EDN Network

Silicon Valley-based tech incubator Plug and Play, which launched its semiconductors program in 2023, has joined hands with Synopsys to provide selected startups with access to its design tools and software licenses. The tie-up aims to lower the barrier to next-generation chip design, accelerate time to market, and reduce development costs across various chip-design stages.

Professional-grade EDA tools often represent a significant investment at the early stage, so this partnership will lower the barrier to entry for cutting-edge chip design. “This collaboration with Synopsys brings their trusted and comprehensive chip design software and solutions directly to the startups we support,” said Rouzbeh Borhani, head of Plug and Play Semiconductors.

The partnership also marks a significant move by Synopsys to win over chip startups while helping them innovate at the earliest stages of chip development. Moreover, Plug and Play can provide Synopsys access to deep tech firms beyond semiconductor startups.

Plug and Play has recently welcomed over 150 startups in its first Silicon Valley batch of 2025. These upstarts will gain hands-on support from industry experts, investors, and corporate partners in a 3-month program, which includes mentor-led workshops, private deal-flow sessions, and curated introductions designed to accelerate their business growth.

The program will also allow startups to secure pilots, proofs of concept, and potential investment without equity requirements. In this batch, 13% of startups are based in Silicon Valley, while 61% are from other parts of the United States. The remaining 26% of startups come from Canada, the United Kingdom, Hong Kong, and several other countries.

Related Content

• Silicon 100: Startups Worth Watching in 2024
• In search of new chip startup funding models
• 10 Chip Startups for AI in Edge and Endpoint Applications
• Startups eye opportunities amid analog industry consolidation
• Inference Sensing and In-Memory Computing Chip Startup Claims 20 TOPS/W

The post Tech incubator leasing EDA tools for chip startups appeared first on EDN.

How likely is it that an AI system will be trained on incorrect information? Please consider the following example.

Thomas Edison achieved a lot of things during his lifetime, but now and then, he was off base. When Edison invited Nikola Tesla to see his new phonograph, and Tesla immediately saw room for improvements, Edison was enraged, and the two of them got into a lifelong feud. Edison’s conflict with George Westinghouse about DC versus AC for public electric power systems was another example where Edison went _______ (fill in the blank for yourself). I’ve read Edison biographies, which pointed out his strengths and his flaws, but those writings seem to have vanished from the internet. Writings today seem only to extol Edison almost as a deity.

I once read that when Edison was trying to choose a material for the filament of his electric light bulb, he was focused on what we would today call low-resistivity materials. Those filaments either didn’t work at all or had very short operating lifetimes. When he was told of Georg Simon Ohm and the fact of Ohm’s Law, Edison spurned that input and announced instead that he was not going to be limited by “Ohm’s silly law.” Eventually, though, he went to the carbon filament with its high resistivity and succeeded. Ohm’s law led him to that success.

Somewhere in the past, I came across a written item to the above effect about Edison, although recently, my trying to find that reference has been an exercise in futility (that deity thing again). However, something did come up during my searches, which I found very disturbing indeed (Figure 1).

Figure 1 Disturbing Google search results when trying to find a reference about Edison and his rejection of Ohm’s law.

When I entered a search term of Edison’s supposed “silly law” remark into Google, some of the search results that came up were full of stuff that was just plain, flat out, totally, utterly, and completely WRONG!

Please take a close look at the three images above and read them carefully. They are denials of the validity and applicability of Ohm’s Law. Vital concepts, as rudimentary as static resistance versus dynamic resistance and thermal coefficients of resistance, are simply ignored. At one point, we read “… the light bulb is not ohmic and does not behave like a resistor.”

The numerous mis-statements in these three screenshots are utterly appalling.

Now, consider some AI system getting trained with the inclusion of the above. That AI system is going to be capable of producing erroneous results. It is going to be capable of “hallucinating” and anyone taking those results as valid will have been grievously misled.

In spite of popular misconceptions and loads of public hype, AI doesn’t think. AI only regurgitates with formatting whatever can be found during its “training”. AI at its best is only an information retrieval tool, not a thinking entity with any level of judgment. Judgment still has to come from human brains.

Scary, isn’t it?

John Dunn is an electronics consultant, and a graduate of The Polytechnic Institute of Brooklyn (BSEE) and of New York University (MSEE).

 Related Content

• The real danger of artificial intelligence
• Is Generative AI Creating a New Three-Body Problem?
• AI vs. AI: Both Friend and Foe in Cybersecurity
• The Unchecked Specter of ‘Anti-Social’ Media

The post AI and the problem of misinformation   appeared first on EDN.

Pick and place equipment is instrumental in many facilities for raising overall output and reducing errors. However, electronics design engineers know that standardization and modularization are among the top requests from decision-makers who use or might soon use this equipment.

Here’s how design professionals can address those standards in pick and place equipment.

After failed efforts to standardize or modularize existing pick and place robots, people may think they should start from scratch to get things right. However, that is not necessarily the case if designers, engineers, and others involved with these projects focus on the most pressing issues and prioritize solving them.

That was the approach for a federal initiative working to improve the Department of Veterans Affairs prescription-fulfillment project. The initial aim was to use 20 modular units, each designed for a specific task.

However, those machines failed numerous tests, and the entity first involved in designing and installing them never finished the project. After assessing matters and implementing solutions, a robotics vendor improved the suction methods that allowed robots to pick up pills and installed funnels around each one to catch dropped pills.

The vendor also implemented technology to train the system to recognize new pill types and shapes, making it future proof. This example shows how electronics design engineers, who identify the most significant issues and determine the best solutions, can get meaningful results and save time and money.

Standard recommendations for training and maintenance

Industrial decision-makers can lengthen the usefulness of pick and place equipment by keeping it well-maintained and ensuring operators know how to use it. These assets are similar to other specialty machines that need ongoing oversight to perform reliably. For example, many laser cutters have auto-focus features and real-time monitoring. Those capabilities tighten quality control but do not remove the need for supervision.

Standardization occurs when managers create training processes, checklists, operator certifications, and other mechanisms to prevent costly mistakes. However, electronics design engineers can influence those resources by suggesting the content of the manuals, warranties, and materials people receive when purchasing pick and place machines.

These engineers can also create guidelines about when to perform specific maintenance measures and stipulate the conditions that may require more frequent maintenance. Those insights can help people standardize internal processes to optimize their equipment.

Traditional standardization opportunities come from purchasing equipment that meets specific industrial standards, such as those associated with pharmaceutical clean rooms. However, some companies purchase automated equipment from a single brand. Although that approach allows standardization, it limits future flexibility.

For instance, recent microelectronics shortages have required creativity to fulfill industrial automation plans. Challenges like this mean standardizing industrial processes may be the best option.

Pursue modularization to enable growth

Electronics design engineers may work for clients who want specific pick and place integrations that address current needs while anticipating future requirements. In those cases, standardization may become limiting, but modularization could create the flexibility required to meet new needs as they emerge.

S&S Activewear’s distribution center is a real-life application of that option. The company uses hundreds of autonomous mobile robots to bring items from shelves to workers. The upgrade saves the employees from time-consuming tasks, like walking up and down aisles to find the desired products.

