EDA
System
IP
SolvNetPlus
Training & Education
The semiconductor industry has followed a familiar script for decades: faster transistors enabled faster processors, and software benefited almost automatically.
That script is now being flipped.
Manufacturing continues to advance, but performance gains are no longer automatic. At the same time, AI workloads are exploding in scale and complexity. This is forcing the industry to rethink what “compute” actually means — and how it must evolve.
That challenge was front and center during a panel discussion at the 2026 Synopsys Executive Forum — part of Synopsys Converge — where leaders from Intel, AMD, Synopsys, and the quantum computing community explored how hardware, software, and systems are being reshaped by AI.
Throughout the course of a wide-ranging discussion, a singular message became crystal clear: the next era of compute will be defined less by raw speed and more by architectural intelligence, the abstraction of hardware complexity, and the co-design of silicon, software, and systems.
Navigate AI Chip Development
Your essential guide to overcoming AI chip complexity and achieving successful silicon outcomes from design to deployment.
Re-architecting software for parallel execution
According to Prith Banerjee, senior vice president at Synopsys, the flattening of CPU clock speeds marked a turning point for the industry. One that forced software teams to confront a new reality.
“Once frequencies flattened around 2 GHz, we were forced into parallelism,” Banerjee said. “Software developers had to rethink algorithms.”
Early approaches such as shared-memory parallelism worked only up to a point. Distributed memory models enabled massive scale, but GPUs — and later FPGAs — introduced entirely new programming paradigms. Each step forward required deeper algorithmic change, not just incremental optimization.
“The best sequential algorithm is not the best parallel algorithm,” Banerjee noted.
What’s the implication for AI-era workloads? Performance gains can no longer be achieved via “compiler magic.” Such gains now hinge on re-architecting software for fundamentally different forms of parallel execution.
Left to right: Ian Cutress (moderator, More than Moore), Manoj Selva (Intel), John Martinis (QoLab), Vamsi Boppana (AMD), and Prith Banerjee (Synopsys) at Synopsys Executive Forum
Addressing new bottlenecks
As traditional workloads struggled to scale, AI models changed the game. Because they thrive on the parallelism GPUs and FPGAs deliver, unprecedented system-level gains were suddenly within reach.
“Instead of 5–7% year-over-year improvements, we’re seeing 3× to 10× gains,” said Vamsi Boppana, senior vice president of AMD’s AI Group.
That scale, however, has placed new levels of strain on other parts of the system. Memory, networking, and storage have all become choke points.
“Everything is a bottleneck,” Boppana said. “The real limit is my CFO.”
While the comment drew laughter, it underscored a serious truth. As GPU prices reach tens of thousands of dollars per unit, value is increasingly measured in return on investment, not component cost. If a system delivers a 10× performance boost, customers will find a way to justify it.
Supporting heterogeneous compute
With systems now integrating CPUs, GPUs, accelerators, and custom silicon, supporting heterogeneity and reducing complexity have become two of the industry’s toughest challenges.
“We need to take away this complexity from the end user,” said Manoj Selva, vice president at Intel. “They don’t care whether something runs on a CPU, GPU, or accelerator. They simply want to run a simulation.”
Achieving that simplicity requires robust abstraction layers, standardized software stacks, and middleware that can adapt as hardware evolves. Maintaining consistency across increasingly large and heterogeneous systems is now as much a software challenge as a hardware one.
Selva also pointed to an emerging helper in this transition: AI agents.
“Agents tolerate rapid changes in compute paradigms,” he said. “They don’t get upset when things shift under them.”
Redefining chips through system level demands
The panel also explored whether chips still define systems, or whether the relationship has reversed. The consensus leaned strongly toward the latter.
“Systems now define chips more than ever,” said Boppana. “Power, cooling, networking, even geographic synchronization — all shape silicon decisions.”
AI will increasingly participate in those decisions, the panelists predicted.
“AI will eventually design systems that design chips,” Boppana said. “We’re not quite there, but we’re close.”
That shift places new emphasis on simulation and emulation, ensuring that increasingly autonomous design processes remain understandable and trustworthy.
Considering the role of quantum computing
While AI dominates much of today’s compute conversation, quantum computing represents a parallel — but very different — trajectory.
“Quantum computers are not replacements for classical systems,” said Dr. John Martinis, Nobel Prize–winning physicist and co-founder of QoLab. “They excel at problems rooted in quantum mechanics.”
Applications such as materials science, chemistry, and drug discovery stand to benefit first. Even modest improvements can have outsized effects.
“A five‑to‑ten percent improvement can have huge economic value,” Martinis said.
Still, large-scale quantum systems remain constrained by manufacturing challenges, limited qubit counts, and economic realities. The path to commercial impact will depend on industrialization — and deep collaboration with the semiconductor ecosystem.
Evolving skills alongside tools
As compute architectures evolve, so do engineering roles. The lines between hardware and software continue to blur.
“Hardware engineers write Python now,” Selva observed. “They don’t draw transistors — they work in EDA tools.”
At the same time, AI is reshaping how engineers work at every level.
“Programming abstraction just keeps rising,” said Banerjee. “Now students must learn how to program through AI tools.”
Rather than replacing engineers, AI is turning them into orchestrators — managing teams of intelligent agents that explore design spaces faster than any human could alone.
Redefining success
Perhaps the most important takeaway from the discussion was that performance alone is no longer the goal.
“Customers care about total cost of ownership,” Banerjee said, “not just peak benchmarks.”
In an era where R&D investment spans trillions of dollars globally, the winners will be those who can accelerate innovation holistically — from silicon to software to system and beyond.
The age of AI is forcing compute to grow up. Faster hardware still matters, but smarter systems matter more.
Continue Reading
