EdgeCortix, a startup focused on making AI inference faster and more power-efficient, is moving toward a chiplet-based design for its third-generation hardware. The shift signals a fundamental change in how edge AI (artificial intelligence) chips are being engineered to handle the latest generation of AI models, particularly transformer-based systems that power large language models (LLMs). Rather than building monolithic chips optimized for a single task, the company is breaking its neural processing unit (NPU) into modular components that can be combined based on specific customer needs .

What Changed Between EdgeCortix's Second and Third Generation?

When EdgeCortix launched its Sakura-II chip in 2024, the company optimized it for vision processing tasks using convolution-type operations, which are common in image recognition and computer vision applications. However, the AI industry has shifted dramatically toward transformer models, which power systems like ChatGPT and other large language models. These models are fundamentally different from vision-focused AI; they are memory-bound rather than compute-bound, meaning the bottleneck is moving data efficiently rather than raw processing speed .

"When we started in 2019, a lot of the world was focused on vision processing. Today, where we are, effectively maybe a couple of years into our journey, we realized that the industry was steadily shifting towards transformers as a model that is a lot more memory-centric rather than compute-centric," explained Sakya Dasgupta, CEO of EdgeCortix.

Sakya Dasgupta, CEO at EdgeCortix

This realization forced EdgeCortix to rethink its Dynamic Neural Accelerator (DNA) architecture, the core processing engine inside its chips. The company's second-generation Sakura-II delivers 60 TOPS (trillion operations per second) in INT8 precision, a metric that measures how many mathematical operations the chip can perform per second using 8-bit integer data. But raw TOPS alone don't tell the full story for transformer workloads, which require different optimization strategies .

How Is EdgeCortix Solving the Transformer Problem?

The move toward chiplets represents a practical solution to a fundamental engineering challenge. Instead of designing a single monolithic chip that tries to handle every possible AI workload, EdgeCortix is building modular components that can be combined in different configurations. This approach allows customers to pay for only the processing power and memory they actually need, rather than buying a fixed package that may be overspecified for some applications or underpowered for others .

The chiplet strategy also addresses supply chain resilience and manufacturing flexibility. EdgeCortix is based in Japan, a location the company views as strategically important for semiconductor manufacturing in 2026 and beyond. Chiplets can be manufactured at different facilities and then assembled together, reducing dependence on any single production line or geography .

Steps to Understanding EdgeCortix's Hardware Evolution

• Vision-First Era: Sakura-I (first generation) was optimized for convolution operations common in image recognition, using integer-only precision to maximize efficiency for vision workloads.
• Transformer Shift: The industry moved toward memory-bound transformer models, requiring architectural changes to minimize data movement rather than maximize compute throughput.
• Chiplet Modularity: Third-generation hardware will use chiplets, allowing customers to configure systems for specific workloads without paying for unused capacity.

EdgeCortix's journey reflects a broader pattern in the AI chip industry. The company started with a "software-first" philosophy in 2019, focusing on how to actually deploy AI models in real-world applications rather than simply maximizing benchmark scores. This approach meant building both hardware and software together, with the compiler stack (the software that translates AI models into instructions the chip can execute) as important as the silicon itself .

"The critical challenge we face when we are deploying a solution for the end customer is not how much TOPS a chip has, but can it really run the frontier model that I want to run? Can it give me the frames per second that I need? And can it do that at the required amount of power consumption?" stated Dasgupta.

Sakya Dasgupta, CEO at EdgeCortix

The company has already proven this approach works in demanding applications. EdgeCortix chips are being used in space and aerospace projects, including work with NASA, where reliability and power efficiency are non-negotiable requirements. These real-world deployments validate the architectural decisions the company is making for its third generation .

Why Does This Matter for the Broader AI Chip Market?

EdgeCortix's pivot toward chiplets and transformer-optimized architectures signals how the edge AI market is maturing. Early AI chips were often designed around theoretical benchmarks or specific use cases like autonomous vehicles or smart cameras. Now, companies are building hardware that reflects how AI is actually being deployed: running inference (the process of using a trained model to make predictions) on devices at the edge rather than sending all data to cloud servers .

The emphasis on latency and power efficiency remains central to EdgeCortix's strategy. Real-time applications, from robotics to medical imaging to industrial monitoring, cannot tolerate the delays of sending data to a distant data center and waiting for results. Edge AI chips must deliver answers in milliseconds while consuming minimal power, especially in battery-powered or thermally constrained environments .

EdgeCortix's third-generation architecture, built on chiplets and optimized for transformers, represents the next phase of this evolution. As AI models continue to grow in complexity and as deployment scenarios become more diverse, the ability to customize hardware configurations will become increasingly valuable. The company's bet on modularity and its focus on solving real customer problems, rather than chasing raw performance metrics, positions it as a meaningful player in the competitive edge AI chip market.

" }