Executives also recognized how pick and place equipment fit into their automation goals. Although picking was once largely manual, specialty equipment has optimized the task and boosted productivity. The company now has 50 modular picker workstations, increasing its initial amount by over 50%. Workers do all their picking there after robots bring them the goods. Additionally, workers place all the picked goods directly into shippable cartons, shortening the process.

This example shows the positive results that can happen when clients, designers, and other concerned parties focus on expansion potential from the beginning. Even if decision-makers are unsure how much their operations might grow, modular installations give them numerous options to implement later.

Increase pick and place equipment adoption

These suggestions can help electronics design engineers use their expertise and problem-solving strategies to encourage industrial leaders to bring pick and place machines into their facilities for the first time. The more decision-makers view automated equipment as aligned with their processes, the likelier they will be to use it.

Ellie Gabel is a freelance writer as well as an associate editor at Revolutionized.

Related Content

• Exploring the Future of Industrial Automation
• Different Sensor Technologies Behind Autonomous Mobile Robots
• Autonomous mobile robot (AMR) design revisited with AI algorithms
• Future of Manufacturing: How to Select Mobile Robots for Heavy Loads
• ‘Wi-Fi on Steroids’ Technology Allows Manufacturing Industry to Finally Cut the Cord

The post Standardization and modularization in pick and place equipment appeared first on EDN.

The R&S UDS series of digital multimeters (DMMs) offers 5.5-digit and 6.5-digit resolution, with the 6.5-digit model delivering a basic DC accuracy of 0.0075%. Replacing the HMC8012, the UDS DMMs provide higher accuracy and an updated user interface to simplify testing.

Streamlining test workflows, UDS models can display up to three measurements simultaneously—such as DC, AC, and statistical data—on a 3.5-inch OVGA color display. They support voltage ranges up to 1000 VDC and 750 VAC, with a current capacity of 10 A.

With a wide range of measurement functions and remote-control interfaces, the multimeters are well-suited for troubleshooting, component testing, and system validation. They also fit well in teaching labs and production environments. In addition to 12 standard measurement functions, the units offer statistical and math capabilities. Interfaces include USB, Ethernet LAN, and IEEE 488 (GPIB) for SCPI-based control.

Configure and request a quote for the UDS digital multimeter using the product page link below.

UDS series product page

Rohde & Schwarz 

The post DMMs improve accuracy and usability appeared first on EDN.

Infineon’s radiation-hardened GaN HEMT is the first in-house manufactured device qualified to the Joint Army Navy Space (JANS) MIL-PRF-19500/794 specification—the highest quality certification issued by the U.S. Defense Logistics Agency (DLA). The company’s new family of radiation-hardened CoolGaN transistors is designed for mission-critical applications in on-orbit spacecraft, manned missions, and deep space probes.

The first three devices in the GaN lineup are 100-V, 52-A transistors with a typical RDS(on) of 4 mΩ and a total gate charge of 8.8 nC. Housed in hermetically sealed ceramic surface-mount packages, they are hardened against Single Event Effects (SEE) up to a Linear Energy Transfer (LET) of 70 MeV·cm²/mg using gold (Au) ions. Two of the devices, while not JANS certified, are screened to Total Ionizing Dose (TID) levels of 100 krad and 500 krad. The third device, also screened to 500 krad, meets the rigorous JANS MIL-PRF-19500/794 qualification.

Engineering samples and evaluation boards are available now, with the final JANS-qualified device set for release in summer 2025. Additional JANS parts will launch soon, expanding the range of available voltage and current ratings. For more information on Infineon’s rad-hard GaN transistors, click here.

Infineon Technologies 

The post GaN HEMT earns DLA-JANS certification appeared first on EDN.

The EA-BIM 20005 battery impedance meter from Tektronix uses electrochemical impedance spectroscopy (EIS) across 20 channels to characterize Li-Ion battery cells. With industry-standard interfaces and a compact 19-in., 3U form factor, it integrates easily into automated battery test systems.

EIS testing provides deep insight into cell quality. Offering broad-frequency capabilities from 1 mHz to 10 kHz and AC stimulus up to 10 A (peak-to-peak), the EA-BIM-20005 supports comprehensive analysis of battery cell behavior under varying conditions. It is well-suited for cylindrical, pouch, and prismatic cells.

Each of the meter’s 20 EIS channels is paired with a 4-wire PT100 temperature channel, enabling simultaneous tracking of impedance and cell temperature. An integrated DC power supply delivers up to ±1 A at 5 V for EIS measurements during cell charging and discharging. Included PC software offers built-in visualizations and analysis tools, while connectivity is provided via a USB port and two CAN bus interfaces.

Access the datasheet and request a quote for the EA-BIM 20005 battery impedance meter using the product page link below.

EA-BIM 20005 product page 

Tektronix

The post Multichannel impedance meter analyzes Li-Ion cells appeared first on EDN.

Two integrated gate drivers from ST provide programmable current control for driving three-phase brushless motors in consumer and industrial equipment. Operating from 6 V to 50 V, the STDRIVE102H supports single-shunt control, while the STDRIVE102BH handles three-shunt control—both configured via two analog pins. Each device can source up to 1 A and sink up to 2 A.

A simple resistor divider programs the gate-drive current, allowing the triple half-bridge drivers to control six external N-channel MOSFETs. This helps optimize power stage performance, including control of switching slew rate, without the need for discrete gate resistors. An integrated charge pump powers the three high-side drivers, enabling continuous on-time for the high-side MOSFETs.

To accelerate development with the STDRIVE102H and STDRIVE102BH gate drivers, the EVLDRIVE102H and EVLDRIVE102BH evaluation boards support field-oriented and six-step control, featuring onboard back-EMF sensing and inputs for position sensors. Standard headers connect to STM32 Nucleo boards, and the X-CUBE-MCSDK toolkit provides the necessary software and code.

The STDRIVE102H and STDRIVE102BH are available in 5×5-mm or 6×6mm QFN packages, priced from $1.20 each in lots of 1000 units.

STDRIVE102H product page 

STDRIVE102BH product page

STMicroelectronics

The post Motor gate drivers enable flexible current control appeared first on EDN.

Comtrend’s PM-1540 powerline data module uses MaxLinear’s G.hn (data-over-powerline) chips to support backend communication in EV charging stations. It transmits data from power meters over existing electrical wiring, eliminating the need for dedicated communication cables, and can also extend connectivity to backend systems in data centers or smart parking environments.

By leveraging existing electrical wiring, the PM-1540 delivers lower latency, higher speeds, and more stable performance than conventional methods. It enables real-time connectivity while reducing costs compared to LAN, Wi-Fi, or 4G systems. The module supports up to 250 nodes within the same powerline domain and transmits signals over distances up to 700 meters, with up to 16 levels of signal repetition for extended reach.

MaxLinear’s G.hn baseband processors and analog front-end chipsets provide reliable, low-latency connectivity over existing wiring, delivering physical data rates up to 2 Gbps with full ITU compliance. Their support for Quality of Service (QoS) and broad media compatibility—including powerline—makes them well-suited for EV charging infrastructure, enabling seamless interoperability and cost-effective deployment.

For detailed information on Comtrend’s PM-1540 G.hn powerline module, click here. An overview of MaxLinear’s G.hn solutions can be found here.

Comtrend

MaxLinear

The post Powerline module enables EV charger data links appeared first on EDN.

