VPU use cases

Here’s evidence that purpose-built chips beat general-purpose compute.

225,000+

VPUs in production

1.1 Trillion+

Minutes of video encoded

1. Build dedicated systems

Purpose-built servers from NETINT and its partners, 20 to 384 channels per rack unit.

CUSTOMER · ZAPPING

75%

fewer servers

Graphic comparing servers: CPU (100 servers, 60 kW) vs VPU (25 servers, 6 kW) with caption: same workload—75% fewer servers and 90% less power.

75% Fewer Servers

Zapping, Latin America’s fastest-growing live OTT television platform, replaced 100 CPU-based encoding servers with 25 NETINT VPU servers and cut encoding power by 90%.

Ideal for: live sports and linear TV at low latency

  • Each VPU server encodes 50 to 60 channels with full 6-to-8-profile ABR ladders in H.264 and HEVC.
  • Servers draw 250 W instead of 600 W, taking the fleet from 60 kW to 6 kW for the same workload.
  • A NETINT server inside each broadcaster’s CDN puts Zapping within seconds of free-to-air, and sometimes ahead of it.

Sources: Transforming Live Streaming TV, One Rack at a Time

PARTNER · ADVANTECH

384

streams in one rack unit

Side-by-side comparison of two edge servers: Quadra Mini Server (left) with 1x Quadra T1M half-rack, 138 W, 20 streams; Advantech VEGA-7141 (right) with 12x Quadra T1U, 100 GbE, 384 streams; caption: Edge node left, origination server right, same silicon.

384 Streams in One Rack Unit

Advantech builds two Quadra systems: the VEGA-7141, a 1RU server with 12 Quadra T1U cards that encodes 384 live 1080p30 channels, and the Quadra Mini Server, a half-rack edge box with one T1M that encodes 20 channels at 138 W.

Ideal for: channel origination, edge CDN nodes, OB vans and remote production

  • VEGA-7141: 384x 1080p30 H.264, HEVC or AV1 on an AMD EPYC 9005 host with 100 GbE and SMPTE ST 2110; TV Technology Best of Show, NAB 2026.
  • Quadra Mini Server: 20x 1080p30 or 5x 4Kp30, 25x 1080p30 decode, 138 W total, 330x160x44 mm, SDI capture built in.
  • Quadra T1M: a full encoder on a 30×60 mm M.2 module under 10 W.

Sources: Beyond General-Purpose Compute · Quadra Mini Server · Quadra T1M

PARTNER · ZIXI

<7 ms

per frame at 32 sessions

Flowchart of video delivery: SDI/ST 2110/SRT/RTMP sources to Zixi Broadcaster, NETINT VPU, then packaging to CDN; Zen Master monitors.

Under 7 ms Per Frame at 32 Sessions

Quadra encodes a broadcast feed to HEVC or AV1 in under 5 ms per 1080p frame and stays under 7 ms with 32 sessions on one card. Zixi Broadcaster carries the signal in with error correction and hitless failover; ZEN Master runs the control plane.

Ideal for: broadcast contribution, remote production, channel origination

  • SDI via Blackmagic DeckLink, SMPTE ST 2110, SRT, RTMP, UDP MPEG-TS in; HDR10, HDR10+, HLG at 10-bit; SCTE-35 and SCTE-104 pass-through; CEA-608/708 and DVB subtitles.
  • One Quadra Video Server encodes 40x 4Kp60 or 160x 1080p60.
  • At 32 concurrent 1080p HEVC sessions, Quadra holds 19 fps per session against 8.5 on NVENC and 2.5 on x265 medium.

Sources: Reliable Live Video Meets Efficient Processing · 2026 VPU Buyer’s Guide · Broadcast page

PARTNER · I3D.NET

8 ms

cloud gaming latency

P2P DMA flow: CPU host orchestrates game logic → GPU tier renders AI inference → Quadra T2A outputs video to players via i3D.net network (1080p30, H.264/HEVC/AV1, 40 W).

