Kazu Gomi, president and CEO of NTT Research, told EE Times that the chip’s standout feature is its ability to perform real-time object detection on full-resolution 4K video streams at under 20W—a very attractive proposition for battery-powered aerial platforms.

By contrast, conventional systems trade accuracy for efficiency. Even when they accept 4K input, they typically downscale frames to around 600×600 pixels before running object detection models like YOLO (short for “You Only Look Once,” a family of real-time object detection algorithms that process an entire frame in a single pass).

How NTT’s chip works

Drawing on decades of expertise in video coding, NTT developed a different approach. Instead of downscaling, the chip divides each 4K frame into multiple 600×600 blocks, runs YOLO detection on each block in parallel and then fuses the results with a holistic “overview” process. This stitching operation resolves cases where objects span multiple blocks.

The chip includes a dedicated image-processing engine for tasks like downscaling and tiling, and a separate inference engine for high-speed detection. Results from these engines are fused in software, giving developers flexibility to adapt to different models.

The architecture also exploits temporal redundancy between frames. By tracking slow-moving objects across a 30-fps stream, it reduces calculations and power use. “We can maintain high-resolution accuracy and keep power under 20 watts because we reuse motion vector information,” Gomi explained.

Efficiency is achieved through fixed-function hardware blocks for video inference, while flexibility comes from software control. Latency performance is equally noteworthy: the chip processes each 4K frame quickly enough to keep pace with 30 fps video without dropping frames — a critical factor for navigation, target tracking and obstacle avoidance.

To put this new chip in perspective, EE Times spoke with Ian O’Connor, distinguished professor at École Centrale de Lyon, who specializes in AI chip architectures. “The ability to run real-time inference on full-resolution 4K streams without compression is rare—and that translates into more robust detection, especially for small or distant objects,” he said.

O’Connor also noted the chip’s use of dynamic precision control in its inference engine. “NTT is the first I’ve seen apply variable bit-precision at real-time 4K scale on the edge,” he said. “That’s a technical advantage worth watching.”

Why drones could benefit most

The chip’s ability to detect objects at greater distances than conventional downsampled systems could unlock new drone capabilities: swarming, coordinated missions or dynamic rerouting based on terrain and target recognition. “4K real-time object detection means you can detect objects further away and classify them more accurately—whether it’s a tree, a tank or a human,” Gomi said.

O’Connor added that another attractive feature for unmanned aerial vehicle (UAV) applications is the onboard 4K processing, which reduces reliance on external data links. “By processing video locally, you cut latency and spectrum congestion—which is critical for swarm operations, where dozens of UAVs may need to coordinate autonomously,” he said. Keeping data local also improves security, since sensitive imagery does not need to leave the platform.

At present, the chip is optimized for video inference workloads. However, it also integrates a CPU and Ethernet interface that could eventually support navigation or communications.

While competition is strong, there may still be room for NTT. O’Connor pointed out that Qualcomm’s Snapdragon Flight RB5 combines 4K capture with onboard ML acceleration at low power, but it typically does this on compressed streams. Nvidia’s Jetson platforms also target edge AI, but very broadly. “If NTT can prove uncompressed 4K inference at 20 W in the field, it will clearly standout from the rest,” O’Connor said.

From R&D to market

NTT unveiled the chip only a few months ago, with general availability planned for early 2026. So far, the chip does not even have a name, and no customers are using it in products today. Not to worry, according to Gomi. This gives potential partners time to influence the final product.

NTT expects to develop UAV-specific adaptations, such as ruggedization and avionics compatibility, with partner input. The company also needs to tailor its thermal management at 20W to compact airborne systems—one more thing they hope to do with outside help.

“We’re interested in working with different partners for different applications,” Gomi said. “This is the stage where we can adapt the design to market needs.” For drone system integrators, NTT will release an SDK to expand support for inference models. The SDK will give integrators control to extend inference beyond the built-in YOLO-style algorithms.

O’Connor said that while Qualcomm and Nvidia bring mature ecosystems, NTT’s opportunity lies in specialization. “If they can ship a robust SDK and prove reliability under drone conditions, they’ll carve out a defensible niche,” he added.

Beyond this chip, NTT is also exploring frontier technologies like biodegradable electronics and optical-based processors that could one day replace GPUs. These innovations aim to cut power requirements enough to enable advanced AI, even generative AI, on mobile platforms.

“In five or six years, it may be possible to mount a GenAI platform on a flying object using our optical chips,” Gomi said.

The project reflects NTT’s tradition of using R&D to leap into new markets. Some of us are old enough to remember that in February 1999, the company launched i-mode, which brought internet to the handsets of millions of Japanese users long before most people in developed countries even had a cellphone. NTT now sees edge AI as its next big opportunity.

Given the expanding commercial drone market expected to reach billions in revenue and millions of units annually, there remains substantial room for NTT to take a meaningful share, especially in industrial, infrastructure and specialized drone applications that demand high-performance AI at low power, according to O’Connor. “It helps that NTT is increasingly being recognized as a major technology and digital business leader,” he said.

NTT hopes the convergence of market demand and technical capability will position it as a key enabler of UAV autonomy. If the technology proves itself in the field, Gomi believes it could become a standard component in both civilian and defense drone systems.

And with commercialization just months away, the clock is ticking for potential partners to engage. “The market tells you what to do,” Gomi said. “We have something that’s almost good enough—and we can make it better.”

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved