Kebnekaise Expansion: A Welcome Addition for all Users

  • Posted on: 29 May 2018
  • By: bbrydsoe

Kebnekaise Expansion: A Welcome Addition for all Users

We are currently installing and verifying the expansion of Kebnekaise that we announced at the HPC2N Open House last fall. The expansion is a welcome addition to a heavily loaded Kebnekaise and will improve queueing times, particularly for GPU jobs. The expansion consists of a combination of standard Intel Xeon Gold 6132 (Skylake) CPU nodes and nodes also equipped with NVIDIA V100 Volta GPUs. This makes HPC2N one of the first accademic HPC sites in Europe to have the V100 GPUs in production.

The much faster CPUs in the expansion will improve the total performance of Kebnekaise with almost 30%. The CPU performance of Intel Skylake processors is up to 60% better compared to the broadwell CPUs in the original nodes. The NVIDIA V100 GPU cards have a significantly higher performance than the K80 GPU cards, installed in the original GPU nodes. With 640 Tensor Cores, NVIDIA V100 is the world’s first GPU to break the 100 TeraFlops barrier of deep learning performance.

We are at the moment installing optimized software packages to take advantage of these new architectures. 

More information on how to use the expansion will be available when the nodes are opened for production in the near future.

Hardware Specifications of the Expansion

  • 52 nodes with Intel Xeon Gold 6132 (Skylake-SP), dual 14 cores, 2.6 GHz. Each node has 192 GB RAM. Interconnect is Mellanox EDR (100 Gb) Infiniband.
  • 10 nodes with the same specification as above, with the addition of 2 NVIDIA V100 (Volta-based) GPU cards on each node.
  K80 (original GPU nodes) V100 (new GPU nodes)
# of CUDA cores 4992 5120
# of Tensor cores - 640
Memory 24 GB GDDR5 16 GB HBM2
Memory bandwith 480 GB/s 900 GB/s

Performance comparision of the GPUs in Kebnekaise

  K80 V100
Single-precision 8.73 TF 14 TF
Double precision 2.91 TF 7 TF
Tensor - 112 TF

See our Kebnekaise page for a full description of the system, including both old and new nodes. 

 

Updated: 2018-06-13, 13:20