Supercomputing Part 2

This is a continuation of my previous post on Supercomputing 2022.

To say, the supercomputing space and the Supercomputing event (held in has changed would be a massive understatement.?

Like the performance of the systems that make up the Top 500 list, the event was so much (much) bigger. The conference’s scope too has expanded. Not only did this year’s attendees come from academia/research and HPC affiliated teams in the enterprise, but I saw quite a few attendees from adjacent groups. This includes teams responsible for AI/ML and data management from life sciences, manufacturing, and more. The Supercomputing conference has become the showcase for the broader HPC eco-system and the latest in AI/ML.

For example, WEKA partner Cerebras showcased their technology behind their Andromeda system delivering 1 Exaflop of AI compute and 120 Petaflops of dense compute. They spoke about their 13.5 million-core AI supercomputer (yes, we’re talking millions). This system delivers near perfect linear scaling for the largest GPT-class language models.?

So, why am I writing about this??

First, I am excited by the potential WEKA has to help organizations with not only their storage needs for traditional HPC workflows (ex: physics research, CAD/CAE, etc), but also the emerging AI/ML workflows (ex: dealing with PBs and PBs of data used to build and reinforce models, large scale interference workloads like fraud detection, and more). It was evident that the AI/ML teams in attendance were realizing the complexity of data related challenges they would be facing with machine learning and deep learning:

1. They not only use traditional CPUs, but also rely on accelerated compute like GPUs and IPUs.?
2. They deal with different types of data at a much different scale – terabytes to petabytes per project.
3. They have to support a very diverse workflow – ingest and transform data, random read and write millions of small files in addition to sequential reads and writes, deal with metadata of thousands to millions of files, and more.
4. They deal with different clouds – their workflows span different clouds. It’s not uncommon to burst from on-prem/private clouds to public clouds to handle a surge of learning activity.?

Second, to marvel on how this conference has evolved as the perfect? place for organizations to experience all parts of the ecosystem come together.? As someone responsible for building a vibrant solutions and ecosystem at WEKA, I saw almost all of our ecosystem partners in full force–from accelerated compute and networking like AMD, Cerebras, Habana & Intel, NVIDIA, and more to ISVs dealing with AI/ML, Data Management, and Workflow such as Ansys, Altair, Run:ai, and more to our strategic cloud and system partners like AWS, Dell, Google Cloud, HPE, Hitachi Vantara, Lenovo, Oracle, Supermicro, and more. An example of this in action is our partner Run:ai, who were connecting with and successfully educating organizations to talk about AI Computing and Machine Learning initiatives including areas like ML Operations (aka MLOps) and getting the most out of their GPU investments across different clouds.?

Third, because of the role our WEKA team played in educating attendees about the state of the art and future of Supercomputing – not only on storage, but across the entire stack. WEKA hosted dozens of partners at our booth theater at the conference, who spoke about industry trends, state of the art technologies, joint solutions, and real world examples – all intended to help organizations understand how they can tackle their next generation of application workflows.?

I’ll cover the learning I took from these sessions in my next post.