llmda newsletter ad (2)
Semiconductor Expert Forum

Posted on

CEO Interview with Juniyali Nauriyal of Photonect

CEO Interview with Juniyali Nauriyal of Photonect
by Daniel Nenni on 02-22-2026 at 2:00 pm
Categories: CEO Interviews, Silicon Photonics

Juniyali

Juniyali Nauriyal is the CEO and Co-Founder of Photonect, a photonics startup focused on commercializing advanced fiber-to-chip attachment technologies.  Juniyali is the co-inventor of Photonect’s core technology, which forms the foundation of the company. As CEO, she leads Photonect in translating cutting-edge photonic packaging research into scalable, real-world solutions.

She has participated in prestigious accelerator programs including Activate, the Luminate Accelerator. Under her leadership, Photonect won the Grand Prize at the New York State Business Plan Competition (2022). Juniyali has received multiple competitive awards and scholarships, including the Corning Women in Optical Communications Scholarship (2022) from Optica, the SPIE Optics and Photonics Education Scholarship (2022), the Harvey M. Pollicove Memorial Scholarship (2018) from Optica, and the Best ISA Student Award (2016) from ISA–Maharashtra. She currently serves as Vice Chair of IEEE Women in Engineering (WIE) Rochester and was named Emerging Technology Woman of the Year (2025). Juniyali holds 7 patents and has authored 20 technical publications, including 2 first-author papers.

Tell us about your company?

Photonect is a Rochester-based startup addressing a critical bottleneck in photonic packaging: slow, costly, and unreliable fiber-to-chip attachment driven by epoxy. Attaching an optical fiber is as delicate as aligning two strands of hair, yet current approaches rely on glue. Photonect’s technology includes a patented chip architecture design called the oxide mode converter and laser fusion process that forms permanent glass-to-glass bonds in under a minute. This improves coupling efficiency from ~50% to ~80%, maintains <1 dB loss, increases throughput by 10×, and reduces per-device cost by 50%. This technology is being advanced into PIX-Attach, a Rochester-designed laser splicing system built for high-volume photonic integration set to launch at OFC 2026.

Please share a little bit more about how company was founded?

Photonect was founded by Dr. Juniyali Nauriyal as a spin-off from the University of Rochester, rooted directly in her doctoral research in integrated photonics. Originally an engineering student in India, Juniyali came to the U.S. to pursue an MS in Optics at the University of Rochester, where she joined the lab of Dr. Jaime Cárdenas. That collaboration evolved into her PhD work and eventually started Photonect itself.

During her PhD, Juniyali co-invented Photonect’s core technology along with Dr. Cárdenas, who is now the company’s CTO and Co-Founder. Rather than taking a traditional path into a large tech firm, Juniyali chose to commercialize her research to ensure it could deliver real-world impact. Her work led to a breakthrough approach to fiber-to-chip light coupling, improving efficiency while addressing the growing sustainability challenges driven by AI and data-center expansion.

Inspired by issues such as the rising energy demands of data centers and even their migration to colder regions to manage heat, Juniyali founded Photonect with a clear mission: to build high-performance photonic technologies that scale responsibly and support a more energy-efficient future.

What problems are you solving?

We address the limitations of epoxy-based fiber-to-chip attachment, which is slow, costly, and inherently unreliable. Today, epoxy is commonly used to connect/attach optical fibers to chips even though these joints are critical to data transfer efficiency, often resulting in up to 50% signal loss, increased heat generation, and long, labor-intensive assembly times that severely limit scalability. Our laser-assisted attachment technology replaces epoxy with a glass-to-glass bond, like soldering. This enables fiber-to-chip attachment in seconds instead of minutes, while delivering long-term reliability and stability across wide temperature ranges. In parallel, our proprietary mode converter technology significantly improves coupling efficiency, achieving <1 dB loss per chip facet and reducing overall link budget requirements by ~25%.

What Industries is your solution a good fit for ?

Our solution is a strong fit for industries undergoing a shift toward dense, high-performance optical integration, driven by the speed, power, and reliability demands of modern computing. This includes AI data centers and cloud infrastructure, high-performance computing, telecommunications, and emerging markets like quantum technologies. As copper and conventional photonic interconnects hit physical limits, Photonect’s technology delivers a step-change in performance, scalability, and energy efficiency, enabling these industries to keep pace with rapidly growing system demands.

What keeps your customers up at night?

Our customers worry about scaling photonic products to high volume without blowing up cost, power use, or yield. Today’s fiber-to-chip attachment is slow, energy-intensive, and requires specialized manpower, often forcing teams to choose between scalability and efficiency putting timelines, margins, and reliability at risk specially now as AI and data-center demand accelerates.
Q6.What does the competitive landscape look like and how do you differentiate?

In fiber-to-chip attachment landscape, existing solutions include wire bonding, fiber packaging with epoxy, and passive alignment approaches some players offer high-precision automated assembly systems that still rely on active alignment and adhesive bonding, none of which have been able to fully replace epoxy. Photonect takes a fundamentally different approach by adhesives through its proprietary laser-assisted attachment process. Our technology enables fiber-to-chip attachment in few seconds, achieves <1 dB optical loss, and delivers throughput of up to 60 attachments per hour.

In addition, our proprietary mode converter designs further reduce coupling losses by 15–20%, while maintaining high power compatibility and reliability across extreme temperature ranges.

What new features/technology are you working on?

We’re working on PIX-Attach, a next-generation laser-assisted, epoxy-free fiber-to-chip attachment platform that dramatically speeds up photonic packaging, cutting attach time to ~60 seconds per unit, enabling ~60 units per hour, improving alignment stability, reducing energy use, and significantly increasing production throughput while lowering the number of machines needed on the factory floor. Our value proposition lies in creating a premium and customized fiber attach equipment that fit’s the customers need. We are all set to launch this in mid-March 2026.