USB Power Delivery (USB-PD) now offers faster, more efficient, and more versatile power handling solutions. As we can all see, it’s an exciting advancement that significantly enhances the capabilities of USB connections.

This mechanism uses the USB configuration channel (CC) to allow a device to request a specific voltage. While this might seem complex at first, it’s pretty easy to utilize in practice.

Figure 1 The module has several jumpers to set the DC output voltage at multiple levels. Source: Author

What makes it easy nowadays is that we can buy compact USB-PD Trigger/Decoy modules that do the complicated background tasks for us (Figure 1). You can see such a module has a number of jumpers to set the DC output voltage to 5 V, 9V, 12 V, 15 V or 20 V.

This module acts as a trigger or decoy to request specific power profiles from USB-PD power sources such as USB-C chargers, power banks, and adapters. So, with this module, you can trigger USB-PD protocols and thus, for example, charge your laptop via a PD-capable USB-C power supply.

Note at this point that a USB-PD Trigger, sometimes called a USB-PD Decoy, is a small but clever circuitry that handles the USB-PD negotiation and simply outputs a predefined DC voltage.

Some USB-PD Trigger/Decoy modules are adjustable with a selector switch, or cycle among voltages with a pushbutton press, while others deliver a fixed voltage, or will have solder jumpers (or solder pads to install a fixed resistor) to select an output voltage. The output connection points on these modules are typically just two bare solder pads, or small screw terminals in certain cases (Figure 2).

Figure 2 The output connection points are shown on the modules. Source: Author

For just a few bucks each, these smaller and slenderer USB-PD Trigger/Decoy modules are useful to have in your tool chest, both for individual projects and for use in a pinch. In my view, for most applications, the fixed voltage type power provider is preferable, as this prevents accidental slips that could destruct the power consumer.

I recently bought a set of these fixed voltage modules. As you can see, the core part of these single-chip modules is the IP2721 USB Type-C physical layer protocol IC for USB Type-C input interfaces.

Figure 3 IP2721 is a USB Type-C PD protocol IC for USB input port that supports USB Type-C/PD2.0/PD3.0 protocols. Source: Author

The USB Type-C device plug-in and plug-out process is automatically detected based on CC1/CC2 pins. The chip has an integrated power delivery protocol analyzer to get the voltage capabilities and request the matched voltage.

Figure 4 The schematics shows a design use case built around the USB Type-C PD protocol IC. Source: Injoinic Technology

Surprisingly, the newly arrived module—designed for a single, fixed-voltage output—features the IP2721 controller in a bare minimum configuration without the power-pass element.

Figure 5 The module features the IP2721 controller in a bare minimum configuration. Source: Author

Hence, the output voltage will be whatever VBUS is, and this could be 5 V during initial enumeration or stay at this voltage in case negotiations failed. Luckily, for many applications, this will not be much of an issue. But on paper, to comply with the USB power delivery specifications, the device is supposed to have a high-side power MOSFET as the power-pass element to disconnect the load until a suitable power contract has been negotiated.

For this writing, I needed to test the output of my module. So, below you can see a little snap taken during the first test of my IP2721 USB-PD trigger 9-V module; nothing but the process of testing the module with a compatible power source and a DC voltmeter.

Figure 6 DV voltmeter shows the output of the IP2721-based USB-PD module. Source: Author

Here are some final notes on the power delivery.

• USB-PD is a convenient way of replacing power supply modules in many electronics projects and systems. Although USB-PD demands specialized controller chips to be utilized properly, easily available single-purpose USB-PD Trigger/Decoy modules can be used in standalone systems to provide USB-PD functionality.
• Interestingly, legacy USB can only provide a 5-V power supply, but USB-PD defines prescriptive voltages such as 9 V, 15 V, and 20 V in addition to 5 V.
• Until recently, the USB-PD specification allowed for up to 100 W (5 A@20 V) of power, called Standard Power Range (SPR), to flow in both directions. The latest USB-PD specification increases the power range to 240 W (5 A@48 V), called Extended Power Range (EPR), through a USB-C cable. So, if a device supports EPR expansion commands, it can use 28 V, 36 V, and 48 V.
• Since the most recent USB-PD specification allows to realize up to 240 W power delivery through a single cable, it’s possible to provide ample power over USB to multiple circuit segments or devices simultaneously.
• Electronic marking is needed in a Type-C cable when VBUS current of more than 3 A is required. An electronically marked (E-Marked) cable assembly (EMCA) is a USB Type-C cable that uses a marker chip to provide the cable’s characteristics to the Downstream Facing Port (DFP). It’s accomplished by embedding a USB PD controller chip into the plug at one or both ends of the cable.
• The USB-PD Programmable Power Supply (PPS) was implemented with USB PD3.0. With PPS, devices can gradually adjust the current (50-mA steps) and voltage (20-mV steps) in the range from 5 V to 20 V. PPS can directly charge a battery, bypassing the battery charger in a connected device.
• Adjustable Voltage Supply (AVS) was implemented with USB PD3.1 and extended with PD3.2, allowing it to work within SPR below 100 W, down to a minimum of 9 V. AVS is similar to PPS in terms of function, but the difference is that it does not support current-limit operation, and the output voltage is adjusted in 100-mV steps in the range from 9 V to 48 V.

Note that USB-PD, which is combined with USB-C, takes full advantage of the power supply and multi-protocol functions over USB-C. Implementing USB-C for portable battery-powered devices enables them to both charge from the USB-C port as well as supply power to a connected device using the same port.

So, devices using a single or multicell battery charger can now be paired with a USB-C or USB PD controller, which enables the applications to source and sink power from the USB-C port. Below is an application circuit based on MP2722, a USB Type-C 1.3 compliant, highly integrated, 5-A, switch-mode battery management device for a single cell Li-ion or Li-polymer battery.

Figure 7 The application circuit is built around a 5 A, single-cell buck charger with integrated USB Type-C detection. Source: Monolithic Power Systems (MPS)

In the final analysis, it’s important to recall that the USB-PD is not just about the power delivery-related negotiations. Feel free to comment if you can help add to this post or point out issues and solutions you have found.

T. K. Hareendran is a technical author, hardware beta tester, and product reviewer.

Related Content

• Understanding USB Power Delivery 3.2
• USB Type-C PD 3.0 Specification, Charging and Design
• Providing USB Type-C connectivity – What you need to know
• USB Power Delivery: incompatibility-derived foibles and failures
• USB-C PD 3.1 EPR: A Full System Design Solution, Wall to Battery

The post A quick and practical view of USB Power Delivery (USB-PD) design appeared first on EDN.

Frequent contributor Nick Cornford recently assembled an ensemble of cool circuit designs incorporating linear-in-pitch VCOs (LPVCOs). 

• “A pitch-linear VCO, part 1: Getting it going”
• “A pitch-linear VCO, part 2: taking it further”
• “Revealing the infrasonic underworld cheaply, Part 1”
• “Revealing the infrasonic underworld cheaply, Part 2”

These elegant and innovative designs (standard fare for Nick’s contributions) were perfectly adequate for their intended applications. Nevertheless, it got me wondering how difficult it would be to implement an LPVCO with a range covering the full 10-octave audio spectrum, from 20 Hz to 20 kHz. I even decided to try for extra credit by going for a tri-wave output suitable for direct drive of one of Nick’s famous squish-diode sine converters. Figure 1 shows the result.

