A VPU — Video Processing Unit — is dedicated silicon designed exclusively for video transcoding, distinct from a GPU (which is general-purpose graphics hardware that also happens to encode video) and from an ASIC (which is single-purpose silicon for one application). The VPU is built to do video work — encode, decode, scale, filter — and nothing else, which makes it dramatically more power-efficient and dense than GPU-based or CPU-based encoding for large-scale transcoding workloads.
The leading commercial VPU vendor is NETINT, whose Quadra series ships in PCIe cards and rack-mount appliances. A typical NETINT Quadra T1U card transcodes ~80 simultaneous 1080p H.264 streams or ~32 simultaneous 4K HEVC streams at around 50W — work that would require multiple CPU servers consuming 400W+ each. AMD acquired Pensando's DPU/VPU silicon and Xilinx's FPGA tools, building their own VPU offering. Marvell, Samsung and a few smaller vendors compete for hyperscale customers. Inside major streaming platforms, custom VPU/ASIC silicon (YouTube's Argos, Meta's MSVP) handles the bulk of internal transcoding.
For a product team, VPUs are the answer when you've outgrown CPU and GPU economics for video transcoding but don't have the budget or scale to build custom silicon like YouTube. The decision usually hits around 10,000–50,000 hours of video per day: below that, cloud CPU and GPU instances are cheaper to operate; above that, a NETINT or AMD VPU rack pays for itself within months in power, cooling and rack-space savings. Most product teams will never need VPUs, but creator platforms, sports broadcasters, video conferencing operators and OTT services at scale should evaluate them as their volume grows.