How do customers normally engage with your company?

We’re present at all major conferences such as OFC, Photonics West, and ECOC. You can also reach out to us directly at info@photonectcorp.com, or submit a query or request a meeting through our website.

 

CONTACT PHOTONETC

Also Read:

CEO Interview with Aftkhar Aslam of yieldWerx

CEO Interview with Elad Raz of NextSilicon

VSORA Board Chair Sandra Rivera on Solutions for AI Inference and LLM Processing

Podcast EP332: How AI Really Works – the Perspectives of Linley Gwennap

Podcast EP332: How AI Really Works – the Perspectives of Linley Gwennap
by Daniel Nenni on 02-20-2026 at 10:00 am
1 Comment

Daniel is joined by Linley Gwennap, technology analyst and author of the new book “How AI Really Works: The Models, Chips, and Companies Powering a Revolution.” Linley was the long-time editor of Microprocessor Report and chaired the popular Linley Processor Conferences.

Dan explores what impact AI is having on the market and the population with Linley, who offers some straight-forward explanations of how AI is impacting our world. Trends in AI model size are discussed, as well as competitive dynamics. Linley offers perspectives on how trends today and in the future will impact the market size and the key players. Trends in AI model size are also reviewed, along with some predictions about how AI will impact the overall workforce, both now and in the future.

You can get a copy of Linley’s new book here.

The views, thoughts, and opinions expressed in these podcasts belong solely to the speaker, and not to the speaker’s employer, organization, committee or any other group or individual.

Posted on

What is the 3nm Pessimism Wall and Why is it An Economic Crisis?

What is the 3nm Pessimism Wall and Why is it An Economic Crisis?
by Mike Gianfagna on 02-20-2026 at 8:00 am
Categories: ClockEdge, EDA

What is the 3nm Pessimism Wall and Why is it An Economic Crisis?

Chip design is getting more difficult as technology advances. Everyone knows that. A lot of the discussion around these issues tends to focus on the demands posed by massive AI workloads and the challenges of shifting to heterogeneous multi-die design. While these create real problems, there is an underlying effect that is making the situation much worse than it needs to be: The ROI on advanced-node scaling is compressing in ways most teams do not yet quantify.

For three decades, Moore’s Law was an economic engine. Today, at 3nm and below, that engine is slowing. While foundries promise massive PPA (power, performance, and area) gains, the reality for most design teams is a “Performance Mirage.” Despite multi-billion-dollar investments in 3nm Gate-All-Around (GAA) and FinFET migrations, a large portion of the promised performance of these advances can be out of reach. It is often being sacrificed to “margin” reserved solely to compensate for modeling uncertainty. Let’s refer to this structural inflation of clock margin as the “Pessimism Wall”.

The good news is that this margin is not a law of physics.  It can be safely reclaimed and redirected toward real silicon limits. More on that shortly. But first let’s answer the question, what is the 3nm pessimism wall and why is it an economic crisis? The answer begins with understanding how margin accumulates – and why that accumulation has become economically consequential.

Anatomy of the Crisis

At 3nm, clock sign-off guard bands have exploded to 25–35% of the total clock period. This is not a choice; it is a structural consequence of abstraction-based sign-off methodologies. The following data highlights the mechanisms driving this structural margin inflation.

The data below reflects trends widely observed across advanced-node programs. While exact values vary by design, the structural pattern is consistent.

  • The 2.5x Over-Design Trap: Applying 28nm-era sign-off assumptions to 3nm designs forces designers to over-design clock networks by 2.5x. You are often paying for buffers, area, and routing complexity that the silicon does not need.
  • The Near-Threshold Danger Zone: As voltages approach device thresholds, delay behavior becomes exponential and non-linear. Standard static timing analysis (STA) over-linearizes these effects, forcing an “uncertainty tax” of 8–12% of the clock period just to remain “safe”.
  • The Jitter Black Hole: Power-supply-induced jitter (PSIJ) and simultaneous switching now consume 5–10% of the margin. Traditional tools treat this as a static guess.

All these effects hide useful margin behind the pessimism wall.

A Closer Look at the Pessimism Wall

Every picosecond of unnecessary margin has a direct impact on the project’s bottom line. The following table breaks down the contributors that can cumulatively drive total clock margin toward ~25–35% range:

The contributors above are individually defensible and grounded in advanced-node physics. What creates the pessimism wall is their cumulative stacking.

In abstraction-based sign-off flows, voltage sensitivity, jitter, aging, and variability are typically evaluated independently and conservatively. Worst-case assumptions stack because electrical interactions are not jointly resolved in time and voltage.

The silicon did not become 35% worse. Our abstractions became cumulatively more conservative. To be clear, the issue is not transistor device models themselves. The structural pessimism arises from abstraction-based timing methodologies and independently stacked worst-case assumptions that approximate electrical behavior rather than directly solving it.

The Economic Consequences – A Crisis in the Making

Leaving 10–15% recoverable clock margin on the table is not a modeling inconvenience – it can be a massive competitive liability. Let’s look a bit closer at what’s involved.

  • The Power Penalty: Because dynamic power scales with the square of voltage, a 10% reduction in margin translates to a ~18–20% reduction in dynamic clock power. Given that clock networks consume 30–40% of SoC power, this often determines whether a design leads its segment or thermally limits its own performance.
  • The Revenue Loss (SKU binning): Reclaiming ~10% margin enables a 300 MHz boost on a 3 GHz target. In high-volume production, shifting even 10% of volume into a premium performance bin can represent hundreds of millions of dollars in incremental revenue currently sacrificed to uncertainty.
  • Area Inefficiency: Abstraction-driven margin forces aggressive cell upsizing, leading to a 10–15% increase in clock tree area. This bloats die size and increases per-unit cost across millions of chips.
  • Field Failures: The industry’s reliance on broad “Guard Bands” actually increases risk:
    – Masked Failures – Broad margins “hide” specific electrical failures—like rail-to-rail or duty-cycle issues – until they hit the field.
    – Aging Roulette – Applying “Global Aging Taxes” ignores path-specific stress, leading to chips that pass tapeout but degrade prematurely in the field.

