When John Shalf stepped onto the stage for his award presentation at SC25, the atmosphere was charged with the sense of a field at a pivotal moment. His talk, titled "A Retrospective on Science-Driven System Architecture and Grand Challenges for the New Century," traced his professional journey, which reflected the evolution and challenges of modern high-performance computing.

From Early Experiments to Scientific Breakthroughs

Shalf recounted his early days at Virginia Tech, where he was first introduced to reconfigurable computing and DNA sequence comparison, drawing him into the high-performance computing (HPC) ecosystem. This foundation led him to Oak Ridge, where he contributed to materials-prediction codes and experienced the renowned "frog memo." He later transitioned to the National Center for Supercomputing Applications (NCSA), where he worked on some of the most ambitious codes of that era.

He mentioned key projects such as Enzo for cosmology, Cactus for general relativity, and the SC95 initiative, which linked supercomputers across the continent to create a single distributed machine simulating colliding galaxies in real time.

These milestones were not mere footnotes; they represented significant advances leading to one of the greatest scientific achievements of the century: the detection of gravitational waves. Shalf emphasized that the 15-year gap between the prediction and confirmation of these waves was not a delay, but rather a testament to the kind of sustained, disciplined computation that truly defines scientific progress.

Designing for the Workload, Not the Hype

A key theme of Shalf's talk was clear and direct: future systems should be designed around real workloads rather than aspirational benchmarks. He pointed out the development of the Sustained System Performance (SSP) benchmark at Berkeley Lab, which aims to accurately represent system performance instead of flattering it.

This philosophy relies on workload analysis, the use of algorithm/application matrices, and collaboration with applied mathematicians. This transition from focusing on synthetic performance to emphasizing actionable intelligence signifies a significant evolution in high-performance computing (HPC) thinking.

Bandwidth, Wires, and the Growing Crisis Beneath the Surface

Shalf explored the fundamental technical challenge of the post-Dennard era: as transistors continue to shrink, the reliability of wires decreases. While bandwidth increases, so does congestion. Memory channel speeds may improve, but latency issues remain significant. The once-reliable engineering playbook is now under strain due to power constraints and interconnect limitations.

He reviewed earlier efforts involving heterogeneous architectures, Cell processors, multi-core AMD designs, and initial flexibly assignable switch topologies. This research culminated in the hybrid H-FAST network approach, which reduced packet switching by leveraging persistent communication patterns.

This wasn’t mere tinkering; it provided evidence that structural change is achievable.

The Green Flash Era and the Rebirth of Co-Design

Shalf revisited the Green Flash project, an early initiative aimed at implementing deep hardware/software co-design for climate modeling. Rather than forcing scientific codes to conform to standard architectures, Green Flash directly optimized kernels for the hardware, automatically tuning them across various architectures and prototyping custom accelerators well before the modern RISC-V renaissance.

The project's influence also reached into the industry. Shalf pointed out that Google's TPU lineage owes a conceptual debt to the early co-design culture that was established during this period.

A Historical Link: Berkeley Lab's Breakthrough in Data Transfer

At one point in his talk, Shalf paused to highlight a significant milestone in networking, which was documented in Steve Fisher's report from July 3, 2002, about Berkeley Lab's demonstration of 10-gigabit Ethernet. https://www.supercomputingonline.com/latest/924-berkeley-lab-proves-10-gigabit-ethernet-data-transfer-is-a-reality 

Long before terms like "AI clusters" and "hyperscale fabrics" became common, Berkeley Lab demonstrated that multi-gigabit, wide-area data movement was not merely a concept of science fiction but a crucial component of scientific infrastructure. Their demonstration achieved data transfer at unprecedented speeds, validating 10GbE as a practical backbone for research networks and paving the way for the distributed science workflows we often take for granted today.

This achievement marked the beginning of an era where the bottleneck shifted decisively from computation to communication, an insight that resonates with Shalf's warnings even today.

Energy: The Defining Constraint of Our Time

Shalf's central thesis made a significant impact: energy is the real barrier.

He noted the end of Dennard scaling, emphasizing that wire delays are now surpassing transistor improvements. Additionally, hyperscale AI is driving consumption at the grid level.

Shalf argued that the next wave of innovation will not come from brute-force scaling but rather from specialization, advanced packaging, and chiplets. Most importantly, he highlighted the need for an expanded definition of co-design that integrates materials, circuits, architecture, and algorithms into a unified approach.

