MyTechDailyNews Hardware Scorecard: Deepx NPUs, Physical AI and the Limits of TOPS

For hardware buyers, the most difficult part of physical AI is no longer finding the highest number on a datasheet. It is deciding whether a low-power NPU can move from a laboratory demo to cameras, robots and industrial systems with repeatable power, thermal and software behavior. The analysis below summarizes a Deepx interview with EE Times Asia, in which founder and CEO Lokwon Kim argues that silicon starts the sales process, while tools, portability and production support determine whether a design stays selected after evaluation.

Why physical AI hardware evaluation differs from cloud or simple edge AI

Cloud deployment can often assume relatively standardized compute, networking and storage. Physical AI faces a fragmented base: cameras, robots, vehicles, drones and industrial machines vary in power envelope, heat dissipation, sensor type, operating system and real-time behavior. This makes fragmentation a major barrier between demonstration and volume production.

The three-stage buyer path, from first design win to ecosystem lock-in

Kim describes the market in a sequence that maps well to hardware procurement: performance secures initial interest, mass-production qualification protects the project, and familiar tools reduce long-term switching cost. The company summarizes the path as follows:

“We see the physical AI semiconductor business developing in three stages. First, differentiated hardware performance must earn the customer’s initial selection. Second, a company must expand design wins and mass production, commercialize at scale, and build a defensible position through a strong portfolio of foundational IP. The third stage is to create an ecosystem in which developers can continuously build and deploy AI models and software using tools that are already familiar and broadly adopted,

Beyond TOPS: what a hardware review must test

TOPS and performance per watt remain useful for battery-powered or thermally limited boards. Yet they do not reveal whether a customer can port a model, meet a bill-of-materials target or qualify supply. Deepx often cites Nvidia as a reminder that a single generation of silicon rarely defines a platform; CUDA built a development environment around hardware and gave teams workflows that persisted across products. Physical AI has no comparable settled position, and the interview emphasizes the opening.

“Physical AI semiconductors do not yet have an undisputed incumbent with that level of ecosystem dominance. That makes this period particularly important. Whoever succeeds in building an open developer ecosystem and establishing a de facto standard early can shape the market for years to come,

Comparison table: benchmark data versus production evidence

Hardware scorecard: screening metrics versus verifiable production evidence

Decision area	Traditional screening metric	Physical AI evidence	Procurement check
Compute capacity	TOPS and performance per watt	Power and thermal behavior inside the product envelope	Run the intended vision or robot workload, not a synthetic peak only
Model fit	Benchmark accuracy on reference models	Conversion, optimization and accuracy retention for customer models	Measure redevelopment effort before signing a design-in
System cost	NPU unit price	Complete bill of materials, including memory, host processors, power management and cooling	Compare total product cost targets, not chip price alone
Portability	Availability of a development board	Same execution environment from prototyping to industrial validation	Test a prototype migration before selecting volume hardware
Scale	Evaluation-kit demonstration	Supply, yield, quality control and long-term software support	Require evidence from pilot production and field deployment

Data source: production framework described by Deepx in the EE Times Asia interview.

Four major barriers between proof of concept and volume deployment

Kim frames production as a set of hard engineering constraints. A system must operate consistently across thousands or millions of units, meet a defined power budget, stay within thermal limits, support the required models, meet cost targets and provide predictable long-term supply and software support. Deepx groups the failure modes into four major barriers: power and thermal constraints, model compatibility, total system cost and production readiness.

• Power and thermal limits: the processor must fit inside the product’s physical and thermal boundaries, not just the evaluation kit.
• Model compatibility: a strong benchmark can disappear if converting and optimizing customer models reduces accuracy or demands extensive software redevelopment.
• Total system cost: customers review the complete bill of materials instead of the NPU price alone.
• Production readiness: a successful board is only the first step; dependable supply, manufacturing yield, quality control and long-term support must follow.