The Solution: Full-Clock Physics Enforcement

The crisis stems from one fact: Models have stopped keeping up with physics.

The most direct way to address structural pessimism is to replace timing abstractions and estimates with electrical resolution by performing detailed, accurate SPICE analysis on the entire clock. Up to now, this wasn’t practical for two reasons. First, standard SPICE runs on networks of this size would take an unreasonable amount of time and consume vast (and expensive) compute resources. And second, standard SPICE can’t even load networks of this size.

The good news is that these barriers are now gone. The ClockEdge Veridian suite delivers a family of SPICE-accurate analysis engines for timing, power, jitter, and aging. And Veridian delivers sign-off precision at real-world speed, revealing interactions that conventional flows miss. This enables full-clock waveform fidelity across timing, power, jitter, and aging interactions.

Veridian engines enable billion-transistor, unreduced SPICE analysis performed overnight. Some of the benefits of this include:

  • Eliminate Abstraction-Driven Guesswork: Enforce Kirchhoff’s Current and Voltage Laws across the entire netlist to eliminate table-lookup errors
  • Expose Hidden Failures: Veridian identifies rail-to-rail and duty-cycle failures that traditional STA “masks” with margin until it is too late
  • Path-Specific Aging: Stop applying global derates. Measure actual aged behavior to recover margin safely

The question is no longer whether the pessimism wall exists – physics proves it does. The question is whether your methodology is capable of exposing it before your competitor does.

 At advanced nodes, competitiveness is increasingly determined not by how much margin can be added, but by how much unnecessary margin can be safely removed.

The 3nm Pessimism Wall is not a silicon limitation – it is a modeling one.

The teams that resolve physics directly rather than approximate it will reclaim performance, power efficiency, and yield that others continue to surrender to uncertainty.

To Learn More

ClockEdge recently published a very informative white paper titled Reclaiming Margin in Advanced Nodes – Why Abstraction-Based Sign-Off Is Becoming the Dominant PPA Limiter at 3nm and Below.

This white paper is essentially a master class in how to preserve margin, performance and profits at advanced nodes. If you find yourself becoming a “slave” to ever-increasing design margins, this white paper is must-read. You can access your copy here. And that’s what the 3nm pessimism wall is and why it is an economic crisis.

Also Read:

The Risk of Not Optimizing Clock Power

Taming Advanced Node Clock Network Challenges: Jitter

Taming Advanced Node Clock Network Challenges: Duty Cycle

Posted on

CEO Interview with Aftkhar Aslam of yieldWerx

CEO Interview with Aftkhar Aslam of yieldWerx
by Daniel Nenni on 02-20-2026 at 6:00 am
Categories: CEO Interviews, EDA, Photonics, yieldWerx
1 Comment

IMG 9364 3

Aftkhar Aslam is the Co-Founder and Chief Executive Officer of yieldWerx and a semiconductor industry veteran with more than 30 years of experience spanning manufacturing, test engineering, yield management, IP strategy, and enterprise digital transformation.

Under his leadership, yieldWerx has become a trusted data and yield analytics platform supporting semiconductor companies across fab, assembly, test, advanced packaging, photonics, and AI-driven device manufacturing. The platform enables organizations to unify fragmented manufacturing data into scalable, actionable yield intelligence.

Prior to founding yieldWerx, Aftkhar held senior leadership roles at Texas Instruments, where he served as Worldwide Director of Test & Yield Management Solutions & Director of Digital Transformation in the space of Design and Delivery systems and solutions across HW and SW.

He also served as a Director within Accenture’s Industry X (IX) practice, where he advised leading global technology organizations including Intel, GlobalFoundries, Qualcomm, Lam Research, Microsoft, STMicroelectronics, and Skyworks. His consulting work focused on bridging the Design-to-Manufacturing divide — architecting Digital Thread and Digital Twin strategies that connected product design, IP management, manufacturing execution, test, and enterprise systems into unified operational frameworks.

Aftkhar holds patents and possesses deep expertise in intellectual property management and protection. His experience spans semiconductor IP lifecycle governance, secure data architectures, and protecting high-value design assets across complex global supply chains.

Tell us about your company.

yieldWerx is a semiconductor-focused data and enterprise yield analytics platform. We help manufacturers unify data across fab, assembly, test, inspection, and advanced packaging into a single environment where engineers can extract real insight — not just generate reports.

What makes us different is that we tend to operate where the problems are hardest. We work with highly specialized, niche products and manufacturing flows — whether that’s heterogeneous integration, chiplets, co-packaged optics, MicroLED with billions of pixel-level data points, or silicon photonics requiring optical and electrical correlation. These aren’t simple wafer-yield problems; they’re multi-domain, multi-stage challenges that traditional tools struggle to handle.

We’re purpose-built for semiconductor manufacturing at extreme scale and extreme complexity. Our platform is designed to manage unconventional data models, massive datasets, and deep traceability requirements without breaking performance or usability.

At a high level, we help companies move from fragmented data silos to a connected digital thread — accelerating yield learning, reducing ambiguity, and enabling smarter, faster engineering decisions in some of the industry’s most advanced and specialized product environments.

What problems are you solving?

The biggest problem in semiconductor manufacturing today isn’t lack of data — it’s fragmentation.

Data lives in MES systems, testers, inspection tools, spreadsheets, homegrown databases, and separate analytics platforms. Engineers spend enormous time manually stitching it together before they can even begin root-cause analysis.

We solve that by unifying the data model and enabling cross-domain correlation — electrical + optical, wafer + module, socket + silicon, defect + yield, and so on.

