Researchers from EPFL, AMD, and the University of Novi Sad have uncovered a long-standing inefficiency in the algorithm that programs millions of reconfigurable chips used worldwide, a discovery that could reshape how future generations of these are designed and programmed.

Many industries, including telecoms, automotive, aerospace and particle physics rely on a special breed of chip called the Field-Programmable Gate Array (FPGA). Unlike traditional chips, FPGAs can be reconfigured almost endlessly, making them invaluable in fast-moving fields where designing a custom chip would take years and cost a fortune. But this flexibility comes with a catch: FPGA efficiency depends heavily on the software used to program them.

Since the late 1990s, an algorithm known as PathFinder has been the backbone of FPGA routing. Its job: connecting thousands of tiny circuit components without creating overlaps.

For decades, it worked so well that it became the standard. However, as circuits grew larger, engineers began encountering frustrating slowdowns and occasional outright failures. Designs that should have worked were often labeled “unroutable.”

Now, with colleagues from the University of Novi Sad and the technology company AMD, researchers from the Parallel Systems Architecture Laboratory (PARSA) in the School of Computer and Communication Sciences have come one step closer to untangling the inner workings of this classic algorithm.

In their paper, which received the Best Paper Award at the 33rd IEEE International Symposium on Field-Programmable Custom Computing Machines, they revealed why these failures happen and how PathFinder’s limits can be overcome.

Cracks in the algorithm

“In fact, it’s not surprising that PathFinder sometimes fails,” explained Shashwat Shrivastava, Ph.D. student with PARSA and first author of the paper.

“Very early on, researchers showed that the problem behind FPGA routing is extremely hard. Later, the creators of the original algorithm, together with a few collaborators, found cases where PathFinder would never succeed—but they noted such cases wouldn’t appear in practice.”

For decades, it seemed they were correct—PathFinder worked surprisingly well.

“PathFinder worked so well, in fact, that when it failed, people rarely questioned the algorithm. Instead of venturing inside to see what was going on, they tweaked its parameters, modified circuits, or switched to larger FPGAs,” added Stefan Nikolić, an EPFL alumnus and now a professor at the University of Novi Sad.

“Part of the reason for this is that it is rather difficult to understand what PathFinder is actually doing on examples of practical importance. Modern circuits are so large that their signals form veritable on-chip jungles.”

Enter the forest

“So, we really needed to look at the individual trees in that jungle,” continued Shrivastava, “and I really mean trees. Each signal—a connection that carries information between circuit components—must reach multiple destinations without overlapping other signals. FPGA routing is essentially about building one tree for each signal on the chip.”

While working on another project that relied on PathFinder, the team kept seeing results that defied intuition. At first, they blamed external factors, not the algorithm itself. Eventually, they realized they needed controlled examples: small, tricky cases where a solution definitely existed, and in which PathFinder should succeed.

“We needed real, practical examples, and lots of them, to understand what was really going on,” Shrivastava explains. “So, we built a framework to automatically extract small, hard problems from real circuits. Watching how PathFinder struggled with these helped us uncover issues that had remained hidden for a very long time.”

Power in partnership

“This breakthrough would have been much harder without industry support,” said Mirjana Stojilović, Shrivastava’s Ph.D. advisor. “From the start, we collaborated with Chirag Ravishankar and Dinesh Gaitonde from AMD. They helped us model FPGAs as close as possible to commercial devices, ensuring our findings had real-world impact.”

Once the framework was ready, things moved quickly. The team found that PathFinder often built routing trees larger than necessary, increasing the risk of overlaps. The problem came from the order in which it created and added new branches to the trees.

“In retrospect, this is intuitive, but somehow it went largely unnoticed for many years,” Shrivastava said. “Our first solution was simple: try different orders and pick the one that results in the smallest tree. Experimentally, it worked surprisingly well.”

The team is now exploring more scalable solutions. “I am especially proud that Summer@EPFL interns have been contributing significantly. One of them, Sun Tanaka, is also a co-author of the paper,” added Stojilović.

“Our discovery could reshape how millions of FPGAs are programmed and influence the design of future generations of these reconfigurable chips.”

Despite increasing awareness, AI shopping assistants have yet to achieve widespread adoption in the US. While 43 per cent of Americans are aware of these tools, only 14 per cent have used them. Usage is highest among Gen Z (24 per cent) and parents of children under 18 (21 per cent), whereas just 7 per cent of baby boomers have tried them, according to a new YouGov study.