Figure 1 An LPVCO with 10-octave (20 Hz to 20 kHz) tri-wave output comprises antilog pair Q1 and Q2, two-way current mirror Q3 and Q4, integrator A1b, comparator A1a, and buffer A1c. Resistors R1 and R2 are precision types, and T1 is a Vishay NTCSC0201E3103FLHT (inhale!).

Wow the engineering world with your unique design: Design Ideas Submission Guide

Vin is scaled by the tempco-compensating voltage divider, ((R1+T1)/R2 + 1) = 28:1, and applied to the Q1 and Q2 antilog pair, where Q1 level shifts and further temperature compensates it. Then, with the help of buffer A1c, it’s antilogged and inverted by Q2 to produce Ic2 = 2(2Vin) µA = 1 µA to 1 mA for Vin = 0 to 5v.

From there, it goes to the two-way current mirror: Q3 and Q4. A description of how the TWCM works can be found here in “A two-way mirror—current mirror that is.”

The TWCM passes Ic2 through to the integrator A1b if comparator A1a’s output is zero, and mirrors (inverts) it if A1a’s output is high. Thus, A1b ramps up if A1a’s output is at 0v, and down if it’s at 5v, resulting in sustained oscillation.

The C1 timing ramp has a duration in each direction ranging from 25 ms (for Vin = 0) to 25 µs (for Vin = 5v). The triangular cycle will therefore repeat at Fosc = 2(2Vin)µA/(25nCb)/2 = 20(2(2Vin) ) Hz.

 So, there’s the goal of a tri-wave LPVCO with an output span of 20 Hz to 20 kHz centered at 640 Hz, and it wasn’t so terribly messy to get there after all!

My thanks go to Nick Cornford for introducing the LPVCO to Design Ideas (DIs), and to Christopher Paul and Andy I for their highly helpful simulations and constructive criticisms of my halting steps to temperature-compensating antilogging circuits. I also thank editor Aalyia Shaukat for her DI environment that makes such teamwork possible for a gang of opinionated engineers, and mostly accomplished without actual bloodshed! 

Mostly.

Stephen Woodward’s relationship with EDN’s DI column goes back quite a long way. Over 100 submissions have been accepted since his first contribution back in 1974.

Related Content

• Seven-octave linear-in-pitch VCO
• A pitch-linear VCO, part 1: Getting it going
• A pitch-linear VCO, part 2: taking it further
• Revealing the infrasonic underworld cheaply, Part 1
• Revealing the infrasonic underworld cheaply, Part 2

 

The post 10-octave linear-in-pitch VCO with buffered tri-wave output appeared first on EDN.

My audio gear

Back in July 2019, I told you about the combo of Massdrop’s x Grace Design Standard DAC:

and its companion Massdrop Objective 2 Headphone Amp: Desktop Edition (Massdrop is now just Drop, by the way, and is now owned by Corsair):

that I’d recently acquired for listening to computer-sourced audio over headphones in a quality-upgraded fashion beyond just the DAC (and amp-fed headphone jack) built into my Mac. That same two-device stack:

remains on my desk and to my right to this very day, albeit subsequently joined by an even higher quality balanced audio stack to my left:

combining a Topping D10 Balanced DAC:

and a Drop + THX AAA 789 Linear Headphone Amplifier:

The Monolith

But I digress. Today’s dissection showcase is of none of these. Instead, I’ll be analyzing the guts of the Monolith by Monoprice Liquid Spark Headphone Amplifier by Alex Cavalli:

It’s comparable in size to the Massdrop Objective 2 Headphone Amp: Desktop Edition I mentioned at the beginning of the writeup:

And what about a companion digital-to-analog converter? That’s a story all by itself. It originally had one, the unsurprisingly named “Monolith by Monoprice Liquid Spark DAC by Alex Cavalli”:

based on an Asahi Kasei Microdevices (AKM) Semiconductor DAC chip. However, in October 2020, just as COVID was in general throttling the tech economy, audio equipment suppliers got a double-whammy: a massive three-day fire at AKM’s semiconductor facility in Japan, which clobbered its output. Some AKM customers, such as Fiio and Schiit (the latter, for example, redesigning and renaming its Modi 3 DAC as the Modi 3e, with “e” short for ESS Technology), redesigned their systems to use chips from other suppliers instead. Others, like Monoprice, threw in the towel. That said, since the Monoprice Liquid Spark headphone amp has conventional RCA (unbalanced) analog line inputs, you can use it with any standard DAC.

A bit of background info

A few definitions before proceeding with the dissection. “Monolith” is Monoprice’s audio products brand. Alex Cavalli is a now-retired, well-known audio amplifier designer who, in addition to selling both self-branded Cavalli Audio equipment (now repaired by Avenson Audio since his retirement) and gear branded by Monoprice (obviously) and Massdrop/Drop, also published complete design documentation sets for others to use in building their own gear, DIY style. Alex is a contemporary of another audio amplifier “wizard” whose name may be more familiar to you: Nelson Pass.

And finally, why do I categorize it as being for “entry-level audiophiles”? The feature set, for one thing. I’ve already noted that it doesn’t offer balanced inputs and outputs, for example, the magnitude-of-benefits of which are debatable, anyway. That said, unlike the Massdrop Objective 2 Headphone Amp: Desktop Edition, it does include preamp outputs, the benefit of which I’ll elaborate on shortly. And its performance is nothing to sneeze at:

• Manufacturer-published spec sheet (PDF)
• Manufacturer-published test measurement graphs (PDF)
• Independently published testing results (Audio Science Review)
• Independently published testing results (Super Best Audio Friends)

And the price, although that’s an imperfect-at-best barometer of quality. That said, Schiit’s current high-end solid-state Mjolnir 3 headphone amp (the company also sells tube-based products) goes for $1,199-$1,299, depending on color. Conversely, when the Liquid Spark Headphone Amplifier was introduced in 2018 (a year after Alex Cavalli announced his retirement, interestingly), Monoprice sold it for $99. Its list price is now $129. But (in explaining how I first came across it) I’ve long subscribed to Monoprice’s periodic promotional emails, and back in March of last year, I stumbled across a smokin’ deal; $32.49 each plus a further 25%-off discount. I bought two at $50.29 total (with tax), one for a buddy’s birthday, the other for me.

What I’ll be taking apart today is neither of these devices, however. Last October, while searching for a Liquid Spark DAC mate to my headphone amplifier, I stumbled across “as-is” Liquid Spark amps on eBay for $29.99 plus tax and $9.99 for shipping. The seller notes said:

Pulled from a professional working environment. Tested for power, no further testing was done. Due to lack of knowledge and having the proper equipment to fully test these units, we are selling AS-IS for parts/not working. Unit shows some signs of scuffs/scratches all around the unit. Please refer to the photos for more detailed information on the cosmetic condition.

As I’ve mentioned (and exemplified) many times before, such “for parts only” listings are perfect for teardown purposes. I ended up getting one for $20.99 (plus the aforementioned sales tax and shipping). I’m not sure how the seller “tested for power”, since it didn’t come with the requisite “wall wart”. And as you may have already noticed from the back panel “stock photo” shown earlier, it’s an uncommon one, outputting 36V at 1.25A min (that said, at least it’s got a DC output; Schiit’s are all just AC transformers).