Another major problem is scale. Modern devices generate massive datasets. Traditional tools weren’t designed for billions of data points. Ultimately, we reduce the time from anomaly detection to root cause — and that directly impacts yield, cost, and time-to-market.

What application areas are your strongest?

We’re strongest in environments where complexity is high and data volumes are extreme.

That includes:

  • Advanced packaging (2.5D/3D, chiplets, CPO)
  • Silicon photonics
  • MicroLED and display technologies
  • AI and high-performance compute devices
  • Automotive and high-reliability semiconductor manufacturing

Anywhere there’s multi-stage manufacturing with complex traceability requirements — that’s where we add the most value.

What keeps your customers up at night?

Three things:

  1. Yield ramp speed — especially for new technologies. Every week of delay is expensive.
  2. Escapes or overkill at test — failing good parts or shipping marginal ones.
  3. Lack of traceability when something goes wrong.

They worry about whether they truly understand where yield loss is originating — is it the wafer, the packaging step, the bonding process, the socket, the test program?

If the answer takes weeks to figure out, that’s a problem. Our goal is to make that answer visible in hours or days.

What does the competitive landscape look like and how do you differentiate?

There are traditional yield tools, BI tools, and homegrown systems.

Traditional yield tools often focus on wafer-level analysis but struggle with cross-domain traceability. BI tools are flexible but require heavy customization and don’t inherently understand semiconductor manufacturing.

We differentiate in three ways:

  1. Semiconductor-native data model — we understand wafers, panels, bonding, pixel maps, optical lanes, serialized modules.
  2. Extreme scalability — billions of records without performance degradation.
  3. Closed-loop capability — we don’t just visualize data; we enable correlation across design, manufacturing stages to drive actionable decisions.

We’re not just another dashboard — we’re the infrastructure layer for yield intelligence.

What new features or technology are you working on?

We’re expanding heavily into:

  • Pixel-level and device-level analytics for MicroLED and advanced displays
  • Optical + electrical unified analysis for photonics and CPO
  • Advanced spatial analytics and pattern recognition
  • AI-assisted anomaly detection and predictive yield modeling from the start of design
  • Deeper integration with test hardware and equipment for closed-loop optimization

We’re also strengthening genealogy and digital-thread capabilities to support next-generation packaging and heterogeneous integration.

The industry is moving toward system-level understanding, not just wafer-level — and that’s where we’re investing.

How do customers normally engage with your company?

Most engagements start with a specific pain point — slow yield ramp, fragmented data, lack of traceability, or scaling a new technology.

We typically begin with a focused pilot or proof-of-value around a real manufacturing dataset. Once customers see how quickly we can unify and analyze their data, the engagement expands into a broader enterprise deployment.

We also work closely with equipment providers, OSATs, and ecosystem partners, because yield today is collaborative — not isolated.

At the end of the day, we’re a long-term partner. Once we’re embedded in the manufacturing data flow, we become part of the operational backbone.

CONTACT yieldWerx

Also Read:

CEO Interview with Elad Raz of NextSilicon

VSORA Board Chair Sandra Rivera on Solutions for AI Inference and LLM Processing

CEO Interview with Dr. Raj Gautam Dutta of Silicon Assurance

Posted on

Intelligent Networks: Power, Reliability, and Maintenance in Telecom — Webinar Preview

Intelligent Networks: Power, Reliability, and Maintenance in Telecom — Webinar Preview
by Daniel Nenni on 02-19-2026 at 2:00 pm
Categories: Analytics, Events, proteanTecs

Intelligent Networks semiwiki ads v7 400x400px

The upcoming webinar “Intelligent Networks: Power, Reliability, and Maintenance in Telecom” will focus on how telecommunications networks are adapting to growing demands for efficiency, resilience, and scalability. As telecom operators expand 5G deployments, integrate cloud-native architectures, and prepare for AI-driven services, the need for intelligent, automated network management has never been greater. This webinar aims to explore how intelligence embedded across the network can help operators address three critical challenges: power consumption, network reliability, and maintenance optimization.

One of the key topics to be addressed is power management in modern telecom networks. With increasing network densification, edge deployments, and higher-capacity equipment, energy usage has become a major operational and financial concern. The webinar will examine how intelligent networks can leverage real-time data, analytics, and automation to optimize power usage across sites and network elements. By aligning power consumption with traffic demand and operational conditions, operators can reduce energy waste, lower operational expenses, and support sustainability goals without compromising performance.

The webinar will also highlight reliability as a core requirement for next-generation telecom services. As networks support mission-critical applications, ranging from emergency communications to industrial automation, downtime and service degradation are no longer acceptable. Speakers are expected to discuss how intelligent networking technologies enable proactive reliability strategies, moving beyond traditional reactive fault management. Topics will include the use of advanced telemetry, AI-driven anomaly detection, and predictive analytics to identify potential failures before they impact service, helping operators maintain high availability and consistent quality of experience.

Another major focus of the session will be the evolution of maintenance practices in telecom environments. Conventional maintenance models, based on fixed schedules or manual inspections, can be inefficient and costly, particularly in large-scale, geographically distributed networks. The webinar will explore how intelligent networks support predictive and condition-based maintenance approaches. By continuously monitoring network health indicators such as power systems, cooling infrastructure, and hardware performance, operators can anticipate issues and intervene at the optimal time. This approach reduces unnecessary site visits, minimizes service disruptions, and extends the lifespan of critical assets.

Automation and orchestration are expected to be recurring themes throughout the discussion. As telecom networks grow in size and complexity, manual management becomes increasingly impractical. The webinar will examine how intelligent network platforms can automate routine tasks such as fault correlation, power optimization, and service recovery. Centralized visibility and intelligent orchestration enable operators to respond faster to issues, improve operational efficiency, and scale their networks with confidence.

