- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
2024.2 DPC++/C++ is ready with the newly released Intel® oneAPI Base Toolkit 2024.2
Highlighted features:
- Extended CPU Feature Support: New checks for CPU features (APX, USER_MSR, AVX10.1) allow software to utilize these features if available, enhancing performance.
- Optimized ABI Implementation: Improved handling of stack alignment to reduce performance impact, especially at program entry points.
- Low-Level Hardware Control Enhancements: Added support for WRMSRNS, MSRLIST, and PBNDKB intrinsics and CPUID enabling, providing more control for developers.
- CCMP and CTEST Optimizations: New support for these optimizations reduces code branches and minimizes performance issues caused by branch misprediction.
- Compress Loop Optimization: Improved handling of specific loop conditions to enhance performance and reduce cache pollution.
- SYCL Level Zero Backend Improvement: Removal of a manual workaround simplifies usage and improves performance, particularly for deep learning applications.
- Profiling with Graphs: Warnings (instead of errors) will be issued for certain profiling properties, with future full support planned to help optimize deep learning workloads.
- Enhanced Memory Management: New strategy to handle large private memory usage in kernels, improving memory management and performance.
- Handling Coordinates in Packed VNNIed Matrix: Coordinates now return in VNNIed matrix format, with the shape expressed in the non-VNNI logical matrix.
- Floating-Point Accuracy Control: Users can now control the accuracy of floating-point math functions, balancing performance and accuracy.
- Enhanced SLM Functions: Improved functions to accept larger data types, offering more flexibility for developers.
- Removed FPGA Support: FPGA support has been removed; refer to specific release notes for FPGA-related updates.
- Floating-Point Model Update: The default floating-point model now honors NaNs and infinities, changing the default optimization behavior and potentially affecting performance. Users can revert to the previous behavior if needed.
Download the Toolkit here.
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
The Intel® one API Base Toolkit 2024.2 introduces a variety of enhancements in both CPU support and performance optimization. Key improvements include support for new CPU features (APX, USER_MSR, AVX10.1), optimized ABI implementation, and enhanced memory management, all aimed at improving performance and developer control. It also features better handling of floating-point accuracy, loop optimizations, and extended support for deep learning applications through SYCL Level Zero Backend. While FPGA support has been removed, the toolkit offers better profiling, low-level hardware control, and new optimizations for minimizing branch mispredictions.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page