Two RCC staffers attended SupercomputingAsia (SCA) and HPCAsia 2026, held in Osaka, Japan from 26–29 January, alongside more than 2,500 participants from academia, industry and government.
Read reports below from both RCC Director Jake Carroll and Research Systems Projects and Delivery Manager Sarah Walters (jump to Sarah's report) about their experience at the landmark event for the Asia-Pacific supercomputing community.
This year’s conference theme — “Everything with HPC: AI, Cloud, QC and the Future Society” — reflected a major shift: advanced computing is no longer siloed. Instead, AI, high-performance computing (HPC), and quantum technologies are rapidly converging into a unified ecosystem.
The conference is co-organised by HPC centres from Australia, Japan, Singapore, and Thailand, and has become a key networking and collaboration platform for the regional and global supercomputing community.
Increasing convergence across computing technologies
By RCC Director Jake Carroll
I observed the following key themes at SCA/HPCAsia 2026:
1. The convergent ecosystem is here:
The primary takeaway from Osaka was the convergence of High-Performance Computing (HPC), Artificial Intelligence (AI), and Quantum Computing (QC).
We are seeing a definitive shift from traditional siloed computing to a unified ‘compute–intelligence’ ecosystem.
- AI as a Scientist: Keynotes, particularly from Hiroaki Kitano (Sony CSL), emphasised AI not just as a tool for data analysis but as an autonomous agent in the scientific discovery process (AI for Science). We saw multiple real-life demonstrations of AI guiding lab operations, search, functionality and experimentation both in the “dry lab” and “wet lab.”
- Sovereign AI Infrastructure: A major theme was the push for "Sovereign AI” nations (led by Japan's $6.3b initiative) building domestic foundation models and the specific hardware required to ensure data autonomy local to the institutions. Super critical: institutions are building their own bridges. Institutions that have the strategy settings right are setting their own pace. Vendor-led is decreasing in the institutions because the very things the vendors are leading are not financially tenable for us. To put it in real terms: we cannot afford AI-hyperscale training systems and we likely never will. Instead, our sector will likely model integrated approaches such as LUMI’s own EU-AI factory built as a cohesive part of its HPC. This reflects the true convergence of AI and HPC in research that is more aligned to the way our researchers and their workflows operate. We aren’t trying to build an “AI factory in a box”.
- HPC+QC Integration: The transition from "experimental quantum" to "integrated quantum" was evident, focusing on hybrid workflows where quantum processing units (QPUs) act as accelerators within classical HPC clusters.
"Institutions are setting their own pace. The 'AI factory' model is not financially right, nor is it workflow-optimised for most universities."
2. Significant trends:
- Sustainability & "Green AI": With power demands soaring, there was a heavy trend toward energy-aware scheduling. RIKEN and the LRZ (Germany) highlighted collaborations on "Energy Efficient HPC State of the Practice," emphasising carbon-aware computing.
- Arm Architecture Dominance: The "Post-Fugaku" era in Japan is doubling down on Arm. The FUJITSU-MONAKA processor and its successor, MONAKA-X, are being positioned as the energy-efficient backbone for both HPC and enterprise AI.
- The rise of alternative interconnects such as Ultra Ethernet. Professor Torsten Hoffler from ETH Zurich gave the opening keynote on the topic. It is now very clear that alternatives are on the way to challenge the market leader's dominance in interconnects. UQ will likely be an early tester of these options with some of our hardware collaborators and industry partners.
“It is now very clear that alternatives are on the way to challenge the market leader's dominance in interconnects. UQ will likely be an early tester of these options.”
3. Research impact and breakthroughs:
Research at the 2026 conference showed a maturing of AI-HPC integration:
- LLM-Driven Code Generation: One of the most discussed papers was "ChatMPI," which explores using Large Language Models to automatically generate and optimise MPI (Message Passing Interface) code, potentially solving the long-standing "productivity gap" in parallel programming.
- Climate Resilience: Building on the ORBIT-2 research, teams demonstrated exascale vision foundation models that can perform planetary-scale climate downscaling at sub-kilometre resolution. This is a vital tool for regional disaster planning.
- Mixed-Precision Innovation: The RAPTOR tool was highlighted for its ability to profile scientific applications to see where they can tolerate lower precision (like FP8 or FP4) without losing accuracy, significantly boosting efficiency on modern AI-centric hardware.
- There was a lot of contentious debate about just how low precision could be before accuracy and validation fell apart. Some GPU makers argue they can restore precision later and emulate as found in initiatives such as the Ozaki Scheme. The research community is cautious at this stage. Some GPU makers are hedging their bets, offering GPUs that offer balanced performance for both very low precision (AI training/inference) and traditional accelerated scientific codes for FP32 and FP64. Time will tell who “wins” – but it may take years to know.
Practical takeaways for researchers
By Sarah Walters, RCC Research Systems Projects and Delivery Manager
For researchers, the SCA/HPCAsia 2026 conference reinforced several practical considerations:
- AI: Do you want to understand, or do you want a result? For industry, they often just need an answer that is "good enough", and AI gets them there faster. For research, that "good enough" answer is an accelerator – it gets you to the point where you can develop more accurate simulations to reach an answer you can understand and explain.
- FP64 precision has been the default for years – "if in doubt, allocate a double" – but if we can be more efficient with selective use of double precision we can work more efficiently with lower-precision architectures, and that translates to real performance gains on GPU. Similar discussions happened back in the 70s when high precision was hideously expensive, and we have circled back around again.
- The industry is shifting towards inferencing and fine-tuning rather than training. In academia we still need foundational models for particular workloads, but there is a lot that can be done with the models we already have.
- On digital sovereignty, ask "what is your threat?" Is it the security of your data? Is it your ability to operate at all? If there is a conflict affecting US infrastructure, does your cloud potentially fall over entirely? Does that compute get co-opted for other priorities, leaving Australian institutions struggling to operate?
The next year is likely to be evolutionary – we’re not expecting a revolutionary leap like the introduction of ChatGPT this year.
Getting your data right is critical – you need to know whether the data is applicable before feeding it into an AI model. If you do not know, you still need more traditional simulations.
- Everybody is talking about KV cache, which is how to store data between requests to an AI so that you do not have to inference repeatedly in extended conversations. Getting storage for this right is a hard problem.
- There is widespread anxiety about the growing cost of compute hardware.
- Quantum computing continues to progress, although it is not yet ready for broad deployment in most university environments. We saw this first-hand at the RIKEN facility, where an operating quantum computer required extensive environmental isolation just to maintain accuracy — a vivid illustration of how far from everyday deployment it remains.