In addition, the webinar will touch on network resilience and security, recognizing that reliability is not limited to physical infrastructure or power availability. As networks become more software-driven and interconnected, they also face greater exposure to cyber threats. Intelligent networks can enhance resilience by identifying abnormal behavior, supporting rapid mitigation, and maintaining service continuity in the face of both physical and digital disruptions.

Overall, “Intelligent Networks: Power, Reliability, and Maintenance in Telecom” is positioned to provide valuable insights into how intelligence, automation, and data-driven decision-making are shaping the future of telecom operations. By addressing power efficiency, reliability, and maintenance as interconnected challenges, the webinar will offer a holistic perspective on building networks that are more sustainable, resilient, and prepared for future demands. For telecom operators, vendors, and industry stakeholders, the session promises to outline practical strategies and emerging best practices for navigating the next phase of network evolution.

REGISTER HERE

Also Read:

Accelerating NPI with Deep Data: From First Silicon to Volume

Failure Prevention with Real-Time Health Monitoring: A proteanTecs Innovation

Thermal Sensing Headache Finally Over for 2nm and Beyond

 

Posted on

Custom IC Design using Additive Learning

Custom IC Design using Additive Learning
by Daniel Payne on 02-19-2026 at 10:00 am
Categories: EDA, Siemens EDA

Additive learning engine

Custom IC design has demanding technical requirements to produce accurate simulation results for timing and power analysis in the shortest run times. EDA vendors have been rushing to use AI and ML technology to meet these analysis requirements. I attended a webinar from Siemens on accelerating iterative design cycles with Solido additive learning techniques to understand their approach to benefit custom IC designers.

Mohamed Atousa, Product Management Manager at Siemens started with an overview of their custom IC platform with tools spanning from schematic capture, variation-aware design, physical layout, library characterization, IP QA, and SPICE simulation to generative and agentic AI.

Within the Solido Design Environment are multiple analysis tools and the focus of this webinar was on the additive learning features used in PVTMC Verifier and the High-Sigma Verifier tools to achieve a 3X to 20X speed-up on incremental and iterative runs.

High-quality IC designs require variation-aware verification flows, yet the traditional Monte Carlo simulation methods run too slowly, but using extrapolation techniques cannot find outliers or model non-Gaussian behavior, and most simulation jobs are iterative. Making minor changes in a design or PDK require re-verification, creating verification flows that are too time consuming.

The new AI technologies enable SPICE-accurate variation-aware verification that is much faster than previous approaches. In the Solido PVTMC Verifier there is full coverage verification across PVT corners plus Monte Carlo, showing speedups of 2X to 10X, allowing it to find outliers that other methods can’t. AI used in Solido High-Sigma Verifier is able to produce 6-sigma yield verification in only thousands of circuit simulations, resulting in speed-ups from 1,000X to a Billion X faster than brute-force, all while maintaining SPICE accuracy.

Jayne Alexander, Technical Product Manager at Siemens talked next about how additive learning technology retains and reuses AI models to speed-up iterative workflows by 3X – 20X. In a traditional workflow it is common to verify a design, then have to re-verify because of changes in: PDK revision, transistor sizing, simulator versions, adding more corners, etc. With the new iterative workflow there is still the first run, but subsequent verification run times are much reduced as previous results are stored for re-use.

With the Solido Additive Learning Engine you experience accurate verification results every time, automatically. Here’s the internal flow chart to achieve this benefit.

This additive learning technology retains and reuses AI models from previous jobs to speed up iterative runs, making it fast and accurate while automating the process, while requiring no user-input or AI knowledge for the tool users. Under the hood there is an AI datastore that is designed to be light-weight and optimized, supporting multiple users at once, all while using small disk space.

From Microchip we heard from Amit Bansal, Technical Staff Engineer and their old iterative design process was taking 20-30 days. They ran PVTMC Verifier on bandgap reference circuits and RC oscillators, with and without Additive Learning (AL). For the bandgap reference circuit they made a design change on the pre-layout netlist then verified at 3 sigma across 21 PVT corners. The results were 3.7X fewer simulations, 4.1X wall clock speedup:

  • Base run – 885 simulations, 2hr 16min

  • AL off – 1,170 simulations, 3hr 24 min

  • AL on – 315 simulations (3.7X), 49min (4.1X)

In the RC oscillator example they changed the trim cap value in the post-layout netlist then verified at 3 sigma across 1 PVT corner. More impressive results with 20X fewer simulations, 18X wall clock speedup:

  • Base run – 300 simulations, 14 hr 16 min

  • AL off – 300 simulations, 11hr 40min

  • AL on – 15 simulations (20X), 39min (18X)

At DAC 2025 Microchip and Siemens presented their results in the Poster Gladiator Competition and won.

Summary

With verification times taking too long for custom IC designs with many iterations there has to be a better way than using brute-force Monte Carlo simulations and then waiting for answers. Thanks to AI-powered features like additive learning, the Solido tools have been shown to reduce the number of simulations by producing faster analysis answers, all while maintaining accuracy. Users can be productive quickly with no learning curve or AI expertise required, it’s out of the box easy with minimal to no user input required.

Watch the archived webinar online here.

Related Blogs

Posted on

SiFive’s AI’s Next Chapter: RISC-V and Custom Silicon

SiFive’s AI’s Next Chapter: RISC-V and Custom Silicon
by Daniel Nenni on 02-18-2026 at 2:00 pm
Categories: Events, IP, SiFive

AI’s Next Chapter RISC V and Custom Silicon

In the rapidly evolving world of artificial intelligence and semiconductor design, open-standard processor architectures are gaining unprecedented traction. At the center of this shift is SiFive, a company founded by the original creators of the RISC-V ISA, which champions an open, extensible, and license-free alternative to proprietary architectures like x86 and Arm. A webinar titled “SiFive AI’s Next Chapter: RISC-V and Custom Silicon” encapsulates the company’s vision for how RISC-V and tailored silicon platforms will power the next wave of AI innovation, from edge devices to large-scale data centers.