Rendered frames move GPU to VPU without a trip through host memory.

GPU Renders. VPU Encodes.

Encoding on a GPU steals the compute that rendering and AI need. Quadra T2A pulls finished frames straight from GPU memory and encodes them, so one server renders, runs inference and streams at the same time. i3D.net delivers that architecture over its low-latency network with regional edge capacity.

Ideal for: cloud gaming, AI-enhanced sports production, GPU inference pipelines

  • T2A reads rendered frames from NVIDIA or AMD GPU memory over P2P DMA and encodes 64x 1080p30 at 40 W.
  • Cloud gaming: up to 400 players per 2RU server, 720p30 to 4Kp60, latency as low as 8 ms.
  • On-chip DNN engines (15 to 36 TOPS per card) handle ROI weighting and content-adaptive encoding without the GPU.

Sources: Scaling Performance-Critical Video Workloads · 2026 VPU Buyer’s Guide · Cloud Gaming · Quadra T2A

2. Rent it

Capacity by the hour, no hardware on your books.

PARTNER · AKAMAI

320

streams per instance

PlanVPUsvCPUsMemoryStorage$/hr
Quadra T1U x1 Small1816 GB200 GB$0.42
Quadra T1U x1 Medium11224 GB300 GB$0.53
Quadra T1U x2 Small21224 GB300 GB$0.73
Quadra T1U x8 Small84896 GB300 GB$2.93
Quadra T1U x8 Medium864128 GB300 GB$3.36
Quadra T1U x8 Large896192 GB300 GB$4.22

320 Streams Per Instance

Akamai is the first cloud provider to offer VPUaaS. Its Accelerated Compute instances run NETINT Quadra T1U VPUs, so streaming platforms rent encoding density by the hour instead of buying hardware.

Ideal for: live events and VOD at scale, with no CAPEX

  • One-, two- and eight-card plans, from $0.42 to $4.22 per hour. The 8-card plan runs more than 320 parallel streams per instance.
  • In Akamai’s own benchmark, a Quadra T1U drew 13 W against 82 W for a comparable GPU on a full ABR ladder: 4.7x the energy efficiency.
  • AV1, HEVC and H.264 encode and decode up to 8Kp60 with HDR, controlled through FFmpeg and GStreamer.

Sources: Akamai 8-card VPU plan · Akamai VPU vs GPU benchmark · Accelerated Compute pricing · NETINT Akamai partner page

PARTNER · NETACTUATE

15

global locations with hosted VPUs

Red central hub labeled Quadra T1U VPU-as-a-Service connected by dashed lines to multiple blue nodes, representing a network."Central service hub connecting to network nodes"

Hosted VPUs in 15 Global Locations

NetActuate rents Quadra T1U VPUs by the hour from 15 points of presence across its global network, with a free two-week proof of concept. Teams add encoding capacity before the next hardware cycle, in the region where the audience is.

Ideal for: regional capacity, burst events, trying VPUs before buying

  • 15 PoPs live with Quadra VPUs out of 50-plus worldwide, more on request; once an account is open, NetActuate reports a VPU VM deploys in under a minute.
  • 110-plus VPU appliances listed in NetActuate’s Media Marketplace.
  • Quadra T1U VM from $0.237 an hour, about $173 a month, with a free two-week proof of concept on real workloads and no hardware purchase.

Sources: VPU Capacity Without a Hardware Cycle · NetActuate expands to 15 locations

3. Add it to servers you own

A VPU card under 20 W turns an existing host into an encoding node.

PARTNER · ARCADIAN

50%

encoding OPEX cut

Bar chart comparing ABR ladders per 1RU and watts per stream for x264 CPU, Quadra T1A, and Quadra T2A, showing increasing power: ~41 W for x264 CPU, ~6.8 W for Quadra T1A, ~5.1 W for Quadra T2A.

