Prepare for a Diversity of Compute Architectures: 2026 ISC Keynote
Get ready for a diversity of architectures in HPC spanning CPUs, AI accelerators, and quantum, as the AI boom has triggered an insatiable demand for computing, Technical University of Munich Professor Martin Schulz said during his opening keynote at the 2026 International Supercomputing Conference in Hamburg, Germany.
“We have demand that’s growing faster and faster than we can keep up with our traditional scaling mechanisms,” Schulz said during his keynote session following a short introduction by 2026 ISC chair Program Chair Rosa Badia. “AI has really shifted the center of gravity.”
Traditional modeling and simulation–the bread and butter of what runs on HPC systems–occupies a big chunk of the workload running on supercomputers, and there’s also high-performance data analytics (HPDA), Schulz said. But it’s really AI and agentic workloads that are driving this need to develop converged computing architectures that can handle this new demand.
“They’re really coming together as one, sometimes forced, sometimes not forced, but nevertheless coming together and working together,” he said.
Professor Martin Shulz delivered the opening keynote at ISC 2026
Heterogeneity is nothing new in today’s HPC data centers. It already exists in most of the systems, thanks in large part to the GPUs that accelerate traditional mod/sim and emerging AI workloads, Schulz said. But the level of diversity is just going to increase going forward, he said.
“There is not a fundamental change in paradigm. There are still von Neumann-based sequential execution streams,” the professor said. “We have still these limits that we’re hitting in HPC. ..It’s kind of amazing where we are right now, considering what we said five, 10 years ago where we could never go exascale. And now we are here, and far beyond that, and it still works. So we know how to push things to the limits, but at some point, they will also get slower and slower, perhaps even stop.”
That slowing in scaling will drive architects to experiment with new systems and new processors. While we will still rely on general-purpose computing as well as GPU-accelerated systems, we need will to find something else to keep driving the scale, as workloads demand.
“Is the GPU the savior for everything? Or do we need something else?” Schulz asked. “I think we’re going to see more and more alternatives–not replacing GPUs, but augmenting the environment there as well. So we are going to see more and more technologies come up that augment our portfolio. We’re going from one single accelerator to a portfolio of accelerators, a portfolio of compute.”
We’ll have a plethora of accelerators, Schulz said. Some companies will rely on ASICs, while others look to FPGAs. Some will depend on GPUs. “In any case, massive parallel compute close to that, to actually work with the data coming out of the CPUs,” will be needed, he said.
Photonics, neuromorphic processors, and DNA storage all will play a role in HPC and AI’s future
The challenge will be making this portfolio of accelerators work harmoniously, both at a hardware and at a software level. There will be a need to mask this complexity to the end user, as exposing them to the full brunt of technological diversity “would be too painful,” Schulz said.
“If you have a large compute, you need to identify the right kernel for the right compute, for the right place, for the right workflow. And this is a huge optimization problem by itself,” Shulz said. “And that’s why we need the right software for to make this work in this heterogeneous environment.”
Another accelerator for HPC is quantum. Schulz described some of the work he’s doing with Munich Quantum Valley, a consortium of universities, research institutions, and quantum startups in the area to develop quantum systems. The goal of Munich Quantum Valley is to develop different quantum modalities and bridge it with existing HPC systems, Shulz said.
“We really believe in that quantum computers will not replace HPC, because there’s so much else to do,” he said. “But they really are accelerators. And as part of the accelerator, they are HPC already. They’re part of our high performance computing portfolio.”
Another way to accelerate workloads on HPC systems is neuromorphic computing. Shulz mentioned the work that’s being done by the Advanced Processor Technologies (APT) group at the University of Manchester dubbed the Spiking Neural Network Architecture, or SpiNNaker, as well as the SpiNNaker2 project at the University of Dresden.
“Here the idea is to kind of take inspiration from how the brain works and put this into a compute element, and work in a similar way to how neurons work in the brain,” Schulz said. “It’s specialized to certain problems, but [has the potential] to be very, very effective.”
Another fairly exotic technology that could become an HPC accelerator are photonic systems, Shulz said. “We know photonics already from optical interconnects,” Schulz said. “But also it can be used for certain linear algebra optimizations and acceleration paths.” DNA storage and DNA computing also holds promise.
Prepare for a diversity of processors, Professor Schulz said
None of these technologies by themselves will solve the scaling problems that we face, Schulz said. But hopefully when we combine them, we can develop systems that meet our needs.