The Strategic Importance of RISC-V for AI

At its core, RISC-V is a modular ISA that lets designers choose only the instruction subsets they need, and extend the base set with custom extensions suited to their applications. This openness dramatically reduces barriers to entry and enables highly specialized designs that can be optimized for power, performance, and area, crucial for AI and machine learning workloads. Unlike closed ISAs where licensing fees and fixed capabilities constrain flexibility, RISC-V allows custom silicon to be tailored from the ground up for specific AI use cases, from inference at the edge to large model training in the cloud.

A webinar on this subject would likely begin by framing why open ISAs are now receiving serious attention: as AI workloads grow in size and complexity, traditional CPU designs can become bottlenecks. Custom silicon chips designed with specific AI functions built directly into the silicon can accelerate key operations like matrix multiplication, tensor processing, and low-precision arithmetic. RISC-V’s flexible ISA makes it easier to implement such features efficiently. Moreover, as traditional leaders in processor design (like Arm) face increasing licensing constraints or strategic shifts, an open foundation like RISC-V offers an attractive alternative for companies wanting to future-proof their hardware roadmaps.

Custom Silicon: Tailoring Processors to AI Workloads

The “custom silicon” part of the webinar title refers to the creation of chips specifically architected for particular AI demands. Rather than using generic CPUs and off-the-shelf components, custom silicon can embed accelerators, optimize memory hierarchies, and integrate unique instruction extensions that speed up AI computations at lower energy consumption. In a field where efficiency and performance per watt are critical, these gains matter.

For example, SiFive’s own products, such as its Intelligence and Performance families, integrate vector and matrix computation units into RISC-V CPUs, allowing these cores to act as accelerator control units that manage AI workloads more efficiently than general-purpose processors alone. This approach drastically reduces overhead and can enable better AI performance on devices ranging from autonomous sensors to cloud servers.

Another important theme is ecosystem enablement. A custom silicon strategy only succeeds if a robust toolchain, including compilers, libraries, and runtime support, enables developers to target these designs. SiFive and partners have been building out support for major AI frameworks and compilers so that developers can deploy models efficiently on RISC-V platforms without sacrificing software compatibility or developer productivity.

Industry Collaboration and Co-Design

Webinars on RISC-V often include discussions about ecosystem partnerships and co-design approaches. For instance, recent announcements highlight collaborations between SiFive and companies such as NVIDIA, integrating technologies like NVLink to enable coherent, high-bandwidth CPU-to-accelerator communication, a major innovation for AI data centers where latency and bandwidth can dramatically impact scaling and throughput.

Similarly, the adoption of RISC-V by other ecosystem players, including major cloud providers and AI accelerator developers, underscores a broader industry shift toward heterogeneous computing architectures where CPUs, GPUs, and custom accelerators work in concert. These partnerships demonstrate how open ISAs and custom silicon are no longer niche, they are becoming central to next-generation AI infrastructure design.

Takeaways for Developers and Architects

A webinar like “SiFive AI’s Next Chapter: RISC-V and Custom Silicon” serves multiple audiences: hardware architects seeking insights on cutting-edge silicon design; software developers interested in how AI workloads can be optimized on RISC-V; and industry strategists evaluating open standard architectures against incumbent designs. Key takeaways would include:

  • How RISC-V’s modular ISA facilitates tailored processor designs for specific AI models and workloads.
  • The advantages of custom silicon in boosting performance and efficiency for AI machine learning functions.
  • Case studies or technical deep dives showing how SiFive’s RISC-V IP can be applied across edge, embedded, and data center use cases.
  • A look into emerging collaborations and ecosystem developments that broaden the practical applicability of RISC-V.

Bottom Line: This webinar represents not just a technical briefing but a reflection of a broader industry narrative: open, customizable hardware built on RISC-V is steadily transforming the AI computing landscape. As AI models continue to grow in complexity and deployment scenarios diversify, processor architectures that offer flexibility, efficiency, and extensibility, hallmarks of RISC-V and custom silicon , are set to play a foundational role in the future of AI.

Also Read:

SiFive to Power Next-Gen RISC-V AI Data Centers with NVIDIA NVLink Fusion

Tiling Support in SiFive’s AI/ML Software Stack for RISC-V Vector-Matrix Extension

RISC-V Extensions for AI: Enhancing Performance in Machine Learning

Posted on

Smarter IC Layout Parasitic Analysis

Smarter IC Layout Parasitic Analysis
by Daniel Payne on 02-18-2026 at 10:00 am
Categories: EDA, Synopsys

ParagonX flow

IC layout parasitics dominate the performance of custom digital, analog and mixed-signal designs, so the challenge becomes how to identify the root causes and to quantify the effects of parasitics during early design stages. The old method of iterating between layout, extraction, SPICE simulation, followed by manual debug and analysis is just too slow and error prone to be relied upon. A smarter approach has been developed with an EDA tool called ParagonX from Synopsys, so I attended a recent webinar to become more informed. ParagonX came from start-up Diakopto which was acquired by Ansys, then Synopsys acquired Ansys.

Rob Dohanyos of Synopsys opened up the webinar with an overview of ParagonX and then most of the time was spent in a live demo, something kind of rare with EDA vendors. The ParagonX tool can be used by a circuit designer to analyze, debug, visualize and even optimize their IC layout parasitics for any technology node and any circuit design style. Typical users of the tool are designing high-speed or high-precision circuits, like: SERDES, optical transceivers, ADC, DAC, SRAM, clocks. Parasitic sensitive designs also benefit from this analysis: Analog, power nets, PMICs, ESD networks, guard rings. Smaller nodes benefit even more from ParagonX, all the way down to 3nm. Instead of spending weeks and months debugging parasitic effects, a designer can reduce that time to hours or minutes.

