HPC 2023 Predictions

CXL

CXL has been in the works for a number of years now. It is backed by a consortium of industry leaders including 英特尔 , AMD , 英伟达 and a host of others. Although version 1 of the specification was first announced in March of 2019, it is only the latest generation of CPU’s released recently (Ie Sapphire Rapids and Genoa) that offer support for it. This is partly due to delays caused by COVID-19, but there were also numerous technical hurdles which needed resolving before such a groundbreaking technology could be put in front of consumers.

So what does it do? To put it simply, CXL allows a computer to more effectively manage its resources. In todays #HPC world, peripherals attached to a server often come with their own memory and/or co-processor. A good example of this is the NVIDIA Bluefield DPU, which is in essence a fancy network card, capable of offloading the grunt work of moving packets around from the CPU, thus freeing up resources which could be used for more vital things like crunching numbers. But if your #DPU isn’t being used to its full potential, CXL can take the unused resources and allow them to be used elsewhere in the system. This opens up a world of possibilities, not just for HPC, but for anyone who wants to maximise the potential of their hardware and leave no byte unturned.

Although CXL 1.1 is only now gaining traction among manufacturers, version 2.0 and 3.0 have already been ratified by the consortium, meaning companies can start preparing now for the next generation of interconnected devices. With energy prices being what they are, it’s in everyone's best interests to utilise their systems to the fullest, and there’s no better example of that than in HPC.

Quantum

2023 could be the year that quantum computing starts to become mainstream. Quantum computing has been a niche interest for most of us for several years now. Some big players like IBM, Google, and Boeing have famously been running quantum computers for years, but any tangible gains beyond research-level breakthroughs have yet to be announced. The difference now is that people are starting to understand quantum better; what it can't do, and more importantly, what it can do.

While state-of-the-art quantum computers are a bit big to live on or under your desk, Boston are working to put small QC’s in the classroom, equipping the next generation of computer scientists with the knowledge they need to properly harness the power of this next iteration of computer science and information theory.

Currently, #quantum computers are very good at a few specific things. What people have started to realise is that the next big breakthroughs will happen when we start using quantum to complement our existing compute resources. The good news is that there are already tools on the market which allow us to do this. You can even test your own quantum theories in your browser! In 2022 Boston announced a groundbreaking partnership with Oxford Quantum Circuits (OQC) , so once you’re ready to bring your theories to the real world, you can be confident we’re here to support you on your journey.

DPU

“Data is cheap to generate, but expensive to move”

A few years ago, the above statement was vastly untrue. But as we’ve started to reach the limits of Moore’s law, #manufacturers have been working on speeding up everything surrounding the CPU, including the network. When the average network card speed was 1Gb/s, it took very little processing power to move data around. This has now changed as #network speeds have increased. NVIDIA have already released their NDR switches, capable of reaching speeds of 400Gb/s per port, and their roadmap shows 800Gb/s (XDR) and 1.6Tb/s (GDR) hardware coming in the next few years.

This is great news for anyone with a large dataset (I’m looking at YOU machine learning lot!), because in the world of true HPC, we don’t want to waste our CPU cycles sending and receiving data when they could be put to better use, discovering new gene therapies or predicting hurricanes!

Enter the DPU. The data processing unit, also known as a Smart NIC, is a device specifically designed to move data for you. While not particularly new (NVIDIA are already on to their 3rd generation DPU), I believe that the proliferation of machine learning, large datasets, and ever-increasing network speeds, will take them from a ‘nice to have’ to a ‘must have’, for anyone wishing to stay ahead of the HPC game in 2023 and beyond.