“So we have this really heterogeneous world. No single architecture wins, but we have a large a large different set of architectures to play with,” he said. “Each of these systems comes with their own programming models, with their own programming philosophy and kind of design points. That is not trivial…So we need to work on how can we bridge some technologies? How can we abstract some technologies and how we can really make this whole gel together? This is one of the biggest problems still.”
Providing an abstraction and a common interface that allows all of these various accelerators to work together, without burdening the user with stifling complexity, is a problem that Shulz is working on. There are various aspects to this problem that need to be worked out, from the access mechanism itself, how to allocate jobs, and how to allocate resources so that these various systems work cohesively in a real-time manner (because batch scheduling will not work) will require a lot more work by the HPC community, he said.
“There’s different ways to spin these stacks. But again, you have these layers, that we know also from classic HPC stacks, and we need to map them together,” vsaid. “We end up basically with this final challenge that we really have to do. We have all these stacks for the different places that kind of look the same, but not quite. We have to line them up in the proper way, but we have to support diversity in technologies. We have to support diversity in software stacks, environments, and how can we really run them together now and really make this work as one system?”
Integration standards will be important to ensure that everything fits together. Ditto for resource discovery. Workflows will need to be documented to ensure the right provenance, and locality must be respected once data start moving around. Long-term maintainability cannot be ignored; otherwise the whole thing collapses.
“Heterogeneity is inevitable,” Schulz concluded. “We have it already now and we’ll get more, and it will be the essential mechanism to drive more performance.”
A version of this story first ran in HPCwire.
Related
Hyperion Research Slices and Dices the HPC-AI Market
The market for HPC and AI for science solutions grew a healthy 17% in 2025,…
AWS Unveils New Tools Aimed at Making Enterprise AI Agents More Effective
AWS expanded its enterprise AI portfolio this week with a collection of new tools aimed…
AI Data Center Spending Grew 60% Last Year, Could Hit $1T Next Decade, Intersect360 Says
The world spent more than $300 billion on AI data centers in 2025, representing a…
Five Things to Watch as ISC 2026 Kicks Off in Hamburg
he International Supercomputer Conference is upon us, featuring four days of HPC education, collaboration, and…
Everpure Wants to Improve Data Management for the Coming AI Wave
Everpure (formerly Pure Storage) may have changed its name a few months back, but the…
If Reproducibility in AI Is Important, Try Model Flows
Reproducibility is absolutely critical in science, but it’s a troublesome characteristic when it comes to…
Broadcom Announces VMware Cloud Foundation 9.1
PALO ALTO, Calif., May 5, 2026 — Broadcom Inc., a global technology leader that designs,…
NVIDIA Announces Financial Results for 1st Quarter Fiscal 2027
SANTA CLARA, Calif., May 21, 2026 — NVIDIA (NASDAQ: NVDA) has reported record revenue for…
Anthropic Unveils ‘Project Glasswing’ as Claude Mythos Targets Software Vulnerabilities
April 9, 2026 — Anthropic has announced Project Glasswing, a new initiative that brings together Amazon…
Google Unveils Gemini Enterprise Agent Platform, Expands Vertex AI into Full Agent Stack
At Google Cloud Next 2026 in Las Vegas this week, the company announced major changes…
Cerebras Systems Announces Launch of Initial Public Offering
SUNNYVALE, Calif., May 4, 2026 — Cerebras Systems, Inc. today announced that it plans to commence…
What Stanford’s HAI Report Says About AI in Science
Progress in artificial intelligence continues to accelerate across a range of expert disciplines, according to…
Qualcomm and Hugging Face Expand Relationship to Advance Open, Developer-Driven AI from Device to Cloud
NEW YORK, June 26, 2026 — Qualcomm Technologies, Inc. has announced the expansion of its strategic…
Nebius AI Cloud 3.6 Strengthens Developer Experience and Governance for Production Operations
AMSTERDAM, June 25, 2026 — Nebius has unveiled Nebius AI Cloud 3.6, adding new developer experience,…
Mistral Unveils OCR 4 for Enterprise Search, RAG and Document Processing
June 25, 2026 — Mistral has announced the release of Mistral OCR 4, featuring bounding boxes,…
Runlayer Raises $30M Series A to Help Enterprises Become AI-Native
Felicis and Khosla Ventures double down to help Runlayer bring AI enablement and governance to…
Qualcomm Strengthens Data Center AI Push with Modular Acquisition
NEW YORK, June 25, 2026 — Qualcomm Incorporated has announced that it has reached an agreement…
CoreWeave Partners with Conapto to Expand AI Cloud Capacity in Sweden Powered by Renewable Energy
CoreWeave’s purpose-built AI cloud delivers further immediate capacity for European customers STOCKHOLM, June 25, 2026…
