making the system HWA-agnostic and versatile.
The HWA support via DMC empowers developers to harness the full potential of Hardware Accelerators, providing a streamlined and efficient solution for high-performance computing and machine learning applications.