Overview

I have no idea if this one actually works, and I’m not going to chance zapping my personal amplifier’s functional PSU to find out. That said, here it is in all its cosmetically imperfect glory, as usual accompanied by a 0.75″ (19.1 mm) diameter U.S. penny for size comparison purposes (the unit has dimensions of 4.6″ x 3.7″ x 1.5″/117 x 94 x 38 mm and weighs 9.6 oz./271g):

Left-to-right are the power switch, a ¼” TRS headphone jack, a 3-or-6 dB gain switch (to accommodate headphones of varying impedance), and a rotary volume control knob. Now, about that back panel (following up on my earlier “teaser” comment about the RCA output set):

Most headphone amps in this price range have unbalanced RCA inputs, but their only output is an unbalanced TRS headphone jack (of varying diameter) up front. But this one also has a pair of unbalanced RCA outputs. And they’re not just simple line level pass-throughs, either; they route through the internal preamplifier first, although they (obviously) then bypass the headphone amplifier circuitry. Why’s this nice? Well, you can connect them to an external power amplifier to drive a set of speakers from the same audio source. And, because the preamp is still in the loop, the headphone amp’s volume control manages speaker volume, too.

The one thing I don’t know (and haven’t tested with my unit) yet, and the user manual doesn’t clarify, is whether the two output sets operate simultaneously or (as is the case with Schiit’s device equivalents) in a one-or-the other fashion. Said another way, when you plug in some “cans”, does this also mute the sound that would otherwise come out the connected speakers?

Onward: the left and right sides:

The top:

and bottom:

complete with a label closeup:

I admittedly enjoyed fondling this device, both in an absolute sense and relative to the scores of predominantly plastic-based products I’ve taken apart in the past (and will undoubtedly continue to do so in the future). It’s heft…the solidity of its all-metal construction…very nice!

Teardown time

Speaking of that all-metal construction, let’s start by getting the front panel off, starting with the Torx head screws on both ends:

Part of the way there…

Let’s see what’s behind that volume knob, which was snugly attached but pulled off after a bit of muscle-powered coercion:

Unscrew the nut, remove the washer:

and this part of the total task is successfully completed:

Check out that gloriously thick and otherwise solid PCB!

Now for the back panel. Six screws there and another one below:

And the panel-still-attached PCB slides out the rear:

Voila!

Jumping forward to the future for a moment, I went back and perused the product page after the teardown-in-progress and found this, which I hadn’t noticed earlier:

It wasn’t surprising. It was, conversely, validating. Truth be told, even before I took this amp apart, I’d suspected I’d encounter a discretes- (vs op amp-) based design. And when I saw the horde of tiny ICs scattered all over the top of the PCB, my in-advance hunch was validated.

Before diving in, let’s first flip the PCB over and take a look at the other side:

Not as much to see here, aside from this closeup:

The largest two ICs shown, which curiously don’t have their own PCB-marking notations, versus the resistors and capacitors surrounding them (perhaps the marks are underneath the chip packages) are labeled as follows, along with what I think is a STMicroelectronics logo:

071I
GZ229

Any idea what they are, readers? Back to the front for a close-up of the most interesting section:

The largest packaged parts on this side are a mix of what I believe to be QJ423 p-channel and QJ444 n-channel MOSFETs, both curiously identified in online specs as intended for automotive applications. And look, Alex even brands his PCBs!

I’ll close with a few side views of the solid-construction circuit board:

And that’s all I’ve got for you today. I’ll hold onto the disassembled device for a while in case you have any specific questions on the markings on some of the other, tinier ICs and/or passives. And, as always, I welcome your thoughts in the comments. Bonus points for anyone who is able to dig up an Alex-authored DIY schematic that corresponds to this design!

—Brian Dipert is the Editor-in-Chief of the Edge AI and Vision Alliance, and a Senior Analyst at BDTI and Editor-in-Chief of InsideDSP, the company’s online newsletter.

 Related Content

• High quality and lossy: Audio upgrades don’t need to be costly
• FiiO transitions: Technology evolutions
• Putting an audiophile “tweak” to the test
• Decoding and virtualization bring surround sound to the masses
• How to stop worrying and love compressed audio
• Microphones: An abundance of options for capturing tones

The post Seeing inside entry-level audiophile desire: Monoprice’s Liquid Spark Headphone Amplifier appeared first on EDN.

The need for unified networks serving artificial intelligence (AI) training and inference is reaching an unprecedented scale. Broadcom’s answer: The Tomahawk 6 switch delivers 102.4 Tbs of switching capacity in a single chip, doubling the bandwidth of any Ethernet switch currently available on the market.

AI clusters—scaling from tens to thousands of accelerators—are turning the network into a critical bottleneck with bandwidth and latency as major limitations. Tomahawk 6, boasting 100G/200G SerDes and co-packaged optics (CPO) technology, breaks the 100-Tbps barrier while facilitating a flexible path to the next wave of AI infrastructure.

Figure 1 Tomahawk 6’s two-tier network structure, instead of a three-tier network, leads to fewer optics, lower latency, and higher reliability. Source: Broadcom

Ram Velaga, senior VP and GM of Core Switching Group at Broadcom, calls Tomahawk 6 not just an upgrade but a breakthrough. “It marks a turning point in AI infrastructure design, combining the highest bandwidth, power efficiency, and adaptive routing features for scale-up and scale-out networks into one platform.”

First, the Tomahawk 6 family of switches includes an option for 1,024 100G SerDes on a single chip, allowing designers to deploy AI clusters with extended copper reach. Moreover, Broadcom’s 200G SerDes provides the longest reach for passive copper interconnect, facilitating high-efficiency, low-latency system design with greater reliability, and lower total cost of ownership (TCO).

Second, Tomahawk 6 is also available with co-packaged optics, which lowers power and latency while reducing link flaps. Tomahawk 6’s CPO solution is built upon Broadcom’s CPO versions of Tomahawk 4 and Tomahawk 5.

Third, Tomahawk 6 incorporates advanced AI routing capabilities that encompass features like advanced telemetry, dynamic congestion control, rapid failure detection, and packet trimming. These features enable global load balancing and adaptive flow control while supporting modern AI workloads, including mixture-of-experts, fine-tuning, reinforcement learning, and reasoning models.

Figure 2 Cognitive Routing 2.0 in Tomahawk 6 features advanced telemetry, dynamic congestion control, rapid failure detection, and packet trimming. Source: Broadcom

The capabilities outlined above provide essential advantages for hyperscale AI network operators. They also allow cloud operators to dynamically partition their XPU assets into the optimal configuration for different AI workloads. Broadcom claims that Tomahawk 6 meets all networking demands for emerging 100,000 to one million XPU clusters.

Figure 3 Tomahawk 6 can accommodate up to 512 XPUs in a scale-op cluster. Source: Broadcom

While Tomahawk 5 has proven itself in large GPU clusters, Tomahawk 6 takes it a step further in terms of bandwidth, SerDes speed and density, load balancing, and telemetry. Tomahawk 6, compliant with the Ultra Ethernet Consortium, also supports arbitrary network topologies, including scale-up, Clos, rail-only, rail-optimized, and torus.