Encoding OPEX Cut in Half

Arcadian takes Quadra VPUs from lab test to production on existing FFmpeg and GStreamer pipelines with a ten-step method, and its customers cut encoding operational cost by more than 50%. Only the encode stage changes; ingest, packaging and delivery stay as they are.

Ideal for: migration from software encoding, capacity on servers you already own

  • Per 1RU dual-socket host: x264 on CPU, 8 ladders at 41 W per stream; Quadra T1A, 26 ladders at 6.8 W; Quadra T2A, 52 ladders at 5.1 W.
  • Capped CRF and rate-control profiles tuned against real traffic, so CDN cost is predictable.
  • Add a VPU card under 20 W to an existing server: 10 cards in 1RU give 320 encodes and 480 decodes.

Sources: From Concept to Production · Arcadian partner page · Get More From the Servers You Already Own

TECHNIQUE

45%

below CBR at 4K

Line chart showing bitrate over a live program; CBR is 16 Mbps and a cap sits at 8 Mbps; the blue bitrate line stays below the cap across segments (talking head, fast action, replay). -- Informative chart note: quality constant; bitrate capped; average well below CBR.

Capped CRF: Predictable CDN Bills

Capped CRF holds quality constant and caps bitrate at a ceiling you set, so easy content costs less to deliver and hard content never bursts past budget. Quadra encodes those ladders at a fraction of the watts a CPU needs.

Ideal for: live OTT, CDN origin, FAST channels

  • Capped CRF at 4K cut bitrate 45% against CBR at the same VMAF (7.8 Mbps against 16 Mbps); talking-head H.264 dropped 76%.
  • 52 full ladders per 1RU on Quadra T2A at 5.1 W per stream, against 8 ladders at 41 W on x264.
  • Runs on servers you already own: keep the FFmpeg or GStreamer pipeline, swap only the encoder stage.

Sources: From Concept to Production · Capped CRF 4K test · Capped CRF with Quadra

CUSTOMER · MAYFLOWER

$8.6M

hardware saved

Horizontal bar chart comparing servers for the same workload by cost and power. CPU: 500 · M · 325 kW. NVIDIA T4: 105 · .36M · 112 kW. NETINT T432: 67 ·

Mayflower’s four platforms: servers, hardware cost, power draw.

$10M to $1.4M in Encoding Hardware

Mayflower serves 10,000 incoming live streams and more than a million outgoing streams on its own low-latency CDN. Replacing 500 CPU encoding servers with 38 Quadra T2 servers cut hardware spend 86% and power draw 90%.

Ideal for: UGC and live-commerce platforms, operator-owned CDNs

  • 500 Dell R940 CPU servers ($10M, 325 kW) became 38 Quadra T2 servers ($1.44M, 34 kW) for the same ladders.
  • Colocation fell from $750,000 to $57,000 a year; hardware cost per stream landed at $141.
  • Each input is one WebRTC feed to a five-rung ladder at 1 to 2 seconds of latency.

Sources: From CPU to GPU to ASIC: Mayflower’s Transcoding Journey · Understanding the Economics of Transcoding

PARTNER · V-NOVA (LCEVC)

23–53%

lower bitrate, same encoder

Two-bar chart comparing bitrates: Base codec only 16.58 Mbps (gray) vs Base + LCEVC 7.79 Mbps (blue).

Matched-encoder 4K retest, Universidade de Brasília: 53% less bitrate at equal quality.

23 to 53% Lower Bitrate, Same Encoder

LCEVC adds a small enhancement layer on top of Quadra’s H.264, HEVC or AV1 output instead of replacing the codec. Viewer tests show 23 to 53% bitrate savings at equal quality, so CDN bills fall without a player migration.