Reversible Logic and the Frontier Beyond Thermodynamics

In one of the most progressive sections of the talk, Shalf introduced reversible computing and topological materials as promising avenues for achieving ultra-low-energy computation. By eliminating the need for bit erasure, reversible logic avoids thermodynamic limits altogether, challenging the conventional belief that computation must always incur an energy cost. This serves as a reminder that future breakthroughs may look very different from the incremental improvements of the past decade.

Closing: A Field Ready for Reinvention

Shalf concluded on an optimistic note, albeit with a sense of urgency. He argued that the future of high-performance computing will not be dominated by massive machines or sheer scaling alone. Instead, it will belong to systems designed with humility, shaped by genuine scientific needs, and developed through collaboration rather than stagnation.

The applause that followed was not simply in recognition of a successful career. It was a response to a vision, a call for computing to reinvent itself, as it has done throughout history whenever science has demanded it.

At the bustling halls of the SC25 conference in St. Louis, thousands of high-performance computing (HPC) professionals gathered in America's Ballroom for what may become a pivotal moment in the evolution of digital life. Futurist and author Thomas Koulopoulos delivered a keynote titled "Gigatrends: The Exponential Forces Shaping Our Digital Future", and the tone was nothing short of electric.

Koulopoulos opened with a challenge: the world is not merely moving forward, it is accelerating. From healthcare to workspaces, from the rise of digital selves to emergent trust frameworks, he mapped a future where computing is no longer a tool but a partner. "We're entering an era where the machine forest breathes alongside us," he said his phrase a metaphor for aligning human intent and computational systems.

He laid out three central pillars of change:

• Exponential connectivity and intelligence: As compute density and network reach expand, machines and humans will co-evolve. The keynote emphasized how existing architectures must shift from linear updates to fractal, self-optimising systems.
• Human-machine symbiosis: Beyond "augmented human," Koulopoulos depicted a world where the digital self (our data profile, our algorithmic twin) participates in decision-making, personal, corporate, societal.
• Trust as an infrastructure: He argued that in the upcoming "digital life" epoch, trust mechanisms (data governance, identity, transparency) will become as fundamental as power grids or fibre-optic networks.

When the slides lit up with a visualization of interconnected nodes representing human lives, devices, data flows, and decisions, the audience sat up. For many in the room, this was more than another HPC talk; it was an invitation to imagine themselves as builders of society's next operating system.

One memorable line: "If you're just scaling faster, you're not doing the job. You're slipping into the past at a higher velocity." This served as both a warning and a rallying cry for the HPC community gathered at SC25.

The keynote did more than inspire, it landed practical imperatives. Koulopoulos urged attendees to:

• rethink how HPC infrastructure supports not just simulation but intelligence-at-scale;
• design systems that are ethically transparent and resilient;
• treat data not as a trace but as a partner ecosystem that learns and gives back.

It resonated. As one attendee later remarked in the lobby, "I came expecting hardware specs and benchmarks; I left thinking about human destiny."

The message resonates deeply with the conference's mission: the annual ACM/IEEE Supercomputing Conference (SC) series brings together thousands of scientists, engineers, researchers, and developers working at the bleeding edge of compute, networking, storage, and analysis.

By choosing Koulopoulos as the keynote, SC25 clearly signalled that the conversation is shifting. The focus isn't just "what can we compute?" but "how should we live alongside our machines?" The keynote sets the tone for what promises to be a week where HPC meets philosophy, infrastructure meets intention, a perfect fit for the city of St. Louis and its spirit of innovation.

Final thought:

In an age of compute-saturation and data deluge, the real frontier is alignment: aligning machines with human values, aligning infrastructure with insight, aligning rapid change with timeless purpose. SC25's opening address made that frontier visible and invited us all to step into it.

PCIe Gen5 Muscle for AI and Cloud

• Up to 14.5 GB/s read, 12 GB/s write
• Up to 3.3M / 1.05M random read/write IOPS
• Configurations up to 30.72 TB (X201) and 15.36 TB (D201)

AI Agents on iGPUs, 25× Faster Than Before

• Up to 25× faster AI agent performance
• A drop in latency from 73 seconds to ~4 seconds in one real-world demo, GenAI inference on YouTube video content.

Scaling Toward Extreme Capacity

Industry Voices at SC25

Why SC25 Cares

• Client: AI agents on iGPUs
• Enterprise: X201 for training and HPC
• Cloud/hyperscale: D201 and the ultra-dense D205V series

Availability

• Pascari X201 / D201: Shipping to select enterprise customers and OEMs by end of 2025
• aiDAPTIV+ iGPU systems: OEM rollouts in early 2026
• More details at phison.com