For current users, key benefits include getting answers to product questions (44 per cent), finding specific items (41 per cent), and locating the best deals (34 per cent). For non-users who are open to trying AI, the most appealing features are price comparisons (67 per cent), evaluating similar products (56 per cent), and accessing product information (55 per cent).

However, trust is a major barrier. Forty-one per cent of Americans say they do not trust AI shopping assistants at all, and only 13 per cent mostly or completely trust them—compared to 53 per cent who trust personal recommendations. Privacy concerns, a preference for human assistance, and fears of upselling further fuel scepticism.

AI shopping assistants remain underused in the US despite 43 per cent awareness, with only 14 per cent having tried them.
Gen Z and parents show the highest adoption.
Users seek answers, deals, and product info, but trust remains low due to privacy concerns and scepticism.
Interest is strongest in clothing, electronics, and groceries.
Wider adoption hinges on building trust and proving real value.

When it comes to shopping categories, consumers are most open to using AI for clothing and accessories (20 per cent), consumer electronics (21 per cent), groceries and household essentials (19 per cent), and travel planning (18 per cent). Interest is much lower for purchases involving finances, vehicles, or pet care.

Retailers such as Walmart and Amazon are already deploying AI tools like Sparky and Rufus to improve customer experience. Yet, the study highlights that broader adoption depends on demonstrating genuine value, safeguarding data, and rebuilding trust—especially among more cautious consumers.

Fibre2Fashion News Desk (SG)

OpenAI could now be the world’s most valuable startup, ahead of Elon Musk’s SpaceX and TikTok’s parent company ByteDance, after a secondary stock sale designed to retain employees at the ChatGPT maker.

Current and former OpenAI employees sold $6.6 billion in shares to a group of investors, pushing the privately held artificial intelligence company’s valuation to $500 billion, according to a source with knowledge of the deal who was not authorized to discuss it publicly.

The investors buying the shares included Thrive Capital, Dragoneer Investment Group and T. Rowe Price, along with Japanese tech giant SoftBank and the United Arab Emirates’ MGX, the source said Thursday.

The valuation reflects high expectations for the future of AI technology and continues OpenAI’s remarkable trajectory from its start as a nonprofit research lab in 2015.

But with the San Francisco-based company not yet turning a profit, it could also amplify concerns about an AI bubble if the generative AI products made by OpenAI and its competitors don’t meet the expectations of investors pouring billions of dollars into research and development.

OpenAI CEO Sam Altman has sought to dismiss those concerns, most recently last week, when he toured a massive data center complex being built to run the company’s AI systems in Abilene, Texas.

“Between the ten years we’ve already been operating and the many decades ahead of us, there will be booms and busts,” Altman said after being asked about a bubble. “People will overinvest and lose money, and underinvest and lose a lot of revenue.”

He added that “we’ll make some dumb capital allocations” and there will be short-term ups and downs but that “over the arc that we have to plan over, we are confident that this technology will drive a new wave of unprecedented economic growth,” along with scientific breakthroughs, improvements to quality of life and “new ways to express creativity.”

Just this week, the company launched two different business ventures, one a partnership with Etsy and Shopify for online shopping through ChatGPT and another a social media app, Sora, for generating and sharing AI videos.

OpenAI has been struggling to offer investors and staff the same perks and compensation as the publicly traded tech giants with which it competes. Facebook parent Meta Platforms, in particular, has been on a hiring spree for elite AI engineers and in June made a $14.3 billion investment in AI company Scale that recruited its CEO Alexandr Wang.

OpenAI’s for-profit subsidiary, valued at $500 billion, is technically controlled by the board of OpenAI’s nonprofit and both are still bound to pursue the nonprofit’s charitable purpose.

OpenAI’s partnerships with major companies and its plans to change its corporate structure have drawn the scrutiny of regulators, including the attorneys general of California and Delaware, who oversee charitable organizations that operate or are incorporated in their states.

The company has made big deals in recent weeks with Oracle and SoftBank, its partners on a data center venture called Stargate, and with chipmaker Nvidia, which makes the specialized AI chips those data centers need. At the same time, it has lessened its reliance on longtime backer Microsoft.

In September, OpenAI announced it had reached a tentative agreement with Microsoft about the future stake of its nonprofit in its for-profit corporation but released few details.

It also opened applications for nonprofits to apply for $50 million in funding from OpenAI, an effort it launched in response to the recommendations of an advisory board. The grants will go toward projects that increase public understanding of AI, support the design of AI for uses that communities want and increase economic opportunity. The deadline to apply closes on Oct. 8.

© 2025 The Associated Press. All rights reserved. This material may not be published, broadcast, rewritten or redistributed without permission.