“The first is power and thermal constraints. Physical AI products such as cameras, robots, and drones cannot accommodate data-center-class power consumption or cooling systems. If power and heat exceed the physical limits of the product, no other advantage can compensate. This is a non-negotiable constraint,

The quote explains why a vendor role changes. Kim puts the operational burden this way:

“A semiconductor vendor therefore cannot stop at supplying a chip and demonstrating performance. It must help the customer run the required models, reduce system-level power and cost, and work alongside them from software integration and pilot production through quality validation and volume manufacturing,

Portability: from one development board to industrial systems

Most projects begin with an accessible development board, a camera and a familiar model, then move to industrial hardware. That transition can create large software effort when the environment changes from platform to platform. Deepx says portability should not be judged by moving between two boards; it should preserve the developer’s initial software investment throughout the product lifecycle.

To reduce friction, Deepx connects its NPUs to environments developers already use. The partnerships named in the interview include Ultralytics YOLO, Baidu PaddlePaddle and Raspberry Pi, while the company is also working on support for Nvidia Isaac applications and other software environments. The stated objective is to cut the number of times developers must rebuild software as an application shifts from experimentation to production.

Three lessons from early developer engagement

1. The entry point must be simple: documentation, installation procedures, familiar workflows and technical support carry major weight.
2. Hardware decisions can be influenced during prototyping, well before production.
3. An easy development platform is not enough unless developers can carry models and software into industrial validation and production.

Common hardware-evaluation misconceptions

• Misconception: A high TOPS number proves deployment readiness. Reality: Production needs verified power, thermal, cost, supply and model behavior.
• Misconception: Framework compatibility is optional. Reality: The interview treats familiar tools and portability as ecosystem requirements.
• Misconception: The NPU price is the bill of materials. Reality: Memory, host processors, power management and cooling affect commercial targets.
• Misconception: Cloud connectivity can be assumed. Reality: Robots and industrial machines must keep making immediate decisions when network connectivity is lost or cloud responses are delayed, while data privacy can keep sensitive information on the device.

Open stack: multiple standards, not one winner

Deepx does not expect physical AI to converge on one software platform the way Android became a dominant smartphone operating system. It expects convergence around multiple de facto standards at different layers. The frameworks named are PyTorch, Ultralytics and PaddlePaddle around model development; ExecuTorch for lightweight edge deployment; llama.cpp for generative AI; and ROS 2 plus Nvidia Isaac for robotics development assets.

“The central question is not who owns every layer of the stack. It is who can connect different models, frameworks, and hardware platforms to actual silicon and production products with the least friction. Developers will not accept having to rewrite their software for every new processor,

Distribution as an engineering requirement

As Deepx began mass production of its first commercial product, the company established partnerships with 27 distributors, including Avnet, WPG, Macnica and DigiKey, plus regional partners across several markets. Kim argues that AI moving into industrial applications changes the distributor role. Distribution is no longer limited to inventory, logistics and sales; partners increasingly provide technical support, identify customers, engage in engineering and support design-in activity.

Application channels also differ. Robotics, machine vision, factory automation, security and healthcare have distinct procurement structures and technical-support requirements.

Application fit: where intelligent vision is ahead of robotics

Among physical AI uses, intelligent vision appears closest to large-scale commercialization because many markets already have extensive camera deployments: security, video surveillance, transportation, retail, factory monitoring and machine vision. The AI models used in many of these applications are comparatively mature. The economic case can be direct: AI can automate continuous video monitoring or visual inspection, reduce labor requirements and identify incidents or defects faster. Companies can often add an edge AI processor without replacing the entire camera or infrastructure.

Smart manufacturing is another near-term opportunity. Asia-Pacific’s position in global manufacturing and electronics supply chains, combined with labor shortages and pressure to improve quality, is driving adoption in machine-vision inspection, worker safety, predictive maintenance and logistics automation.

Robotics may eventually represent the largest physical AI market, but it is also considerably more difficult. A vision system mainly perceives and analyzes; a robot must perceive, decide and execute physical action in real time. An incorrect perception can become an incorrect movement or a safety event, raising requirements for reliability, model robustness and system integration.

