NVIDIA opened up ISC 2018 with a number of new announcements, although none of them tied to a new GPU. On the pre-briefing call, NVIDIA touted a few of its major supercomputer wins, but also some of the software behind the wins.
NVIDIA’s ISC 2018 Top500 Roster
Perhaps the biggest announcement is ORNL’s Summit supercomputer that utilizes over 27,000 NVIDIA Volta GPUs, IBM Power9 CPUs, and Mellanox EDR Infiniband. We covered this new system in our US DoE Announces Summit with IBM Power and NVIDIA Tesla piece recently. Beyond Summit, the LLNL Sierra supercomputer is another machine from the US DOE’s CORAL program that will be #3 on the June 2018 Top500 list. NVIDIA’s GPUs power five of the top eight on this list.
As a percentage of new FLOPS on the Top500 list NVIDIA is showing that its overall share is growing. There is a question regarding how well traditional CPU and GPU architectures will scale towards an exascale future slated for around 2021. For now, NVIDIA is gaining share.
The HPC community, by adopting the NVIDIA Tesla V100, also gains access to an accelerator that is designed for matrix multiply. That means that Tesla V100 machines have the prowess to enable AI/ deep learning workloads alongside and to augment traditional HPC workloads.
This is an area where two architectures significantly converge to the point that the NVIDIA Tesla V100 offers supercomputer operators a lot of value.
Other News from NVIDIA at ISC 2018
The NGC is adding new HPC containers. These containers make it easy to deploy HPC applications and are maintained by NVIDIA. You can read about that in our piece: NVIDIA GPU Cloud is one important step in democratizing deep learning.
The other highlight from NVIDIA was that its software stack is maturing as well.
This slide essentially shows that NVIDIA has continued to develop hardware and software and that those two combine to improve performance. Of course, these are not raw hardware gains but are hardware and software which is a concept NVIDIA is keen to push.
The NVIDIA Tesla V100 was announced over a year ago so we are getting fairly late in the product cycle at this point. HPC teams know what the NVIDIA Tesla V100 is capable of, and can choose to deploy it or something more exotic. Intel by moving away from the Knights Landing/ Knights Mill architecture on the path to exascale has ceded ISC 2018 to NVIDIA. We are told there is something coming for the exascale future but we need to stay tuned for that.