Ideal for: CDN cost reduction, 4K delivery on existing codecs

  • Enhancement layer costs 9 to 14% of the bitrate; MPEG verification found 24 to 46% savings across AVC and HEVC at HD and UHD.
  • A matched-encoder 4K retest dropped 16.6 Mbps to 7.8 Mbps (53%).
  • QoMEX 2026: 23% better subjective quality (BD-MOS) and 2.6x faster encode.

Sources: Pairing VPUs with MPEG-5 LCEVC · How LCEVC Works

CUSTOMER · API.VIDEO

99.3%

encoding cost cut

Line chart: cost per encoded minute falls as server load increases, with data points at ~0.5%, 5%, 25%, 50% and 90% and a note &apos; alt=

api.video’s published cost model, four-VPU server amortized over three years.

Encoding Cost Cut 99.3%

api.video moved its transcoding from CPUs to Quadra VPUs and the cost per encoded minute fell to one 150th of the previous price. The company then made encoding free at every quality, 4K included.

Ideal for: video API platforms, SaaS encoding services

  • One VPU encodes 2,000 megapixels per second of H.264, enough for 17 minutes of five-rung ladder output every minute.
  • A four-VPU server at $10,000, amortized over three years with about $1,000 a year in power, works out to $0.0002 per minute at 50% load.
  • H.264 quality matched FFmpeg x264 medium: VMAF 95.9 against 95.3.

Sources: api.video Revolutionizes the Industry with Zero-Cost Encoding · api.video: the numbers behind free encoding

4. Surveillance and AI video

Decode density and region-of-interest encoding for camera fleets.

PARTNER · ZONEMINDER

480

camera feeds per rack unit

Diagram of a camera processing pipeline: Cameras → Decode → Scale → Convert → AI (GPU) → Alerts, with note &apos;Quadra VPU · 48 feeds per card&apos; and wattage: GPU 2.5–3.6 W, VPU 0.4–0.7 W; benchmark line below.

480 Camera Feeds Per Rack Unit

AI analytics stall on decode, not on the model. Quadra VPUs turn compressed camera streams back into frames at 48 feeds per card and 480 per 1RU server, so the GPU runs inference instead of decoding. ZoneMinder, the open-source VMS, runs on Quadra VPUs today.

Ideal for: VMS platforms, command centers, AI video analytics

  • Decode 48x 1080p30 per VPU (H.264, HEVC, VP9, JPEG); 480x 1080p30 or 120x 4Kp30 per 1RU at about 500 W.
  • 0.4 to 0.7 W per stream, against 2.5 to 3.6 W per stream on a GPU.
  • Decode, scale and color-convert on the VPU, then hand frames to the AI model; Key Frame Detection drops 90% of static frames before inference.

Sources: Before AI can see: the decode bottleneck · ZoneMinder video · Quadra Video Server · Surveillance page

TECHNIQUE

90%

bitrate cut, same cameras

Two-panel infographic comparing video quality: left shows &apos;Before: full quality&apos; with two blue circles and a blue bar on a dotted grid, labeled 8.1 Mbps; right shows &apos;After: ROI quality only&apos; with red-outlined regions on a grid and &apos;830 kbps&apos; label.

Keep Every Camera. Cut the Bitrate 90%.

Quadra’s on-chip AI finds faces, plates and moving objects in each frame and tells the encoder where quality matters. The rest of the frame compresses hard, so existing 1080p cameras keep streaming at a tenth of the bitrate while new 4K and 8K cameras join the same server, and archives shrink with them.

Ideal for: municipal and campus upgrades, mixed-generation camera fleets, long retention

  • ROI encoding took a 1080p60 H.264 stream from 8.1 Mbps to 830 kbps in the 2026 Security Show demo, faces at full quality.
  • One decoder handles any frame up to 8192×8192: 4K at 125 fps, 8K at 60 fps, 10-bit.
  • Lower bitrate at capture means smaller nDVR archives for the same retention window.

Sources: J21 Security Show 2026 release · Surveillance page · ISC West 2026 page · Quadra T2A