Related Content

• Solving AI’s Power Struggle
• AI Trolls for Data Center Woes
• Broadcom Throws Programmable Switch
• Are we ready for large-scale AI workloads?
• Ethernet adapter chips aim to bolster AI data center networking

The post New AI networking switch breaks the 100-Tbps barrier appeared first on EDN.

The analog-based timing semiconductor world, comprising crystals and phase-lock loops (PLLs), is facing a conundrum. While crystals provide higher performance at lower frequencies, PLLs accommodate higher frequencies with lower performance. An Irvine, California-based timing startup claims to have an answer to this conundrum. It digitally synthesizes timing signals using CMOS technology, thereby replacing legacy analog chains.

Read the full story at EDN’s sister publication, Planet Analog.

Related Content

• It’s time to advance your clock
• Programmable clock, the universal timing solution?
• Moving Time: Bringing Quantum Clocks Out of the Lab
• A Planet Analog Primer: A Clock For Every Timing Need
• RTC design: Real-time clocks still matter in automation, IoT applications

The post The analog-centric timing world takes a digital turn appeared first on EDN.

The Gigabit Multimedia Serial Link (GMSL) technology of Analog Devices Inc. (ADI) is finally heading down the standardization path with the inception of the OpenGMSL Association, a non-profit entity joined by an automotive OEM, tier 1 suppliers, semiconductor companies, and several test and measurement firms.

GMSL—a SerDes technology for automotive applications like advanced driver assistance systems (ADAS), touchscreen infotainment, and in-vehicle connectivity—facilitates high-resolution video links while supporting data transfer speeds of up to 12 Gbps. ADI claims to have shipped more than 1 billion GMSL chips for automotive-grade platforms.

Figure 1 GMSL is a point-to-point serial link technology dedicated to video data transmission; it was originally designed for automotive camera and display applications. Source: ADI

OpenGMSL Association aims to turn this automotive SerDes technology into an open standard for in-vehicle connectivity. “As automotive architectures evolve to meet the growing demands of in-vehicle communication, networking and data transfer, it is critical that the industry has access to open global standards such as OpenGSML to enable ecosystem-led innovation,” said Fred Jarrar, VP and GM of Power and ASIC Business Unit at indie Semiconductor, a member of OpenGMSL Association.

Among the test and measurement companies joining the OpenGMSL Association are Keysight Technologies, Rohde & Schwarz, and Teledyne LeCroy. These in-vehicle network test outfits will help OpenGMSL in facilitating the development and deployment of interoperable and reliable automotive systems through a standardized, open ecosystem for in-vehicle connectivity.

Hyundai Mobis, which has used the GMSL technology in the Korean OEM’s vehicles for many years, has also joined the initiative to standardize GMSL. Then, there is GlobalFoundries (GF), pitching its 22FDX, 12LP+ and 40LP process technologies for GMSL chips targeted at next-generation automotive applications.

Figure 2 OpenGMSL aims to transform SerDes transmission of video and/or high-speed data as an open standard across the automotive ecosystem. Source: ADI

Next-generation automotive platforms like ADAS heavily rely on high-quality video data to make critical, real-time decisions that improve driver safety and reduce accidents. Likewise, touchscreen infotainment systems demand high-speed and low-latency connectivity for seamless, immersive user experiences.

OpenGMSL aims to accelerate innovation across these automotive platforms by cultivating a standardized, open ecosystem for in-vehicle connectivity. ADI is betting that an open standard for video and/or high-speed data transmission built around its GMSL technology will bolster autonomous driving, ADAS, and infotainment applications, and its own standing in the automotive market.

Related Content

• Machine-Vision Architectures for Automotive Apps
• ADI’s GMSL technology eyes a place on ADAS bandwagon
• Using GMSL SerDes Devices in a Dual Automotive (ECU) Application
• Complex but necessary technology for automotive: Serial links re-visited

The post GMSL video link’s quest to become open automotive standard appeared first on EDN.

Editor’s Note: This DI is a two-part series.

In Part 1, Nick Cornford deliberately oscillates the TDA7052A audio power amplifier to produce a siren-like sound and, given the device’s distortion characteristics, a functional Wien bridge oscillator.

In Part 2, Cornford minimizes this distortion and adds amplitude control to the circuit.

When audio power amplifiers oscillate, the result is often smoke, perhaps with a well-cooked PCB and a side order of fried tweeter. This two-part Design Idea (DI) shows some interesting ways of (mis-)using a common power amp to produce deliberate oscillations of varying qualities.

That device is the TDA7052A, a neat 8-pin device with a high, voltage-controllable gain, capable of driving up to a watt or so into a bridge-tied load from its balanced outputs. The TDA7056A is a better-heatsinked (-heatsunk?) 5-W version. (That “A” on the part number is critical; the straight TDA7052 has slightly more gain, but no control over it.) The TDA7052B is an uprated device with a very similar spec, and the TDA7056B is the 5-W counterpart of that. But now the bad news: they are no longer manufactured. Some good news: they can easily be found online, and there is also a Taiwanese second source (or clone) from Unisonic Technologies Ltd.

A simple circuit’s siren song

For the best results, we’ll need to check out some things that don’t appear on the data sheets, but let’s cut straight to something more practical: a working circuit. Figure 1 shows how the balanced, anti-phase outputs help us build a simple oscillator based on the integrator-with-thresholds architecture.

Figure 1 A minimalist power oscillator, with typical waveforms.

This circuit has just three advantages: it’s very simple, reasonably efficient, and, with a connected speaker, very loud. Apart from those, it has problems. Because of the amp’s input loading (nominally 20k) and the variation of drive levels with different loads, it’s hard to calculate the frequency precisely. (The frequency-versus-R1 values shown are measured ones.) R2 is needed to reduce loading on the timing network, but must leave enough gain for steady operation. (A series capacitor here proved unnecessary, as the internally biased input pin is being over-driven.) Its efficiency is due to the amp’s output devices being run in saturation: with no extra heatsinking, the (DIL-8) package warms by ~15°C when driving into an 8 Ω speaker. The square wave produced is somewhat asymmetrical, though good enough for alarm use.

Figure 1 shows a 5-V supply. Raising that to 12 V made only one change to the performance: the output became very, very loud. And it drew around an amp with a 10 Ω load. And it could do with a heatsink. And a TDA7056A/B rather than a ’52.

The Vcon input on pin 4 is not used. Left open, it floats at ~1.14 V, giving the device a measured gain of around 25 dB. Taking it close to ground inhibits operation, so a bare-drain MOSFET hooked on here can give on/off control. Taking it higher gives full gain, with a shift in frequency. If that is not important (and, in this context, why should it be?), logic control through a 22k resistor works fine. When inhibited, the device still draws 8–10 mA.

Feeding Vcon with varying analog signals of up to a few tens of hertz can produce interesting siren effects because changes in gain affect the oscillation frequency. But for a proper siren, it would be better to generate everything inside a small micro and use an H-bridge of (less lossy) MOSFETs to drive the speaker with proper square waves. (We’ve all heard something like that on nearby streets, though hopefully not in our own.)

Fancy sound effects apart, any power amp with a suitable input structure, enough gain, and balanced (BTL) outputs should work well in this simplest of circuits.

