Intel Compute Runtime 26.44.40734.6 optimizes large buffer copies for Xe and Arc

Intel has released version 26.44.40734.6 of its Compute Runtime (CRI), which introduces a performance optimization for buffer copy operations. This update specifically targets Intel Xe and Arc hardware architectures by adjusting how the driver handles large memory transfers.

The change introduces a conditional logic for selecting copy kernels based on the size of the aligned middle region. While smaller transfers continue to use a 16-byte kernel for maximum efficiency, larger transfers will now utilize an 8-byte kernel to optimize throughput for massive data movements.

Buffer copy kernel optimization

• Kernel Selection Logic: The driver now switches to an 8-byte middle-region copy kernel only when the aligned middle region meets or exceeds a specific threshold provided by the product helper.
• Threshold Parameters: For the Compute Runtime (CRI), this threshold is set at 32 MiB; transfers smaller than this value retain the 16-byte kernel to maintain higher speeds for smaller workloads.
• Hardware Specificity: This 8-byte kernel optimization is exclusive to the Intel Compute Runtime and is not applied to other Intel products within the driver stack.
• Workload Impact: The update aims to improve performance specifically for large-scale data copies, addressing the efficiency of the middle-region copy process during heavy compute tasks.

Why it matters

This update improves the efficiency of GPGPU workloads on Intel Arc and Xe-based systems, which is critical for developers working with OpenCL or compute-heavy applications. By optimizing large memory transfers, Intel is refining the performance characteristics of its open-source compute stack for professional and high-performance computing use cases.