beta CUDA 1.1

Today we’ll be posting a beta release of CUDA 1.1 initially on the developer login site.

We appreciate that more of you would like to be on the registered developer program, but we’re sorry we’re so backlogged getting through all the requests and required non-disclosure agreements.

We expect to post at least the Windows beta release for everyone in just a few days as there will then be a public Windows display driver available that supports CUDA. *this is a little different from my initial post: Linux standard display driver still has a little more time to go so the CUDA beta release will stay in the login/NDA area for a little longer. Sorry about that. As CUDA is part of the standard driver this becomes much easier going forward. We’d like to be through this beta in just a few weeks anyway which should cooincide with the Linux display driver release.

CUDA 1.1 has some new features and sample programs in the SDK including:
• NVIDIA standard display drivers will now include the CUDA dll. This is just starting to roll out with all new display drivers, so for now we’ve still included drivers in this beta release.
• Having moved into the standard display driver caused us to rename cuda.dll to nvcuda.dll. The other dll’s included with the CUDA toolkit account now rely on this. Even with this change, most binaries built with 1.0 work fine just running on 1.0. We’ve seen one case where a product specifically checked for the existence of “cuda.dll” and that was easy to work around.
• It’s fine for you to redistribute the dll’s from the CUDA toolkit and SDK so your customers can just use a (newer) standard NVIDIA driver rather than having to install a CUDA-specific one. We just require that when you install our dll’s with your product that you do not install them in a common area - but rather in your own subdirectory. And do not redistribute nvcuda.dll.
• Support for WinXP64
• More signals in profiler
• Graphics interoperability with CUDA across Multiple GPUs is enabled. Please note that this is not as efficient as it could be. Data is transferred back through the host in this release.
• New event management features particularly useful for the new async and stream management
• Asynchronous execution: new memory copy calls with Async suffix return control to the application without waiting for completion (see new asyncAPI SDK example and 1.1 Programming Guide)
• Alpha feature for supporting overlapped memory copies concurrent with kernel execution using new stream management interface. There are specific limitations to this and it will only do overlap on 1.1 architectures (g84/g86/g92), though it will revert to serial operation on 1.0 architectures. See new simpleStreams SDK example and 1.1 Programming Guide). We’re still tuning this, so this capability will remain a beta feature even as CUDA 1.1 becomes a production release.
• There are now 50 source code examples in the SDK.

Features that will be in upcoming releases:
• 3D Texture support
• Vista
• Mac OS X (we’re demonstrating this now and should see beta in January)
• Debugger on GPU (demonstrating now)
• Visual profiler (demonstrating now)
• Optimized peer-to-peer transfers
• Support for double-precision hardware

This beta release does not include support for all Linux distos - only Red Hat 4 & 5 (32 and 64bit). We’ll certainly have the same set we support on CUDA 1.0 when we release this in production.

There are now about 48 Million CUDA-capable NVIDIA 8-series GPUs shipped so thanks everyone for developing CUDA applications that are making the owners of those GPUs even happier! And we really appreciate your bug reports and feedback during this beta release!!