new features in CUDA 4.0
7 views
Skip to first unread message
Lung-Sheng Chien
Mar 2, 2011, 3:44:44 AM3/2/11
to pfacForum
CUDA 4.0 is going to release on March 4.
NVIDIA anounces features in http://gpgpu.org/2011/03/01/cuda-4-0-release#more-3309,
I think some of them are related to PFAC library.
■Thrust C++ Template Performance Primitives Libraries
we use Thrust to do prefix-sum, we may link Thrust to incoming CUDA
4.0
■Multi-GPU Sharing by Single CPU Thread - A single CPU host thread can
access all GPUs in a system. Developers can easily coordinate work
across multiple GPUs for tasks such as "halo" exchange in applications
This is a good feature that we don't need to use OpenMP, or
equivalently, one PFAC context can bind to several GPUs.
■NVIDIA GPUDirect(tm) 2.0 Technology - Offers support for peer-to-peer
communication among GPUs within a single server or workstation. This
enables easier and faster multi-GPU programming and application
performance.
this is useful when one PFAC context wants to utilize more than one
GPU.
■MPI Integration with CUDA Applications - Modified MPI implementations
automatically move data from and to the GPU memory over Infiniband
when an application does an MPI send or receive call.
We don't provide MPI + PFAC library so far. We can consider a HUGE
app, for example DNA analysis and then apply MPI to deal with a large
pattern set.
NVIDIA anounces features in http://gpgpu.org/2011/03/01/cuda-4-0-release#more-3309,
I think some of them are related to PFAC library.
■Thrust C++ Template Performance Primitives Libraries
we use Thrust to do prefix-sum, we may link Thrust to incoming CUDA
4.0
■Multi-GPU Sharing by Single CPU Thread - A single CPU host thread can
access all GPUs in a system. Developers can easily coordinate work
across multiple GPUs for tasks such as "halo" exchange in applications
This is a good feature that we don't need to use OpenMP, or
equivalently, one PFAC context can bind to several GPUs.
■NVIDIA GPUDirect(tm) 2.0 Technology - Offers support for peer-to-peer
communication among GPUs within a single server or workstation. This
enables easier and faster multi-GPU programming and application
performance.
this is useful when one PFAC context wants to utilize more than one
GPU.
■MPI Integration with CUDA Applications - Modified MPI implementations
automatically move data from and to the GPU memory over Infiniband
when an application does an MPI send or receive call.
We don't provide MPI + PFAC library so far. We can consider a HUGE
app, for example DNA analysis and then apply MPI to deal with a large
pattern set.
Lung-Sheng Chien
Mar 5, 2011, 9:19:16 PM3/5/11
to pfacForum
communication of dual GPU can be behind PCIE
http://forums.nvidia.com/index.php?showtopic=194235&pid=1202365&start=&st=#entry1202365
it is great but we need to test it.
On 3月2日, 下午4時44分, Lung-Sheng Chien <lungshengch...@gmail.com> wrote:
> CUDA 4.0 is going to release on March 4.
Reply all
Reply to author
Forward
0 new messages