The NVIDIA GTC2018 Conference was held March 26-29th at the San Jose Convention Center. This is the 7th year that CertaSIM attended the conference and participated in the technical session. Dr. Mindle, CertaSIM’s Director of Sales & Marketing presented a paper entitled “Multi GPU Parallel Processing with NVLINK”. The presentation is available from CertaSIM and NVIDIA also records the audio from the presentation and makes it available to hear online. GPU technology is at the heart of the IMPETUS Afea Solver® massively parallel processing capabilities and so keeping up with the current GPU hardware is a must for CertaSIM.
The 10th Ground Vehicle Systems Engineering And Technology Symposium (GVSETS) was held August 7-9, 2018 at the Suburban Collection Showplace in Novi Michigan.
The conference included participants from DoD, TACOM, Research, Development and Engineering Command (RDECOM), Program Executive Offices (PEOs) for the Army and the Marine Corps for ground vehicles as well as leaders from industry and academia. The trade show has grown considerably over the last 4 years and provides a great venue to meet many of the companies that work with the US Army with regard to ground vehicle development. The overriding tone of the conference is the mission to design and build vehicles to better protect the “Warfighter”. One of the highlights of the non-technical sessions is the panel of warfighters which brings together combat experienced warfighters to discuss the pluses and minus of the current technology and what they need to better carry out their mission.
Ciara Technologies are experts in building overclocked professional workstations. They recently released the KRONOS 840-G4 which includes the latest Intel i9 Processors. The i9 processors provide larger core count, up to 18 cores and the Ciara machines provide machines that run up to 4.6 GHz for all processors all the time, where the base clock speed is 3.3 GHz. The Intel Turbo boost only operates for short durations to maximize performance and it only applies to a limited amount of the cores which means if your application uses too many cores the system defaults back to the base clock speed. Overclocking is quite different, all cores run at the higher frequency all the time which also includes overclocking the memory as well. The KRONOS workstations also include the latest liquid cooling technology for the CPU that insures a reliable system. The combination of GPU Technology for massively parallel processing and high performance CPU processing results in a workstation that rivals cluster based solutions.
GPU computing continues to improve every time NVIDIA releases a new generation of processors. At GTC2018 held at the San Jose Convention center this past March, the CEO of NVIDIA, Jensen Huang, introduced the latest Professional Workstation GPU, the Quadro GV100 powered by the NVIDIA Volta. With 32GB of HBM2 memory, 5120 CUDA Parallel-Processing Cores and 640 Tensor Cores it leapfrogs the previous version, the GP100 that had 3584 CUDA Cores and 16GB of HBM2 memory and no Tensor Cores. To quote the NVIDIA Website:
“The NVIDIA Quadro® GV100 is reinventing the workstation to meet the demands of next-generation ray tracing, AI, simulation, and VR enhanced workflows. It’s powered by NVIDIA Volta, delivering the extreme memory capacity, scalability, and performance that designers, architects, and scientists need to create, build, and solve the impossible.”
CertaSIM presented benchmark results at GTC2018 for a Hypervelocity Impact Scenario to compare performance of the GV100 with the GP100. The models were run with the IMPETUS Afea γSPH Solver which is a full continuum particle model that takes full advantage of GPU Technology. There were 4 resolution models tested, an 800K particle model, 6.5 million particle model, 10 million particle model and 33 million particle model. The GP100 and GV100 are NVLINK capable which means the communication between 2 GPUs is a direct connection that bypasses the CPU to improve performance. The benchmark tests consisted of single GPU, dual GPU and dual GPU with NVLINK.
To compare with a cluster based solution, a similar 800K model took a whopping 20min on a 36 core cluster. Certainly no match for GPU massively parallel processing.