Determining distortion

So much for simplicity and raw grunt. Now let’s take a look at some of the device’s subtleties and see how we can use those to good effect. Distortion will be critical, but the data sheet merely quotes 0.3 to 1% under load, which is scarcely hi-fi. If we remove the load, things look much healthier. Figure 2 shows the unloaded output spectrum when the input was driven from an ultra-low-distortion oscillator, at levels trimmed to give 0 dBu (2.83 V pk-pk) and -20 dBu at the output with a device gain fixed at around 25 dB (Vcon was left open, but decoupled).

Figure 2 The TDA7052A’s output spectra for high and low output levels, taken under ideal conditions and with no output load.

Further tests with various combinations of input level and device gain showed that distortion is least for the highest gains—or smallest gain-reductions—and lowest levels. With outputs less than ~300 mVpk–pk (~-18 dBu) and gains more than 10 dB, distortion is buried in the noise.

That’s unloaded. Put a 10 Ω resistive load across the outputs, and the result is Figure 3.

Figure 3 Similar spectra to Figure 2, but with a 10-Ω output load.

That looks like around -38 dB THD for each trace, compared with better than -60 and -70 dB for the unloaded cases. All this confirms that the distortion comes mainly from the output stages, and then only when they are loaded.

A working one-chip sine-wave oscillator—

This means that we have a chance to build a one-chip Wien bridge audio oscillator, which could even drive a power load directly while still having lower distortion than the average loudspeaker. Let’s try adding a Wien frequency-selective network and a simple gain-control loop, which uses Zener diodes to sense and stabilize the operating level, as in Figure 4.

Figure 4 A simple gain control loop helps maintain a constant output amplitude in a basic Wien bridge oscillator.

The Wien network is R1 to R4 with C1 and C2. This has both minimum loss (~10 dB) and minimum phase shift (~0°) at f = 1 / 2π C2 (R2 + R4), which gives the oscillation frequency when just enough positive feedback is added. When the amplitude is large enough, Zeners D1 and D2 start to conduct on the peaks, progressively turning Q1 on, thus pulling U1’s Vcon pin lower to reduce its gain enough to maintain clean oscillation.

C3 smooths out the inevitable ripple and determines the control loop’s time-constant. R5 minimizes U1’s loading of the Wien network while C3 blocks DC, and R6 sets the output level. The unloaded spectra for outputs of 0 and -10 dBV are shown in Figure 5.

Figure 5. The spectra of Figure 4’s oscillator for 0 and -10 dBV outputs with no load.

—that has problems

While those spectra are half-decent, with THDs of around -45 and -60 dB (or ~0.1% distortion), they are only valid for a given temperature and with no extra load on the output. Increasing the temperature by 25°C halves the output amplitude—no surprise, given the tempcos of the diodes and the transistor. And those 3.3-V Zeners have very soft knees, especially at low operating currents, so they are better regarded as non-linear resistors than as sharp level-sensors.

Adding a 10-Ω resistor as a load—and tweaking R6 to readjust the levels—gives Figure 6.

Figure 6 Similar spectra to Figure 5’s but with the output loaded with 10 Ω.

THD is now around -30 dB, or 3%. Unimpressive, but comparable with many speakers’ distortions, and actually worse than the data-sheet figures for a loaded device.

So, we must conclude that while a one-chip sinusoidal oscillator based on this is doable, it isn’t very usable, and further tweaks won’t help much. We need better amplitude control, which means adding another chip, perhaps a dual op-amp, and that is what we will do in Part 2.

—Nick Cornford built his first crystal set at 10, and since then has designed professional audio equipment, many datacomm products, and technical security kit. He has at last retired. Mostly. Sort of.

Related Content

• Ultra-low distortion oscillator, part 1: how not to do it.
• Ultra-low distortion oscillator, part 2: the real deal
• Distortion in power amplifiers, Part IV: the power amplifier stages
• A pitch-linear VCO, part 1: Getting it going
• A pitch-linear VCO, part 2: taking it further

The post Power amplifiers that oscillate—deliberately. Part 1: A simple start. appeared first on EDN.

Tamper detection devices, commonly built around switches and sensors, employ several techniques according to design specifications and operating environments. T. K. Hareendran examines several anti-tampering device designs to educate and inform users on various aspects of tamper detection circuits. He also presents anti-tampering use cases built around switches and sensors.

Read the full article at EDN’s sister publication, Planet Analog.

Related Content

• Prevent tampering in energy meters
• Achieve tamper-proof capacitive sensing
• What tamper detection IP brings to SoC designs
• Don’t trust your tamper detection circuitry, it may be dumb?
• Anti tamper real time clock (RTC) – make your embedded system secure

The post Basic design considerations for anti-tampering circuits appeared first on EDN.

The fundamental difference between Microsoft and Google’s dueling mid-May keynotes this year comes down sizzle versus steak. And that isn’t just my opinion; I can even quantify the disparity that others apparently also ascertained. As noted in my recent coverage of Microsoft’s 2025 Build conference, the full keynote ran for a minute (and a few seconds) shy of 2 hours:

But The Verge was able to condense the essentials down to a 15-minutes (and a few seconds) summary video, 1/8th the length of the original:

What about Google’s day-later alternative? It was only a couple of minutes shorter in total:

But this time, The Verge was only able to shrink it down to around 1/3 the original length, resulting in a 32-minute (and change) summary video:

Translation: nearly the same keynote duration, but much more “meat” in the Google keynote case. And that’s not even counting the 70-minute developer-tailored keynote that followed it:

That said, in fairness, I’ll point out that Google’s own summary video for the keynote was only 10 minutes long, so…

What did Google’s presenters cover in those 3+ two-keynote hours, and more generally across the two-day event (and its virtual-event precursor)? Glad you asked. In the sections that follow, I’ll touch on what I thought were at least some of the high points. For more, check out Google’s summary blogs for the developer community and the public at large, along with the conference coverage summary pages from folks like 9to5Google, Engadget, The Verge and Wired.

Android (and its variants)

Conceptually similar to what Microsoft had done, Google decided to release some of its news ahead of the main event. This time, though, it was one week prior, not two. And the focus this time was on software, not hardware. Specifically, Google discussed its upcoming Expressive Design revamp of the core Android UI and associated apps, along with planned added-and-enhanced features for the O/S and apps, and related evolutions of the Android variants tailored for smart watches (Wear OS), smart glasses and headsets (Android XR), vehicles (Android Auto), displays (Google TV), and any other O/S “spins” I might have overlooked at the moment. In the process, Google got the jump on Apple, who will reportedly announce a conceptually similar revamp for its various O/Ss in a couple of weeks (stay tuned for my coverage)!

I’ll talk more about Android XR and its associated hardware, as Google did at I/O itself, in a separate topic-focused section to come later in this piece.

Multimodal large language models

Gemini, as I’ve discussed in past years’ Google I/O reports and other writeups, is the company’s suite of proprietary deep learning models, all becoming increasingly multimodal in their supported data input-and-output diversity. There are currently three primary variants:

• Pro: For coding and complex prompts
• Flash: For fast performance on complex tasks, and
• Flash-lite: For cost-efficient performance

Plus, there’s Gemma, a related set of models, this time open source, which, thanks to their comparatively low resource demands, are also useful for on-device inference with edge systems.

