UMD/Nvidia GPU Summit October 27-19, 2014 Fran LoPresti, Deputy - - PowerPoint PPT Presentation

umd nvidia gpu summit
SMART_READER_LITE
LIVE PREVIEW

UMD/Nvidia GPU Summit October 27-19, 2014 Fran LoPresti, Deputy - - PowerPoint PPT Presentation

UMD/Nvidia GPU Summit October 27-19, 2014 Fran LoPresti, Deputy CIO, Cyberinfrastructure and Research IT DIVISION OF INFORMATION TECHNOLOGY Cyberinfrastructure Center in the University Research Park DIVISION OF INFORMATION TECHNOLOGY


slide-1
SLIDE 1

DIVISION OF INFORMATION TECHNOLOGY

UMD/Nvidia GPU Summit

October 27-19, 2014

Fran LoPresti, Deputy CIO, Cyberinfrastructure and Research IT

slide-2
SLIDE 2

DIVISION OF INFORMATION TECHNOLOGY

Cyberinfrastructure Center in the University Research Park

slide-3
SLIDE 3

DIVISION OF INFORMATION TECHNOLOGY

Deepthought2

slide-4
SLIDE 4

DIVISION OF INFORMATION TECHNOLOGY

Deepthought2 Configuration

  • 300 Teraflops peak (theoretical) performance
  • 62 DELL C8000 chassis containing a mix of C8220 CPU-only nodes and C8220X CPU/GPU nodes
  • 416 CPU-only nodes (dual Intel E5-2680v2 CPUs at 2.8GHz, 10 cores per CPU)
  • 9200 total Xeon cores
  • 40 CPU/GPU nodes (2 nVidia K20 GPUs per node)
  • Each K20 GPU has 5GB RAM, 2496 CUDA cores, 1.17TF double precision floating point

performance

  • 200,000 total nVidia cores
  • 128GB RAM per node (6.4GB per core)
  • 4 large-memory nodes (quad Intel E5-4640 CPUs at 2.2GHz, 10 cores per CPU) with 1TB RAM each
  • 1 Petabyte of Lustre storage (20GB/second aggregate throughput)
  • FDR (56 Gigabit) Infiniband networking, configured as 2:1 blocking
  • Power consumption 250kW, 12 racks
slide-5
SLIDE 5

DIVISION OF INFORMATION TECHNOLOGY

Deepthought2 on Campus

  • HPC Boot Camp
  • Lecture series in advanced computing and Visualization
  • Matching visualization hardware grants
  • Curriculum use of Deepthought2
  • Free Deepthought2 allocations based on merit of

proposals

slide-6
SLIDE 6

DIVISION OF INFORMATION TECHNOLOGY

Enjoy the Summit