AI Video Encoding
Fewer bits, same quality.
AI preprocessing combined with in-house ASIC encoding logic delivers matched quality at lower bitrates, plus the low latency live delivery requires.
2026 · REV. 1
AI Video Encoding at a glance
AI preprocessing
Frame analysis directs bits toward visually significant regions.
ASIC encoding
The U250 encoding ASIC keeps latency below GPU pipelines.
Standard codecs
H.264, HEVC, and AV1 integrate with existing workflows as-is.
Low latency
Sub-frame latency, as live delivery requires.
Production-scale performance
Bitrate savings at matched quality (VMAF), plus encoding latency.
How it works
Preprocessing through encoding on one card, scaling per channel.
Real-time encoding
A low-latency encoder that runs inline in the live pipeline.
Unified acceleration
AI preprocessing and ASIC encoding execute on the same card.
Multi-channel scaling
Multi-channel 4K per card; add cards for 8K and heavy traffic.
Against the conventional approach
The same outcome, at a different cost.
Where it applies
The approach differs depending on whether bitrate or latency is the bottleneck.
Re-encoding a catalog to cut delivery cost
As the catalog grows, storage and CDN egress costs accumulate — but lowering bitrate raises quality complaints.
AI preprocessing concentrates bits on visually significant regions, then the catalog is re-encoded with a profile that holds VMAF while lowering bitrate.
- +30–50% lower bitrate at matched quality
- +Storage and CDN egress fall together
- +Standard codec output means no player changes
Adding channels without adding servers
Every new channel means another encoding server, which also means more rack space, power, and operational staff.
Move to a configuration where one card handles multi-channel 4K concurrently, absorbing channel growth through card additions instead.
- +Fewer servers for the same channel count
- +Rack space and power budget reduced
- +Per-channel failover maintains stability
Workloads where latency defines the experience
Cloud gaming, teleoperation, and XR degrade noticeably at even tens of milliseconds of delay — a target software encoders struggle to hit.
The ASIC encoding path secures sub-frame latency, and keeping preprocessing on the same card removes the PCIe round trip.
- +Sub-frame latency through the encoding stage
- +Latency distribution (p99) stays predictable
- +Higher quality within the same latency budget
Key specifications
Technical specifications
Codec, throughput, and operational specifications.
- Video codecs
- H.264 · HEVC · AV1
- Bit depth
- 8 / 10-bit
- Rate control
- CBR · VBR · CRF · Capped VBR
- GOP
- Fixed · Adaptive · Scene-cut detection
- ABR renditions
- Up to 6 concurrent
- Concurrent 4K channels
- Up to 32ch / card
- Maximum resolution
- 8K (multi-card)
- Encoding latency
- p50 6ms · p99 8ms
- Bitrate savings
- 30–50% at matched VMAF
- Mode
- Live · Batch
- Metadata
- SCTE-35 · CEA-608/708 passthrough
- Quality reporting
- Per-channel VMAF and bitrate logs
- Control
- REST · gRPC API
- Redundancy
- Per-channel failover
Frequently asked questions
Anything not covered here, we answer in a technical meeting.
What conditions produce the 30–50% bitrate savings?
It compares the bitrate required to hold identical VMAF against a software encoder. The margin varies with content complexity: it is largest on low-motion, noisy content and smallest on already-optimized high-quality sources. We provide measurements on your actual content during PoC.
How can bitrate drop without quality dropping?
AI preprocessing analyzes each frame and separates what viewers perceive from what they do not. Elements that consume bits without contributing to perceived quality — noise being the obvious one — are cleaned up before encoding, and the remaining bits go where they matter. The result is the same perceptual quality expressed in fewer bits.
When should AV1 be used?
AV1 has the highest compression efficiency, so the delivery cost savings are largest. You do need to check playback device coverage. A common setup emits AV1 for modern devices alongside H.264 for legacy compatibility.
Can we compare quality against our current encoder directly?
Yes. During PoC we feed identical source into both in parallel and measure VMAF, bitrate, and latency together. We agree on the deciding metrics before starting.
How is latency distribution (p99) managed?
The ASIC path is a fixed-function pipeline, so per-frame processing time varies little. Without the scheduling jitter or garbage collection pauses software encoders experience, the gap between p50 and p99 stays narrow.
Your video infrastructure,
one level up.
Start with a single card and scale as needed. We design the demo, the rollout, and the technical review with you.