CUDA 6 is Out

[flowrider]

CUDA 6.0.37 is now out. For OS 10.9 Mavericks - or so it says. I just downloaded it.

http://www.nvidia.com/object/macosx-cuda-6.0.37-driver.html

Lou

 

[crjackson2134]

Put it to the test and let us know the outcome.

 

[m4v3r1ck]

Thanks for the heads-up Lou! Let's see what the review say...

 

[v0n]

Updated yesterday, so far about 50% slower than Cuda 5.5 with web drivers. Quick test with cudaminer confirms about 40-45% drop.

 

[Lone Deranger]

Updated yesterday, so far about 50% slower than Cuda 5.5 with web drivers. Quick test with cudaminer confirms about 40-45% drop.

Slower???!!

 

[Earl Urly]

Which graphics card, v0n? Maybe it's optimized for Kepler class GPUs?

 

[v0n]

GTX 660 and GTX 470.

Using cudaminer on GTX 660 in benchmark mode:
196kH/s with CUDA 5.5.47 and stock Maverics drivers
212 kH/s with CUDA 5.5.47 and Web drivers 331.01.01f02
Struggles to pass 141kH/s with CUDA 6.0.37 and drivers 331.01.01f02.

Installing 5.5.47 back pushes benchmark up to 169 kH/s but won't go back to figures from before installation, something somewhere has been overwritten, maybe reboot required.

 

[layzarc]

I just installed CUDA 6 on my early rMBP (Geforce 650M) and CUDA-Z's single precision GFlops go from 300 to 260. What's interesting is that rebooting to my Windows 7 BOOTCAMP partition I get 400 Gflops (the same as with CUDA 5.5). Is NVIDIA intentionally hamstringing the Mac since Apple dropped them in the nMP?

 

[flowrider]

Finally saw Nvidia's release Notes:

Introduced support for the Maxwell architecture (sm_50). More information on Maxwell can be found here: https://developer.nvidia.com/maxwell-compute- architecture. Although the CUDA Toolkit supports developing applications targeted to sm_50, the driver bundled with the CUDA installer does not. Users will need to obtain a driver compatible with the Maxwell architecture from http:// http://www.nvidia.com/drivers.

Unified Memory is a new feature enabling a type of memory that can be accessed by both the CPU and GPU without explicit copying between the two. This is called "managed memory" in the software APIs. Unified Memory is automatically migrated to the physical memory attached to the processor that is accessing it. This migration provides high performance access from either processor, unlike "zero- copy" memory where all accesses are out of CPU system memory.

Added a standalone header library for calculating occupancy (the library is not dependent on the CUDA Runtime or CUDA Driver APIs). The header library provides a programmatic interface for the occupancy calculations previously contained in the CUDA Occupancy Calculator. This library is currently in beta status. The interface and implementation are subject to change.

The Dynamic Parallelism runtime should no longer generate a cudaErrorLaunchPendingCountExceeded error when the number of
pending launches exceeds cudaLimitDevRuntimePendingLaunchCount. Instead, the runtime automatically extends the pending launch buffer beyond cudaLimitDevRuntimePendingLaunchCount, albeit with a performance penalty.
Support for the following Linux distributions has been added as of CUDA 6.0: Fedora 19, Ubuntu 13.04, CentOS 5.5+, CentOS 6.4, OpenSUSE 12.3, SLES SP11, and NVIDIA Linux For Tegra (L4T) 19.1.

Support for the ICC Compiler has been upgraded to version 13.1.
Support for the Windows Server 2012 R2 operating system has been added as of CUDA 6.0.

RDMA (remote direct memory access) for GPUDirect is now supported for applications running under MPS (Multi-Process Service).

CUDA Inter-Process Communication (IPC) is now supported for applications running under MPS. CUDA IPC event and memory handles can be exported and opened by the MPS clients of a single MPS server.

Applications running under MPS can now use assert() in their kernels. When an assert is triggered, all work submitted by MPS clients will be stalled until the assert is handled. The MPS client that triggered the assert will exit, but will not interfere with other running MPS clients.

Previously, a wide variety of errors were reported by an "Unspecified
Launch Failure (ULF)" message or by the corresponding error codes CUDA_ERROR_LAUNCH_FAILED and cudaErrorLaunchFailed. The CUDA driver now supports enhanced error reporting by providing richer error messages when exceptions occur. This will help developers determine the causes of application faults without the need of additional tools.

Lou