Circuit designers and layout engineers will find the new tool to be easy to use out of the box, with quick run times, while providing insights to any problem areas. Here’s where ParagonX fits into your existing design and layout flow:

This tool can accept netlists that are hundreds of gigabytes in size and analyze nets that have hundreds of millions nodes or resistors, all made possible through it’s own binary netlist database. Users can expect fast netlist loading for top-hierarchy analysis, including power net analysis all while using the interactive GUI. There are six basic analysis features in ParagonX:

Invoking the tool brought up a GUI with choices grouped by analysis features, then they loaded a netlist in a few seconds. All nets could be traversed with a hierarchical widget, plus searching for nets with wildcards made it fast to find a specific net. Point to point (P2P) resistance was shown by selecting start and end points, and then the resistance contributions were displayed by layer type and percentage contribution.

The top four layers had the most resistance, so next they showed sensitivity analysis with visualization of the layout by color-coded resistance.

Each of the many functions in the ParagonX tool are actually Pythons scripts that you can customize to fit your own analysis needs. From the command line of the tool you can see each script being used for a particular function. A function called Rview shows the resistance to all points on one net, and this got 1,000 times faster in the latest tool release. Rob ran Rview on a sample net VIN, showing 240 instance pins with a resistance distribution of 100 to 450 ohms.

Next in the demo the capacitive coupling on a single net was run on net VIN and all aggressor nets, showing that VSS had the most coupling from VIN to diffusion. Metal 3 had the highest coupling on VIN with CLOCK S. Users can tell which nets and layers are causing the most coupling, both numerically and visually.

RC delay was another function demonstrated with a start point of VIN to all instance points, then a device characterization file was generated for use with a specific SPICE tool. Sensitivity of RC delay to parasitics was the next function displayed using a rainbow of colors.

The net matching function was run on a differential pair, VIN and VIP, where the user set the matching criteria and simulation results showed layout areas in red that were not matching, while areas in green were matching properly. Several other functions were demonstrated: Layout Parasitic Screener, What-if analysis, victim noise analysis, glitch analysis.

Summary

Circuit and layout designers now have a much smarter way to quickly analyze IC layout parasitics, find the root cause and even begin to optimize their designs by using the ParagonX tool in their flow. This really is a new EDA tool category and will benefit custom-digital, analog and mixed-signal design projects.

Watch the archived webinar online.

Related Blogs

Posted on

Improving Retrieval Accuracy in AI

Improving Retrieval Accuracy in AI
by Bernard Murphy on 02-18-2026 at 6:00 am
Categories: AI

Agentic RAG expert

While there are big ambitions for virtual engineers and other self-guiding agentic applications, today estimates show 83-90% of AI inferences are for internet searches. On a related note, chatbots are now said to account for nearly 60% of internet traffic. Search and support are the biggest market drivers for automation and unquestionably have improved through AI automation. Search gets closer to what you want in one pass. Chatbots also depend on retrieving domain-specific information following a question. In either case, RAG – retrieval augmented generation – plays an important role in finding the most relevant sources for a search or chat response.

Or so you hope. My experience is that the RAG results I get in a basic search/question are most useful for simple (one-part) questions in areas where I have no expertise. The more expertise I have or the more complex my question, the less useful I find the response. I can do somewhat better by adding context (You are an expert in … My question is …). Asking for citations also helps. But even these tricks don’t always work. The problem is that RAG as originally conceived (2020) has limitations. I thought it would be interesting to look at advances in this field, for variety looking from a business perspective and a healthcare perspective. In AI, it seems that our needs and priorities are not so different.

A business perspective from Elastic and Cohere

Target applications here cover a wide range in business: finance, public sector, energy, media, etc. I found this webinar which presents a combination of these two technologies with particular emphasis on RAG, the basics, challenges, and advances.

First, a quick note on RAG. LLMs are trained on publicly accessible corpora. RAG training derives information from separate and typically internal proprietary sources: PDFs, spreadsheets, images, etc. This information is chunked in some manner (eg. paragraphs in PDF text) and encoded as vectors based on similarity (scalar products of vectors, so related objects are close and unrelated objects are not). Chunks in training data must be expert (human) labeled.

Retrieval then uses a mix of keyword matching and similarity-based search to develop a top-ranked set of responses to your question. RAG is more accurate in retrieval than a general-purpose LLM because it can exploit semantic understanding based on similarity matching between a query and labeled training data.

So far this is naive RAG, with known limitations. These include struggles where the answer needed may require wider understanding of a source, or multiple sources, or the question asked has multiple clauses and requires sequential reasoning.

You know what’s coming next: agentic RAG, also called advanced-RAG. To address these limitations a system must develop a plan of attack, do multiple hops of reasoning, and self-reflect/verify after each step, potentially triggering rework. This is what agentic does. As soon as a question/request becomes even moderately complex, resolution must turn agentic, even in RAG. Tools used to support such agentic flows in business applications might be Microsoft Office, CRM, or SQL databases.

For completeness, a further advance you may find is modular RAG. These systems allow for more building block approaches to blend retrieval and refinement in structuring pipelines.

A healthcare perspective from Kent State and Rutgers

Here I draw on a long but very interesting paper. The authors suggest the following as key applications in healthcare: diagnostic assistance by retrieving information on similar cases; summarizing health records and discharge notes; answering complex medical questions; educating patients and tailoring responses to user profiles; matching candidates to clinical trials; and retrieving and summarizing biomedical literature, especially recent literature, in response to a clinical or research query.

The authors note a range of challenges in retrieving information. Obviously such a system must handle a wide range of data types (modalities), from doctor notes to X-rays, EKG traces, lab results, etc. They must also contend with a wide range of potentially incompatible health record sources, some with technically precise notes (myocardial infarction), some less precise (heart attack). Users face challenges in understanding the credibility of sources (media health articles, versus Reddit, versus respected journals in a field) and how these contribute to ranking conclusions. Familiar challenges even in our field.

