Immersive video is about to turn every live event into millions of personalized streams. The infrastructure answer is not more servers; it is putting the right silicon in the right place.
NETINT VPU ECOSYSTEM · IBC 2026
AT A GLANCE
General-Purpose Compute alone cannot efficiently support the next generation of immersive video. Modern media architectures require CPUs, GPUs, and VPUs working together so each workload runs on the silicon best suited for the task.
This article explores how Advantech and NETINT combine heterogeneous architecture, spatial AI, and dedicated video processing to improve scalability, efficiency, and deployment for immersive streaming.
When professional football adopted the Video Assistant Referee at the 2026 FIFA World Cup, broadcasting quietly crossed a threshold. Synchronized high-definition cameras, processed in near real time, proved capable of reconstructing physical reality with enough fidelity to overturn a referee’s call.
Call it the volumetric truth: not a picture of the game, but a manipulable digital twin of it. If a room full of officials can scrub through a reconstructed moment from any angle, why can’t a fan?
The engineering vision taking shape across the streaming industry is essentially a personal VAR for every viewer: multi-camera capture feeding spatial AI, which uses techniques such as 3D Gaussian splatting to build a live, manipulable model of the event, which in turn delivers personalized virtual viewpoints to millions of concurrent users. Tactical overheads, courtside seats, a camera that follows one player all night.
The barrier is no longer data collection. What stands between the demonstration and the product is not imagination. It is infrastructure economics.
The immersion gap
Here is the uncomfortable asymmetry at the heart of live sports production: the technology used to capture the game is roughly a decade ahead of the technology used to watch it. Broadcasters already track millimeter-accurate, real-time 3D physics on the field. While the fan in their living room receives a flat, directed, one-size-fits-all 2D feed. Capture has gone volumetric; delivery has not.
Closing that gap changes the arithmetic of streaming. In conventional broadcasting, adding viewers meant adding replicas of the same feed. In immersive and personalized broadcasting, one physical event generates thousands, eventually millions, of unique output streams, each of which must be reconstructed, rendered, and encoded. The multiplication effect is the opportunity. It is also the trap.
The trap of borrowed hardware
The industry reflex, trained by a decade of cloud thinking, is to solve every scaling problem with a uniform pool of general-purpose compute. For immersive media, that reflex breaks down. If the pipeline relies on premium general-purpose GPUs to handle both complex spatial rendering and massive high-density video encoding, the total cost of ownership fractures. Expensive silicon built to understand a scene ends up spending its cycles on the routine, repetitive work of compressing one.
That is why the executive question has changed. It is no longer “Which server is fastest?” It is “Which workload belongs on which class of silicon, and how should the system be engineered so that silicon sustains its performance in production?”
This is the practical case for a third category of processor alongside the CPU and GPU: the video processing unit, or VPU.
Start with workload placement
A modern immersive or live pipeline has three materially different jobs, and it helps to give each one a role. The CPU is the orchestrator: it ingests synchronized SDI or ST 2110 feeds, preserves timing and data integrity, and manages workflow and control logic.
The GPU is the artist: it applies computer vision, 3D Gaussian splatting or other reconstruction methods to turn camera feeds into manipulable scene data.
The VPU is the engine: purpose-built silicon that converts the outputs into AVC, HEVC or AV1 streams at the scale, latency and cost the service model can support.
Run all three jobs on one accelerator class, and you strand expensive resources. GPUs are superb at massively parallel mathematics, but encoding thousands of routine renditions competes directly with the reconstruction workload that creates differentiation.
CPUs remain essential for orchestration, yet general-purpose cores are not optimized for sustained, high-density encode. A heterogeneous design is more disciplined: the CPU orchestrates, the GPU reconstructs, and the VPU processes video.
Crucially, this is a system-level decision, not a shopping decision. It implicates PCIe topology, memory and network bandwidth, thermal headroom, power delivery, software interfaces and product lifecycle. An accelerator advertised for high stream counts creates little value if it is starved for data, throttled by airflow or tied to a short recertification cycle.
Route each workload to the silicon that does it best
Figure 1. A heterogeneous media architecture keeps orchestration, spatial compute and codec processing on the silicon best suited to each job.
Three nodes, not one monolith
In deployment terms, the architecture resolves into three distinct nodes. Immersive edge capture lives at the venue, where synchronized feeds demand absolute precision timing, local NVMe storage and resilience; this node is the bridge between the physical camera lens and the digital twin, and its defining requirement is zero data loss.
Spatial AI compute lives in an on-premises data center or production truck, where premium GPU capacity is protected for volumetric reconstruction, inference and scene manipulation, thermally stabilized for continuous load. Media delivery lives at the outbound or cloud edge, where dense VPU capacity absorbs the repetitive codec work and personalized streams fan out to the audience.
The boundaries can move, but the jobs remain distinct, and separating them gives operators a cleaner scaling model. Capture scales with camera count. Spatial compute scales with scene complexity and latency targets. Delivery scales with renditions, viewpoints, and viewers. Decouple those axes and growth in one stage no longer forces overbuilding the entire pipeline.