Latest v2.5 of Gemini Pro and Gemini Flash had both already been unveiled, but at I/O Google touted iterative updates to both of them, improving responsiveness, accuracy and other metrics. Also unveiled, this time first-time, was Gemma 3n, specifically tailored for mobile devices. And also newly announced was Gemini Live, which supports the real-time analysis and interpretation of (and response to) live audio and video feeds coming from a camera and microphone. If you’re thinking this sounds a lot like Project Astra, which I mentioned at the tail-end of last year’s Google I/O coverage (albeit not by name)…well, you’d be spot-on.

AI integration into other Google products and services…including search

Just as Microsoft is doing with its operating system and applications, Google is not only developing user direct-access capabilities to Gemini and Gemma via dedicated apps and web interfaces, it’s also embedding this core AI intelligence into its other products, such as Gmail, various Workspace apps, and Google Drive.

The most essential augmentation, of course, is that of the Google Search engine. It was Google’s first product and remains a dominant source of revenue and profit for it and parent company Alphabet, by virtue of the various forms of paid advertising it associates with search results. You may have already noticed the “AI Overview” section that for a while now has appeared at the top of search results pages, containing a summary explanation of the searched-for topic along with links to the pages used to generate that explanation:

Well, now (as I was writing this piece, in fact!) “AI Mode” has its own tab on the results page:

And similarly, there’s now an “AI Mode” button on the Google Search home page:

Google is even testing whether to relocate that button to a position where it would completely replace the longstanding “I’m Feeling Lucky” button.

It wasn’t too long ago when various tech pundits (present company excluded, to be clear) were confidently forecasting the demise of Google’s search business at the hands of upstarts like OpenAI (more on them later). But the company’s “deft pivot” to AI teased in the title of this piece has ensured otherwise (at least until regulatory entities may say otherwise)…perhaps too much, it turns out. As I’ve increasingly used AI Overview (now AI Mode), I find that its search results summaries are often sufficient to answer my question without compelling me to click through to a content-source page, a non-action (versus tradition) that suppresses traffic to that page. Google has always “scraped” websites to assemble and prioritize search results for a given keyword or phrase, but by presenting the pages’ information itself, the company is now drawing the ire of publishers who are accusing it of content theft.

Rich content generation

Take generative AI beyond LLMs (large language models) with their rudimentary input and output options (at least nowadays, seemingly…just a couple of years ago, I was more sanguine about them!), and you’re now in the realm of generating realistic still images, videos, audio (including synthesized music) and the like. This is the realm of Google’s Imagen (already at v4), Veo (now v3), and Lyria (v2 and new RealTime) models and associated products. Veo 3, for example, kicked off the 2025 Google I/O via this impressive albeit fanciful clip:

Here’s another (less silly overall therefore, I’d argue, even more impressive) one from Google:

More synthesized video examples and their associated text prompts can be found at the Veo page on the Google DeepMind site. Veo 3 is already in public release, with oft-impressive albeit sometimes disturbing results and even real-life mimickers. And combine audio, video and still images, add some additional scripting smarts, and you’ve got the new AI filmmaking tool Flow:

Who would have thought, just a few short years ago, that the next Spielberg, Scorsese, Hitchcock, Kubrick, Coppola or [insert your favorite director here] would solely leverage a keyboard and an inference processor cloud cluster as his or her content-creation toolbox? We may not be there yet, but we’re getting close…

Coding assistants

Coding is creative, too…right, programmers? Jules is Google’s new asynchronous coding agent, unveiled in Google Labs last December and now in public beta, where it goes up against

OpenAI’s recently delivered one-two punch of the internally developed Codex and acquisition (for $3B!) of Windsurf. That said, as VentureBeat also notes, it’s not even the only AI-powered coding tool in Google’s own arsenal: “Google offers Code Assist, AI Studio, Jules and Firebase”.

Android XR-based products (and partnerships)

Google co-founder Sergey Brin made a curious onstage confession during a “fireside chat” session at Google I/O, admitting that he “made a lot of mistakes with Google Glass”:

His critique of himself and the company he led was predominantly two-fold in nature:

• Google tried to “go it alone” from a hardware development, manufacturing and marketing standpoint, versus partnering with an established glasses supplier such as Italian eyewear company Luxottica, with whom Meta has co-developed two generations (to date) of smart glasses (as you’ll soon learn about in more detail via an upcoming sorta-teardown by yours truly), and
• The bulbous liquid crystal on silicon (LCoS) display in front of one of the wearer’s eyes ensured that nobody would mistake them for a conventional pair of glasses…a differentiation which was not advantageous for Google.

Judging from the 2025 Google I/O messaging, the company seems determined not to make the same mistake again. It’s partnering with Warby Parker, Korea-based Gentle Monster, Samsung and Xreal (and presumably others in the future) on smart glasses based on its Android XR platform…glasses that it hopes folks will actually want to be seen wearing in public. Samsung is also Google’s lead partner for a VR headset based on Android XR…the “extended reality” (XR) that Google envisions for the operating system spans both smart glasses—with and without integrated augmented reality displays—and head-mounted displays. And it not only did live demos during the keynote but also gave attendees the chance to (briefly) try out its prototype smart glasses, glimpsed a year ago in the Project Astra clip I mentioned earlier, for themselves.

Google Beam

Two years ago, I noted that the way-cool Project Starline hologram-based virtual conferencing booth system announced two years earlier (during COVID-19 lockdowns; how apropos):

had subsequently been significantly slimmed down and otherwise simplified:

Fast forward two more years to the present and Google has rebranded the 3D-rendering technology as Beam, in preparation for its productization by partners such as HP and Zoom:

And in the process, Google has notably added near real-time, AI-powered bidirectional language translation to the mix (as well as to its baseline Google Meet videoconferencing service, which previously relied on captions), preserving each speaker’s tone and speaking style in the process:

Now there’s a practical application for AI that I can enthusiastically get behind!

OpenAI’s predictable (counter)punch

In closing, one final mention of one of Google’s primary competitors. Last year, OpenAI attempted to proactively upstage Google by announcing ChatGPT’s advanced voice mode one day ahead of Google I/O. This time, OpenAI attempted to suck the wind out of Google’s sails retroactively, by trumpeting that it was buying (for $6.5B!) the “io” hardware division of Jony Ive’s design studio, LoveFrom, one day after Google I/O. Not to mention the $3M allegedly spent on the “odd” (I‘m being charitable here) video that accompanied the announcement:

While I don’t at all discount OpenAI’s future prospects (or Meta’s, for that matter, or anyone else’s), I also don’t discount Google’s inherent advantage in developing personalized AI: it’s guided by the reality that it already knows (for better and/or worse) a lot about a lot of us.

How do you think this’ll all play out in the future? And what did you think about all the news and enhanced technologies and products that Google recently unveiled? Let me (and your fellow readers) know in the comments!

—Brian Dipert is the Editor-in-Chief of the Edge AI and Vision Alliance, and a Senior Analyst at BDTI and Editor-in-Chief of InsideDSP, the company’s online newsletter.

 Related Content

• The 2024 Google I/O: It’s (pretty much) all about AI progress, if you didn’t already guess
• The 2023 Google I/O: It’s all about AI, don’t cha know
• Smart glasses skepticism: A look at their past, present, and future(?)
• Google’s fall…err…summer launch: One-upping Apple with a sizeable product tranche

The post The 2025 Google I/O conference: A deft AI pivot sustains the company’s relevance appeared first on EDN.