There is a longer list from which I’ll call out one widely relevant item: the need to continuously update as new research, drugs and treatments emerge, also need to deprecate outdated sources. In a medical context the authors suggest that manual updates would be too slow and error-prone and that any useful RAG system for their purposes must build continuous update into the system.

They look at tradeoffs between the three RAG architectures mentioned earlier (naive, advanced, and modular). They find naive RAG easy to setup and use, though for their purposes too noisy and risky for high-stakes scenarios. Advanced RAG is more promising in diagnostic support and EHR summarization, striking a balance between factual grounding and speed, but requires significant compute resource (presumably an on-prem datacenter). This method looks most ready today for clinical use, at least in hospitals and large clinics. They see modular RAG as interesting for ongoing research, though training and resource costs make it impractical to consider for near-term deployment.

Relevance to design automation

Accuracy is critical for technical support in our domain, whether internal or external. Our users are very knowledgeable and intolerant of beginner-level suggestions. Experiences above suggest that advanced/agentic RAG may be the most appropriate method to deploy support here.

That guidance should aim to avoid mistakes made in some ambitious all-AI rollouts (Klarna customer support for example). These certainly should include emphasis on “don’t know” for suggestions with low support, explainability for top candidate response offered, and methods to escalate to a human expert when the bot is uncertain. I am starting to see some of this in general customer support.

Meantime, agentic RAG can make a big difference in productivity and user satisfaction for in-house and external users. Most of us would prefer to explore on our own supported by effective agentic RAG, only turning to a human expert when we’re not making progress. That’s technology worth supporting.

Also Read:

Bronco Debug Stress Tested Measures Up

TSMC and Cadence Strengthen Partnership to Enable Next-Generation AI and HPC Silicon

How Memory Technology Is Powering the Next Era of Compute

Posted on

Ceva IP: Powering the Era of Physical AI

Ceva IP: Powering the Era of Physical AI
by Daniel Nenni on 02-17-2026 at 2:00 pm
Categories: Ceva, IP

Ceva IP Powering the Era of Physical AI

Artificial intelligence is rapidly moving beyond the digital domain and into the physical world. From autonomous robots and smart factories to intelligent vehicles and connected consumer devices, AI systems are increasingly expected to perceive their surroundings, make real-time decisions, and act on them instantly. This shift marks the rise of Physical AI, a new generation of intelligence that combines sensing, connectivity, and on-device inference. At the heart of this transformation is Ceva IP, a leading provider of semiconductor and software intellectual property that enables Physical AI at the edge.

Physical AI differs from traditional cloud-centric AI in a fundamental way. Rather than relying on remote servers to process data, Physical AI systems operate directly on the device. They gather data from sensors such as cameras, microphones, radar, and motion detectors, analyze it locally using AI inference engines, and respond in real time. This approach is essential for applications where latency, power efficiency, reliability, and privacy are critical. Ceva’s technology portfolio is purpose-built to address these exact requirements.

Ceva’s role in Physical AI begins with connectivity, a foundational pillar for intelligent devices. The company offers a comprehensive suite of wireless IP solutions, including Bluetooth, Wi-Fi, ultra-wideband, and cellular technologies. These connectivity solutions allow devices to communicate seamlessly with other machines, infrastructure, and ecosystems while maintaining low latency and high reliability. In Physical AI systems, fast and robust wireless communication ensures that sensor data, control signals, and AI-driven decisions flow without interruption, even in complex or dynamic environments.

Equally important is sensing and sensor fusion, another core strength of Ceva IP. Physical AI systems rely on multiple sensors to understand the real world accurately. Ceva provides DSP architectures and software frameworks optimized for handling multimodal sensor data. By efficiently combining inputs from vision, audio, motion, and radar sensors, Ceva-powered systems gain a richer, more contextual understanding of their surroundings. This capability is crucial for applications such as robotics, advanced driver-assistance systems (ADAS), and industrial automation, where precise perception directly impacts safety and performance.

At the center of Physical AI lies on-device inference, and this is where Ceva’s neural processing technologies play a decisive role. Ceva’s AI and neural network IP solutions are designed to deliver high performance at ultra-low power, enabling complex AI workloads to run directly on edge devices. These inference engines support a wide range of machine learning models, from classical AI algorithms to modern deep learning networks. By performing inference locally, devices can respond instantly to real-world events without depending on cloud connectivity, while also protecting sensitive data.

One of Ceva’s key differentiators is its system-level approach. Rather than offering isolated components, Ceva provides an integrated IP ecosystem that allows semiconductor designers to build complete Physical AI platforms. Connectivity, sensing, and inference IP blocks are designed to work together efficiently, reducing integration complexity and accelerating time to market. This modular and scalable approach enables customers to tailor solutions for diverse applications, from tiny battery-powered IoT devices to high-performance automotive and industrial systems.

The impact of Ceva IP is already visible at scale. Ceva technologies have been deployed in tens of billions of devices worldwide, underscoring the company’s influence across consumer, automotive, industrial, and mobile markets. As the demand for smarter, more autonomous systems continues to grow, Physical AI is becoming a defining trend in the semiconductor industry.

Bottom line: Ceva IP is powering the era of Physical AI by providing the essential building blocks that allow machines to sense, connect, think, and act in the real world. By enabling real-time intelligence at the edge, Ceva is helping bridge the gap between digital computation and physical interaction—unlocking a new generation of responsive, efficient, and intelligent devices.

CONTACT CEVA-IP

Also Read:

Ceva-XC21 Crowned “Best IP/Processor of the Year”

United Micro Technology and Ceva Collaborate for 5G RedCap SoC and Why it Matters

Ceva Unleashes Wi-Fi 7 Pulse: Awakening Instant AI Brains in IoT and Physical Robots