Summary
As Software Manager at Circle Optics, Steve Feller focuses on how machines interpret the world, not just capture it. Building on insights from NVIDIA GTC, his work bridges imaging and intelligence to deliver reliable, real-time perception in demanding environments.
Steve Feller, Software Manager, Circle Optics
NVIDIA GTC is a portal to how AI, robotics, and autonomous systems are actually being built. For Circle Optics, the purpose of attending is not visibility. It is to understand how these systems are being designed and where perception fits within that architecture. This year, Software Manager Steve Feller and Senior Computer Vision Engineer Mitchell Baller attended with a shared objective: to evaluate how system constraints, simulation, and edge processing are reshaping the role of visual data. Across sessions, one pattern is clear. Systems are no longer designed to capture as much data as possible. They are designed to operate within constraints and make decisions about what data is necessary. If the input is incomplete or distorted, the system compensates downstream. If the input is correct at acquisition, the system performs more efficiently and with less ambiguity.
When machines need to capture everything, they capture with Circle Optics.
System Constraints Are Driving Design Decisions
The most consistent shift observed across sessions is a move away from maximizing data collection toward operating within system constraints. These constraints include bandwidth, power, and physical system design, and they determine how systems are built and deployed. This is not theoretical. It directly impacts performance and reliability. Steve Feller attended GTC to evaluate how these constraints are being addressed across the ecosystem and how they influence system-level decisions. If the system cannot move or process all available data, then the system must decide what data is necessary.
Shift from Data Volume to Data Prioritization
One of the clearest observations from GTC is the changing relationship between data and computation. “Data is still valuable,” Feller said, “but computation is becoming more important because of the limits of what we can realistically capture.” The constraint is physical. There are limits on how much data can be captured, how quickly it can be transmitted, and how it can be processed. These limits are tied to bandwidth, power, and system size. Feller was explicit about how this shows up in practice. “We’re bandwidth limited. Without processing at the edge, you either oversize the system or slow down how it operates.” Oversizing increases cost and complexity, while slowing down reduces effectiveness. As a result, system architecture is shifting toward edge processing. Data is analyzed at the point of capture, relevant information is identified, and only that subset is transmitted. This pattern is consistent across domains. Not all data can be moved. The system must decide what matters.
Accuracy, Simulation, and Deployment Requirements
For Circle Optics, this reinforces a core design position. Capture must be correct at acquisition so that downstream systems do not need to reconstruct or infer missing context. This improves efficiency and reduces risk. Another consistent signal from GTC is the importance of ecosystem alignment. NVIDIA is defining the frameworks that many systems will use for development and integration, and alignment enables faster development and easier collaboration. At the same time, alignment requires discipline in selecting where integration creates value. A key distinction discussed at GTC is between accuracy and representation. In simulation environments, representation is often sufficient. Systems can train and iterate using approximations. In deployed systems, particularly in defense and autonomy, accuracy remains required. This creates a dual model in which simulation supports development while measured data supports real-world performance.
Feller also pointed to a broader shift that goes beyond system constraints. In some AI workflows, simulation is becoming the primary environment and measured data is used less frequently. Systems are increasingly trained on modeled environments rather than direct observations of the real world. This changes how accuracy is defined and how systems are validated.
“You could get to a point where systems never even consume real data. They operate entirely on simulated or generated inputs.”
This introduces a new design question. When does a system require measured data, and when is a representation sufficient. The answer depends on the application, but the distinction is now part of system architecture. For Circle Optics, the position remains consistent. Provide high-quality data at capture, integrate where it improves performance, and maintain accuracy where it is required. In deployed environments, input quality remains the limiting factor.
From GTC to Rochester: A Conversation on the Future of Perception Systems
Recorded at ROC Vox Studios in Rochester, New York, this episode of Stitchless: Circle Optics Team in Focus captures a real-time conversation with Software Manager Steve Feller and Senior Computer Vision Engineer Mitchell Baller following their return from NVIDIA GTC.
This discussion reflects what is actually shifting in how machines capture, process, and act on visual data.
You can listen to the full episode featuring Steve Feller and Mitchell Baller wherever you get your podcasts.
