NETINT · Creator of the VPU
Video encoding silicon.
Built for scale.
A VPU is a chip that does one job: encode video. NETINT invented it. Run VPUs in servers you own or on partner clouds using open source and commercial software integrations.
225,000+
VPUs in production
1.1 Trillion+
Minutes of video encoded
RECOGNITION
VPU Emmy®, alongside Google, Meta and AMD.
The 2024 Technology and Engineering Emmy® Award recognized NETINT for Design and Deployment of Efficient Hardware Video Accelerators for Cloud. The other recipients built chips for themselves. NETINT sells VPUs to everyone.
Why a VPU?
General-purpose compute can no longer keep up with codec complexity and power limitations.
No video silicon
AV1 and HEVC buy smaller streams with far more computing work. Software encoders rack up the bill in CPU cores. VPUs do the work in fixed-function logic, so AV1 costs no more to encode than HEVC or H.264.
Small video IP block
Platforms can encode hundreds and thousands of streams at once. CPU cost climbs with every stream. GPUs are slightly better because they are built for graphics and AI. VPU servers encode more than 320 streams.
Entire chip dedicated
to video
Racks are capped by watts, not slots. A card that encodes 32 HD streams at 17 W changes what one rack can carry. A 1RU VPU server encodes 320 streams in a power envelope of less than 500 watts.
MEASURED, NOT MODELLED
Tested against CPU and GPU.
VPU compared to x265
7.6xgreater density
HEVC at 32 concurrent 1080p sessions on identical Akamai hosts.
Source: NETINT Akamai Throughput and Quality Report, SDK V5.7.2.
VPU compared to GPU
4.7xfewer watts
Peak VPU power 13 W against 51 to 82 W for an NVIDIA RTX 4000 Ada on a live ABR ladder.
Akamai independent test results, Frankfurt, 19 August 2025.
VPU compared to x264
3xenergy efficiency
15 clips, H.264, HEVC and AV1, energy logged through OpenBMC.
Independent test by G&L Systemhaus, 9 July 2025.
PRICE AND PERFORMANCE
Cost per live HD stream, per month.
Same 1080p channel. Three ways to run it.
Monthly instance list price divided by sustained live 1080p streams: 2 on the CPU plan, 8 on VT1, 18 on the VPU plan. September 2025 prices. Your inputs will differ.
Where VPUs run
Built for the video you run.
Thousands of concurrent encodes, AV1 and HEVC, at a flat cost per stream.
Peak-event capacity with low-delay encoding and no rebuild of the pipeline.
Encoder latency under 5 ms at 1080p for gaming, betting and auctions.
Decode, downscale and run inference on hundreds of camera streams per chip.
One VPU: card, server or cloud.
M.2, U.2 and PCIe form factors. 20 to 64 HD streams per card for servers you own.
1RU with ten VPUs, 320 HD or 80 4K streams. Half-rack Mini Server for the edge.
Send us your stream count, codecs, resolutions, and latency target. A NETINT engineer will spec the cards, servers, or cloud instances to match, and show the power and rack-space math.
LEADERSHIP AND INNOVATION
A decade building the VPU category.
2015
NETINT founded in Vancouver, Canada to build video-only silicon
2019
Codensity G4, Logan ships as the T408 and T432
2021
Codensity G5, the Quadra family adds AV1, 8K and on-chip AI engine
2024
Technology and Engineering Emmy Award, alongside Google, Meta and AMD
2025
Akamai becomes the first infrastructure for video cloud to offer VPUs
2026
More than 225,000 VPUs deployed, encoding greater than 1 trillion minutes
What only a video-first chip can do
Encode AV1 at the same density as HEVC and H.264.
Decode, downscale and run an AI model on the same die.
Move frames straight from a GPU into the encoder.
Fit ten cards in one rack unit, enabling 320 live channels at broadcast quality.
VPU Ecosystem Partners
CDN & Platform
Software
Integrators & Services
Servers & Systems