Density is an operating metric
For conventional video, streams per server was a useful measure. Under real rack, power, and cooling constraints, streams per rack unit and streams per watt are more revealing; in a power-constrained facility, a denser node can be the difference between deploying capacity and waiting for it. Advantech and NETINT illustrate the spectrum with two deployment profiles built on the same VPU silicon strategy.
At the compact end, the Quadra Mini Server packages a NETINT Quadra T1M VPU into Advantech’s VEGA-6321-based 1RU half-rack platform, alongside a 13th Gen Intel Core CPU, DDR5 memory and NVMe storage, in a power envelope of roughly 138 watts. This small powerhouse is capable of processing 20 live 1080p30 streams across H.264, HEVC, and AV1, with optional SDI capture and support for FFmpeg, GStreamer, the NETINT SDK, and Bitstreams media processing framework. It answers a question many teams face: where to put dedicated VPU encoding when a full server deployment does not make sense, whether mobile, in-stadium or at the regional edge.
At the dense end, the Advantech VEGA-7141 supports up to twelve Quadra T1U VPUs on an AMD EPYC platform in a single rack unit. Capable of 384 simultaneous 1080p30 streams, with 100GbE networking and ST 2110 support, it’s easy to see why the system received a TV Tech Best of Show Award at the 2026 NAB Show. The point is not the large number. It is that capacity, networking, cooling and power are engineered together as one production system, which is what makes density deployable rather than theoretical.
One VPU strategy, two deployment profiles
Figure 2. One VPU strategy serves both compact edge and high-density production deployments; the selection criteria are operational, not only computational.
The server determines whether the accelerator delivers
The bottlenecks that kill real deployments rarely appear on benchmark charts. Accelerators rated for massive stream counts can be starved of data because generic PCIe topologies cannot provide sustained, non-contending bandwidth. Video processing runs 24/7, and a chassis designed for ordinary enterprise loads can recirculate heat around multiple accelerators until expensive silicon throttles and drops frames.
This is where the division of labor between the silicon topologies matters. NETINT supplies purpose-built video processing silicon; Advantech integrates it into media platforms where bandwidth, thermals, capture (SDI), and network interfaces are validated. For business leaders, that reframes the total cost of ownership conversation. Acquisition price matters, but so do rack expansion, power, cooling, recertification, deployment delay, and risks. A lower-cost server that forces more racks, or wastes GPU cycles on commodity encode, can easily be the more expensive choice.
Four questions for the next architecture review
- Are scarce GPU cycles reserved for video encoding? If premium GPUs spend most of their time on routine encoding tasks, the workload map deserves another look.
- Can each node scale independently? Camera count, spatial complexity and output streams rarely grow at the same rate. Be ready to increase GPU capacity to accommodate new AI features. At the same time, be sure to leave room in the rack for VPUs when more channels or camera angles are needed.
- Are performance claims sustainable? Ask about data paths, thermals and power under continuous 24/7 load, not only peak lab throughput. The lab is never production. Don’t assume “it worked in the lab” will carry through to the stadium or data center.
- Is the platform lifecycle aligned to the service? Broadcast and streaming products may outlive commercial IT refresh cycles. This means a server that would normally be depreciated over a 5-year time period could last 7 years or longer. Qualification costs must be modeled explicitly.
From accelerator choice to media system design
The broader lesson is that immersive media is unlikely to scale through indiscriminate acceleration but through silicon neutrality, where each workload is routed to the compute engine built for it. VPUs remove a high-volume, repeatable workload from general-purpose CPUs and premium GPUs, making room for richer experiences without collapsing delivery economics. The winners of the immersive era will not simply own more compute. They will place compute more intelligently.
NETINT VPU ECOSYSTEM · IBC 2026
As video workloads continue to grow, adding capacity is no longer a single decision, but a set of practical paths depending on the real bottleneck. This series explores how to scale video processing efficiently using NETINT VPUs, whether through infrastructure partners, existing systems, or dedicated deployments designed around video from the start.
1. The New Video Capacity Problem
2. Three Ways to Add Video Capacity in a Constrained Market
3. VPU-Powered Capacity as a Service
4. Get More From the Servers You Already Own
5. Build Dedicated VPU-Based Video Infrastructure
6. VPU Capacity Without a Hardware Cycle
7. Scaling Performance-Critical Video Workloads
8. Beyond General-Purpose Compute
9. Reliable Live Video Meets Efficient Processing
10. Maximizing Encoding Efficiency in Live Streaming Workflows
11. Pairing VPUs with MPEG-5 LCEVC Delivers The Next Layer of Video Efficiency
12. From Concept to Production: A Field-Proven Methodology for Deploying NETINT Quadra VPUs
Subscribe to the NETINT blog
One email a week with our latest articles on VPUs, encoding efficiency, and video infrastructure.