Kim identifies real-world model reliability, power and thermal constraints, system cost, software compatibility, connectivity and safety as remaining barriers. Physical AI models must maintain performance despite changes in lighting, vibration, weather, camera angles and obstacles.

From edge AI to Federated AI operations

Kim says the next stage of physical AI development will require something resembling MLOps for distributed devices. Cloud AI benefited from mature workflows connecting training, deployment, monitoring and updates. Physical AI remains fragmented, with deployment and operational management often handled independently for individual devices and projects. That model will not scale across millions of cameras, robots and industrial machines.

The five requirements Kim lists are:

• secure cloud-device communication;
• a reliable fleet-wide model;
• software deployment with rollback mechanisms;
• simple tools for managing field equipment;
• unified monitoring of device status, performance, power, errors and model accuracy.

Deepx describes the resulting architecture as “Federated AI,” while distinguishing it from federated learning. In this model, the data center handles training and centralized management, edge devices perform inference and control where latency, security and energy efficiency matter, and gateways connect the two where necessary. The company is working with AWS IoT Greengrass toward an environment where models trained in the data center can be securely deployed to devices, run locally on Deepx NPUs and managed centrally as a fleet.

Power and efficiency data cited in the interview

The interview includes several architecture-level figures. They are not presented as universal benchmarks; they illustrate the leverage possible as more inference moves onto devices.

• Deepx says its internal analysis compared data center GPU systems with on-device AI systems across power, servers, cooling and network resources. The company identified potential for about a 100-fold improvement in resource efficiency in some applications and system configurations.
• In an internal architecture-level simulation involving a 100-billion-parameter mixture-of-experts model, with approximately 10% of parameters active at any one time, Deepx identified the potential for a single 5-W-class AI accelerator to generate more than 10 tokens per second.
• The company notes that the 5-W figure applies to the chip itself, while total system power varies with memory capacity, configuration and operating conditions.
• Workload research suggests that models at or below the 100-billion-parameter class could potentially handle more than 80% of inference requests in the physical AI scenarios being targeted. In that scenario, the data center would handle the remaining, more complex workloads, potentially allowing existing central infrastructure to serve more users or devices.

Physical AI as infrastructure and the mobile-computing analogy

Kim compares the transition to the move from desktop computing to mobile computing. Low-power CPUs moved computation into handheld devices and created a mobility-focused model. Deepx expects ultra-low-power AI semiconductors to produce a similar shift for inference, moving it from centralized data centers into machines and objects operating in the physical world.

The company describes this emerging class of devices as “Free-Token Machines” not because token generation is literally free, but because local inference can occur without a data-center charge for every interaction and without continuous dependence on network connectivity. As models become capable and efficient enough to run locally, cameras, robots and industrial systems can become agentic machines that perceive surroundings, make decisions and act.

The data center remains responsible for frontier-model training and the most demanding workloads. Local devices absorb workloads where latency, privacy, power efficiency and connectivity make centralized processing less practical. When viewed as a distributed compute layer, billions of devices stop being simple endpoints and become part of the infrastructure.

Procurement checklist: action advice for hardware teams

1. Start with the product power and thermal envelope, not the processor datasheet alone.
2. Run a customer model migration test and record accuracy change and redevelopment effort.
3. Score the development entry point: documentation, installation, familiar frameworks, examples and support responsiveness.
4. Check whether the execution environment spans prototyping and industrial validation.
5. Calculate the complete bill of materials, including memory, host processor, power management and cooling.
6. Require evidence of production readiness: supply, yield, quality control, long-term software support and field deployment.
7. Plan fleet operations early, including secure communication, rollback, monitoring and model update policy.
8. Map the ecosystem against the team’s tools before assuming rewrite cost.

The practical takeaway for hardware buyers is narrow: a platform is only as good as its ability to survive transfer from evaluation to production. TOPS and performance per watt still matter, but they are screening tools, not proof. The decisive metrics are model portability, system power, bill of materials, supply and managed fleet operations. If those are verifiable, the physical-AI opportunity becomes repeatable; if they are not, the device remains a well-specified demo.