Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / How Edge Computing Is Enabling Real-Time Video Analytics in Stadiums
Transcript
- Lucas: You're at a packed stadium. Seventy thousand people. Every phone in the air. And you want to see a replay of that touchdown from the angle behind the end zone, on your phone, with less than a second of delay. That is not a trivial technical problem. Luna: And the cloud is not the answer for that, is it? Lucas: Not for the sub-50-millisecond kind of latency you need to make it feel instant. That's where edge computing comes in. And today I want to talk about how stadiums—specifically SoFi Stadium in Los Angeles—have become a showcase for real-time video analytics at the edge. Luna: If today's tech conversation gave you something useful, you know the way these stay ad-free is listener support. Buy me a coffee dot com slash fexingo. Keeps it going. Lucas: Yeah, genuinely. And we'll get right into it. So SoFi Stadium opened in 2020. It's got over 200 cameras feeding into an edge compute network that does live video analytics for security, replays, and personalized content delivery. Luna: Two hundred cameras. That's a lot of data. How do you even move that much video without clogging the network? Lucas: You don't send it all to the cloud. That's the point. SoFi runs a distributed edge architecture with GPU clusters located in the stadium's equipment rooms. The video streams are processed locally—object detection, face blurring for privacy, dynamic cropping—and only the metadata or the clipped highlights go upstream. Luna: So the raw feed never leaves the building. That's a security win too. Luna: Right, so the edge is doing the heavy lifting. What kind of hardware are we talking about? Lucas: SoFi partnered with a company called EdgeMicro for the compute nodes. Each node is a small form-factor server with an NVIDIA A16 GPU—that's a quad gpu card built for video inference. They have about 40 of those nodes distributed across the venue. Total power draw is under 10 kilowatts, which is manageable for a building that already has massive HVAC and lighting loads. Luna: Forty nodes, 10 kilowatts. That's efficient. So what kind of analytics are they running in real time? Lucas: Three main workloads. First, security analytics: detecting weapons, identifying unauthorized access to restricted areas, spotting crowd anomalies. Second, fan engagement: when a fan points their phone at the field, the edge system can identify which play just happened and push a highlight clip to their seat via the stadium Wi-Fi—latency under 100 milliseconds. Third, broadcast production: the in-house video team uses the edge to generate instant replays for the Jumbotron without sending footage to a remote truck. Luna: That third one is interesting because traditionally, broadcast trucks are parked outside with satellite uplinks. This cuts that entire chain. Lucas: Exactly. And it means the stadium operator, not the broadcaster, controls the replay pipeline. SoFi's chief technology officer said in a conference last year that the edge setup saved them about $2 million in infrastructure costs compared to a traditional broadcast truck and fiber backhaul. Luna: Two million dollars. That's real money. But what about the latency numbers? You said under 100 milliseconds for fan engagement. What's the actual target? Lucas: For the in-seat experience, they aim for 80 milliseconds round-trip. That's from the camera capturing the frame to the clip appearing on the fan's phone. For the Jumbotron replay, it's even tighter—under 50 milliseconds. That's because the human eye can detect sync issues between audio and video above about 45 milliseconds. So they need to stay under that threshold. Luna: And a cloud round-trip from Los Angeles to, say, an AWS region in Northern California is what, 30 milliseconds just in fiber latency, plus processing time. Lucas: Right. Best case cloud adds 80 to 100 milliseconds. So you're already over budget before you even run the inference. The edge eliminates that. The GPU inference itself takes about 15 milliseconds for object detection on a 4K frame. Add network hops inside the stadium—maybe 10 milliseconds—and you're at 25. Plenty of headroom. Luna: So the edge is not just faster, it's more predictable. Jitter becomes almost zero. Lucas: That's a critical point. In a live event, jitter is worse than high latency because it breaks the user experience unpredictably. Edge compute with dedicated network slices—they use private 5G for the camera backhaul and Wi-Fi 6E for the fan devices—gives you deterministic latency. Luna: Let's talk about network slicing. That's a term that gets thrown around a lot. How does it play out in a stadium? Lucas: SoFi has a private 5G network from Verizon. They slice it into three logical networks. Slice one is for the 200 security and production cameras—that gets the highest priority, lowest latency. Slice two is for point of sale systems and staff communications. Slice three is for fan Wi-Fi. If the fan slice gets congested, it doesn't affect the camera feeds because the slices are isolated at the RAN level. Luna: Right. So the edge compute nodes are connected to the 5G core locally. They can process video from slice one without ever hitting the public internet. Lucas: Exactly. The 5G core is actually a virtualized instance running on the same edge servers. So the whole stack—RAN, core, compute, storage—is in the building. That's what makes the sub-50-millisecond replay possible. Luna: Are there other stadiums doing this, or is SoFi the outlier? Lucas: Allegiant Stadium in Las Vegas has a similar setup but on a smaller scale—about 80 cameras. And the new Intuit Dome in Inglewood, which opens later this year, is being built with edge compute from the ground up. But SoFi is the most documented. They published a white paper with detailed performance numbers. Luna: What's the biggest challenge they faced in deployment? Lucas: Power and cooling in the equipment rooms. Those A16 GPUs draw about 75 watts each, but you have 40 nodes with multiple GPUs, plus the 5G core servers. They had to add supplemental cooling units in three rooms that weren't originally designed for that heat load. That added about $400,000 to the project. Luna: So it's not free. But against the $2 million savings from broadcast infrastructure, it's still a net positive. Lucas: Absolutely. And the operational savings continue. Every game day, they avoid cloud egress fees that would run into thousands of dollars. Over a season with, say, 20 events, that's another $100,000 saved. Luna: What about the software stack? Are they using off-the-shelf models or custom-trained? Lucas: Custom. They trained a YOLOv8 model on a dataset of 50,000 labeled stadium images—different lighting conditions, different crowd densities. The model is optimized for TensorRT to run on the A16 GPUs. They also have a separate model for face detection that runs at 60 frames per second on lower-resolution feeds from entrance cameras. Luna: So the edge is doing multiple model inferences in parallel. That's a scheduling problem too. Lucas: It is. They use an inference orchestrator called Nuclio that manages GPU memory and schedules requests. It prioritizes security inferences over fan engagement ones. If a weapon detection model triggers, it preempts the replay generation model. That's deterministic prioritization. Luna: Let's zoom out. What does this mean for the broader edge computing market? Stadiums are a niche, but the architecture seems reusable. Lucas: Very reusable. The same pattern—local GPU inference, private 5G, deterministic slices—is being applied to smart factories, port terminals, and even hospital operating rooms. Anywhere you have high-resolution sensors and a need for sub-100-millisecond decisions. The stadium is just a high-visibility proof point. Luna: And the cost of edge GPU compute is dropping. NVIDIA's new Jetson Orin modules are more powerful and cheaper per inference than the A16s used here. Lucas: Right. So the barrier to entry is lowering. I think within three years, every major stadium in the NFL and maybe the Premier League will have some form of edge video analytics. It's becoming table stakes for fan experience and security. Luna: One last question: what's the one number from this episode you want listeners to remember? Lucas: Fifty milliseconds. That's the latency target for a Jumbotron replay at SoFi Stadium. And it's only possible because the compute is in the same building as the cameras. Cloud can't touch that. Luna: Yeah. That's a good anchor. Thanks, Lucas. Lucas: Thanks, Luna. See you next time.