• Parameters with reference type are now supported for __global__ functions. If the object bound to the reference is read or written from the GPU during the execution of the __global__ function, the object must reside in memory that is accessible from the GPU.

• The code base for CUDA-GDB was upgraded from GDB 7.2 to GDB 7.6, which enables support for DWARF 4 and thus enables distributions using GCC 4.8. GDB 7.6 also provides better ARM debugging support: allowing single stepping between 32-bit and 16-bit instructions.
• To mitigate the issue of variables not being accessible at some code addresses, the debugger offers two new options. With set cuda value_extrapolation, the latest known value is displayed with (possibly) a prefix. With set cuda ptx_cache, the latest known value of the PTX register associated with a source variable is display with the (cached) prefix.
• It is now possible for the set cuda break_on_launch option to break on kernels launched from the GPU. Also, enabling this option does not disable deferred kernel launch notifications.
• Remote debugging has been made considerably faster: up to two orders of magnitude. Local debugging is also considerably faster.
• CUDA-GDB can now use optimized methods to single-step a program; these accelerate single-stepping most of the time. This feature can be disabled with set cuda single_stepping_optimizations off.

• The CUDA-MEMCHECK tool now works with applications that use the multi-process server (MPS) to share a single GPU among multiple processes.

• The nvprof profiling tool now works with multiple processes that are sharing a single GPU with the multi-process server (MPS).

• Nsight Eclipse Edition now supports developing (building, debugging, and profiling) CUDA applications on remote systems.

• The Visual Profiler guided analysis system has been updated with several enhancements.
  • It now includes a side-by-side source and disassembly view annotated with instruction execution counts, inactive thread counts, and predicated instruction counts. This view enables you to find hotspots and inefficient code sequences within your kernels.
  • It has been updated with several new analysis passes: (1) kernel instructions are categorized into classes so that you can see if the instruction mix matches your expectations, (2) inefficient shared memory access patterns are detected and reported, and (3) the per-SM activity level is presented to help you detect detect load-balancing issues across the blocks of your kernel.
  • It can now generate a kernel analysis report. The report is a PDF version of the per-kernel information presented by the guided analysis system.

• The cuBLAS library now supports execution of Level-3 BLAS routines out-of-core (that is, out of system memory, even if the total data is too large to fit completely in GPU memory) and across two GPUs in parallel. This functionality is exposed via the cublasXt interface. Note that the GPUs must both be on the same GPU board, such as a Tesla K10 or GeForce GTX 690; furthermore, with the appropriate licensing, more than two GPUs on the same board can be supported (seehttps://developer.nvidia.com/cublasxt).
• CUDA 6.0 includes a new library called NVBLAS, which intercepts certain calls to a host BLAS implementation and routes them to cuBLAS via the cublasXt interface. For an existing application that makes use of BLAS calls, NVBLAS allows for drop-in GPU acceleration with only relinking.
• Conversion LAPACK routines cublas{T}tpttr and cublas{T}trttp, which convert from Packed Triangular format to Triangular and to Packed Triangular format from Triangular, have been added to cuBLAS.

• The cuFFT library now supports execution of batched and single FFTs across two GPUs by using the cufftXt interface. Note that the GPUs must both be on the same GPU board, such as a Tesla K10 or GeForce GTX 690. As of CUDA 6.0, all types of batched transforms are supported, with the exception that strides are not allowed. For single transforms, C2C/Z2Z 1-D, 2-D, and 3-D transforms with power-of-2 sizes are supported.

In CUDA 6.0, support for a new random number generator, Mersenne Twister 19937, has been added. This generator is widely recognized to be very efficient at producing random variates with excellent statistical properties. The generator is callable from the host and generates results in GPU memory. A kernel-callable interface is not available at this time.

• The routines cusparse(X)bsrsv2, cusparse(X)bsric02, and cusparse(X)bsrilu02 have been added to the cuSPARSE library. They perform triangular solve, incomplete Cholesky factorization, and incomplete LU factorization on the matrices stored in block CSR format, respectively. The routines handle arbitrary block sizes; moreover, the user can query the size of working space used internally and pre-allocate the buffer, passing a pointer to it to the routines. If the matrix becomes numerically singular (during factorization), the user can query its properties and use the boost value to proceed with computation.

• Support for over 80 new SIMD instructions (__v*2() and __v*4(), particularly useful in video processing) has been added to the math library.

• The NPP library in the CUDA 6.0 release contains more than 500 new image and signal processing primitives, including BGR/YUV conversions, color transformations using a 3D LUT with trilinear interpolation, optimized Huffman coding for JPEG, filter median routines, error metric computations, and other miscellaneous color and pixel routines.
• In CUDA 6.0, NPP adds a new primitive, FilterMedian, to filter the image using a median filter. The new primitive supports variants with different data types and channel settings.

• The following issue has been fixed in CUDA 6.0. On the GK110, the kernel occasionally may produce incorrect results when, either by loop unrolling or straight lines of code, there are more than 63 outstanding texture/LDG instructions at one point during the program execution.

• Before Nsight 6.0, the CPU architecture was set in the linker and compiler properties. Now this setting is configurable on a per-target system basis.

• In CUDA 6.0, NPP routines are fully validated on the ARM platform. Please report any functional errors via the CUDA registered developer website.
• In CUDA 6.0, NPP has improved the stability of JPEG-codec-related functions in multi-threaded and multi-GPU applications. This includes the DCT and quantization primitives.
• Fixed a bug in the progressive refine encoding mode.
• Optimized the performance of the CPU Huffman decoder code.

• (Mac) When Clang is used as the host compiler, 32-bit target compilation on OS X is not supported. This is because the Clang compiler doesn't support the -malign-double switch that the NVCC compiler needs to properly align double-precision structure fields when compiling for a 32-bit target (GCC does support this switch). Note that GCC is the default host compiler used by NVCC on OS X 10.8 and Clang is the default on OS X 10.9.
• The NVCC compiler doesn't accept Unicode characters in any filename or path provided as a command-line parameter.
• A CUDA program may not compile correctly if a type or typedefT is private to a class or a structure, and at least one of the following is satisfied:
  • T is a parameter type for a __global__ function.
  • T is an argument type for a template instantiation of a __global__ function.
  This restriction will be fixed in a future release.
• (Mac) The documentation surrounding the use of the flag -malign-double suggests it be used to make the struct size the same between host and device code. We know now that this flag causes problems with other host libraries. The CUDA documentation will be updated to reflect this. The work around for this issue is to manually add padding so that the structs between the host compiler and CUDA are consistent.

• (Windows) Using the mouse wheel button to scroll does not work within the Visual Profiler on Windows.
• The Visual Profiler cannot correctly import profiler data generated by nvprof when the option --kernels <kernel filter> is used. Visual Profiler reports a warning, "Some collected events or source-level results could not be associated with the session timeline." One workaround is to use the nvprof option --kernels :::1 to profile the first invocation for all kernels.
• (Mac) Visual Profiler events and metrics do not work correctly on OS X 10.8.5. Please upgrade to OS X 10.9.2 to use Visual Profiler events and metrics.