GTC 2021 Keynote with NVIDIA CEO Jensen Huang

NVIDIA CEO Jensen announced NVIDIA’s first data center CPU, Grace, named after Grace Hopper, a U.S. Navy rear admiral and computer programming pioneer. Grace is a highly specialized processor targeting largest data intensive HPC and AI applications as the training of next-generation natural-language processing models that have more than one trillion parameters.

Further accelerating the infrastructure upon which hyperscale data centers, workstations, and supercomputers are built, Huang announced the NVIDIA BlueField-3 DPU.

The next-generation data processing unit will deliver the most powerful software-defined networking, storage and cybersecurity acceleration capabilities.

Where BlueField-2 offloaded the equivalent of 30 CPU cores, it would take 300 CPU cores to secure, offload, and accelerate network traffic at 400 Gbps as BlueField-3— a 10x leap in performance, Huang explained.

CUDA driver version is insufficient for CUDA runtime version

When you do a “/usr/local/cuda-10.1/extras/demo_suite/deviceQuery”. You might get the errors seemed above

[root@node1 ~]# /usr/local/cuda-10.1/extras/demo_suite/deviceQuery
/usr/local/cuda-10.1/extras/demo_suite/deviceQuery Starting...

CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount returned 35
-> CUDA driver version is insufficient for CUDA runtime version
Result = FAIL

The Issue may cause some confusion. It is not your libraries. But the it is the Power Setting at the BIOS. Most Servers are configured to be balanced. But for GPGPU, you need to put Power to “Maximum Performance”. For example, for HPE Server, you should put “Static High Performance Mode”

Compiling Gromacs-2019.3 with Intel MKL and CUDA


GCC-6.5 Compilers and associates libraries

Intel Compilers and Prerequisites

source /usr/local/intel/2018u3/bin/ intel64
source /usr/local/intel/2018u3/impi/2018.3.222/bin64/ intel64
source /usr/local/intel/2018u3/mkl/bin/ intel64
source /usr/local/intel/2018u3/parallel_studio_xe_2018/bin/ intel64

Create a setup file


Put the following into the

CC=mpicc CXX=mpicxx cmake .. -DCMAKE_C_COMPILER=mpicc -DCMAKE_CXX_COMPILER=mpicxx -DGMX_MPI=on -DGMX_FFT_LIBRARY=mkl
-DCMAKE_INSTALL_PREFIX=/usr/local/gromacs-2019.3_intel18_mkl_cuda10.1 -DREGRESSIONTEST_DOWNLOAD=ON
-DCMAKE_C_FLAGS:STRING="-cc=icc -O3 -xHost -ip"
-DCMAKE_CXX_FLAGS:STRING="-cxx=icpc -O3 -xHost -ip -I/usr/local/intel/2018u3/compilers_and_libraries_2018.3.222/linux/mpi/intel64/include/" 
make install

Testing and Verification

$ source /your/installation/prefix/here/bin/GMXRC
./ all -np 2

Developing a Linux Kernel Module using GPUDirect RDMA

Taken from Developing a Linux Kernel Module using GPUDirect RDMA

1.0 Overview

GPUDirect RDMA is a technology introduced in Kepler-class GPUs and CUDA 5.0 that enables a direct path for data exchange between the GPU and a third-party peer device using standard features of PCI Express. Examples of third-party devices are: network interfaces, video acquisition devices, storage adapters.

GPUDirect RDMA is available on both Tesla and Quadro GPUs.

A number of limitations can apply, the most important being that the two devices must share the same upstream PCI Express root complex. Some of the limitations depend on the platform used and could be lifted in current/future products.

A few straightforward changes must be made to device drivers to enable this functionality with a wide range of hardware devices. This document introduces the technology and describes the steps necessary to enable an GPUDirect RDMA connection to NVIDIA GPUs on Linux.


1.1. How GPUDirect RDMA Works

When setting up GPUDirect RDMA communication between two peers, all physical addresses are the same from the PCI Express devices’ point of view. Within this physical address space are linear windows called PCI BARs. Each device has six BAR registers at most, so it can have up to six active 32bit BAR regions. 64bit BARs consume two BAR registers. The PCI Express device issues reads and writes to a peer device’s BAR addresses in the same way that they are issued to system memory.

Traditionally, resources like BAR windows are mapped to user or kernel address space using the CPU’s MMU as memory mapped I/O (MMIO) addresses. However, because current operating systems don’t have sufficient mechanisms for exchanging MMIO regions between drivers, the NVIDIA kernel driver exports functions to perform the necessary address translations and mappings.

To add GPUDirect RDMA support to a device driver, a small amount of address mapping code within the kernel driver must be modified. This code typically resides near existing calls to get_user_pages().

The APIs and control flow involved with GPUDirect RDMA are very similar to those used with standard DMA transfers.


Read more at:

Compiling OpenMPI-1.8.8 with Intel Compiler and CUDA

Configuration parameters for Compiling OpenMPI-1.8.8 with Intel Compiler and CUDA can be found here.

# ./configure --prefix=/usr/local/openmpi-1.8.8-gpu_intel-15.0.7 CC=icc CXX=icpc F77=ifort FC=ifort --with-devel-headers --enable-binaries --with-cuda=/usr/local/cuda/
# make -j 16
# make all


  1. Compiling OpenMPI 1.6.5 with Intel 12.1.5 on CentOS 6
  2. Building OpenMPI Libraries for 64-bit integers