Recent visits to Europe have been significantly influenced by the FIFA World Cup football (‘soccer’ to my American friends) in Russia. Every bar and restaurant has had additional large screen TVs showing the games and important meetings were carefully scheduled around the tournament schedule. Against this backdrop you would expect Europe’s largest HPC conference and trade show - International Supercomputing Conference 2018 “ISC18” - to struggle for attention. However, nothing could be further from the truth.
Enterprising exhibitors leveraged the tournament schedule with custom World Cup events like our Icelandic colleagues Opin Kerfi who hosted a party to watch what unfortunately turned out to be Iceland’s swansong from the tournament, despite Gylfi Sigurdsson's penalty below (image thanks to Vadim Ghirda/AP):
Additionally, and despite the football, and especially Germany being knocked out by South Korea on the final day of the exhibition, several HPC announcements and new products got the full attention of the ISC community.
Most conferences include the obligatory keynote sessions and ISC is no different. Following a fascinating view of HPC at CERN - home of the Large Hadron Collider (LHC), the world’s largest and most powerful particle accelerator, the Top500 Supercomputer rankings were updated for the 25th year. Unusually 4 of the top 5 places changed this year resulting in the following Top 5:
A few things struck me about the new rankings: the lack of traditional heavy compute iron, the use of accelerators like GPUs and the impressive showing of the new IBM Power 9 servers coupled with their NVIDIA GPUs taking both the first and third position.
I noticed two other major changes in the HPC world order. Firstly, just over a year ago I visited The Stephen Hawking Centre for Theoretical Cosmology (CTC) and teased the exceptional physicists there about their large black-hole research models still being coded in Fortran. It appears that academic physics kudos are not gained for rewriting software and none of them had any intention of rewriting any legacy Fortran. As you would expect from such intellectual giants, they found a solution! At ISC18 I encountered a few computer science PhD candidates, who were rewriting some large legacy science applications in C++ with GPU augmentation and getting their PhD for their efforts. This trend will continue to drive the attractiveness of the CPU server with GPU augmentation for future scientific supercomputers. Secondly, NEC are one of the few, if not the only vendor, selling vector computers. It will be interesting to see if they get increasing or decreasing market traction with that technology versus the CPU/GPU momentum. It’s hard to see such heavy vector iron being offered by the cloud computing community.
Previously I had thought of the AI development and HPC domains as quite separate but related. Each has their own lexicon and conference circuit. At ISC18 I encountered traditional HPC applications using AI Deep Neural Network (DNN) subsystems to accelerate convergence on an answer while using less compute resources. Consider training a DNN on a constrained subset of computational fluid dynamics (CFD) problems, say an aircraft wingtip design, then using inference of that trained DNN rather than the traditional differential equation solvers in the likes of Ansys Fluent to create a new design. The likely short-term consequent of this is even more demand for GPUs, competing against regular AI and the crypto currency miner communities. Over the longer-term AI DNN techniques will become a larger and larger part of general HPC.
A whole new set of HPC performance metrics will be needed to compare the traditional Linpack double precision floating point benchmark against something more appropriate for DNN training. The next few years will be very exciting for both HPC and AI fans.
Sometimes the dropped toast lands butter side up – or at least that’s how it felt when a Director of IT for a Fortune 15 company approached our booth with his bare metal cloud shopping list. His very demanding engineering team had given him a very specific list of HPC requirements and as we described the attributes of our hpcDIRECT product, we covered every one of them. It’s only natural that the large-scale compute clouds must tune their overall cloud infrastructure and resulting cost of goods (COGs) to match their typical customer profile with a competitive service. Providing HPC specific technology for generic webpage applications, e-mail servers or legacy database customers is just not viable in a competitive market. Verne Global's TrueHPC solution with 100% Infiniband connectivity, latest generation servers, high performance storage, optional GPU augmentation and all clustered in our HPC optimised campus in Iceland really hit the spot.
Iceland did a fabulous job at the World Cup out performing countries with populations hundreds or thousands of times larger. Exploit that same type of tenacity for your HPC/AI DNN training applications and benefit from the free-air cooling and 100% green energy, whether hosted HPC or using our very fit for purpose hpcDIRECT bare metal cloud.
Next stop on the HPC trade-show marathon for myself will be the gigantic SC18 - I look forward to seeing you and the supercomputing community in Dallas in November.