GPU dedicated servers for live video
GPU transcoding server.One decode, every profile on NVENC.
A GPU transcoding server is a dedicated machine whose NVIDIA GPU decodes each live channel once and encodes its 1080p, 720p and 504p profiles in hardware, so the processor stays free for your streaming engine. On a production channel in our fleet, one 1080p profile in x264 held 3.7 CPU cores; through NVENC it used about a quarter of one.
- Dedicated GPU, one tenant
- 12 NVENC sessions on GeForce
- Unmetered port options
- Five cities
This is the video transcoding server we sell: the whole machine and the full GPU, at one flat monthly price, with Flussonic, Wowza or FFmpeg installed on request. Count your channels and profiles below and you know how many servers to order.
CPU or NVENC
What live transcoding costs in processor time
Live transcoding turns one incoming channel into several outputs at different resolutions and bitrates, every second, for as long as the channel runs. Done in software, that work lands on the processor. Done on the GPU, a separate block of silicon handles it.
x264 in software
3.7
CPU cores held by one 1080p profile. At three profiles per channel, the processor is spent before the streaming engine starts.
NVENC in hardware
0.25
of a core for the same profile, with NVDEC decoding the source on the GPU too. Measured on a production channel, not taken from a datasheet.
Driver session limit
concurrent NVENC encodes per system on GeForce GPUs. Data-centre GPUs such as the L4, L40S and T4 have no session limit.
The default profile set
1080p for strong connections, 720p for most phones, and 504p so a weak link keeps playing instead of buffering. One decode feeds all three, with keyframes aligned every two seconds so players switch cleanly between them, as Apple's HLS authoring specification asks.
What the free processor is for
With the encode on the GPU, the cores run what the transcoder cannot: the streaming engine, an IPTV or media panel, recordings, DVR and your API. On a CPU-only server those jobs fight the encoder for the same cores.
Sizing
Transcoding server sizing: channels times profiles
Each output profile is one NVENC session. Multiply your live channels by the profiles each one needs, and divide by the 12 sessions a GeForce server allows. Pass-through streams cost nothing.
Twenty channels at two profiles each is 40 sessions, which is four servers. Above that, or for 4K channels, a data-centre GPU with no session limit usually means fewer machines: ask on chat and we size it on encoder throughput.
The GPUs
NVENC transcoding server GPUs, and the two to avoid
Every GPU below has NVENC on the die and is in our catalogue today. What separates them for this job is AV1 output, the number of encoder blocks, and whether the driver's session limit applies. The A100 and H100 have no video encoder at all. Each GPU's monthly price, beside the server it goes in, is on the GPU servers page.
- GeForce12 per systemGTX 1080 TiH.264, HEVC · 11 GB GDDR5XThe standard build. Every measured figure on this page came off this GPU.
- GeForce12 per systemRTX 3080H.264, HEVC · 10 GB GDDR6XAmpere NVENC: a cleaner picture than Pascal at the same bitrate.
- GeForce12 per systemRTX 3090H.264, HEVC · 24 GB GDDR6XSame encoder as the 3080, with 24 GB for filter chains beside the encode.
- GeForce12 per systemRTX 4090DH.264, HEVC, AV1 · 24 GB GDDR6XAda silicon: AV1 output on a GeForce part.
- GeForce12 per systemRTX 5090H.264, HEVC, AV1 · 3 NVENC · 32 GB GDDR7Blackwell with three encoder blocks behind the same 12-session driver limit.
- Data centreno session limitL4 ADAH.264, HEVC, AV1 · 2 NVENC · 24 GB GDDR6Two encoders, AV1 and no session limit, in a small data-centre part.
- Data centreno session limitL40SH.264, HEVC, AV1 · 3 NVENC · 48 GB GDDR6 ECCThree encoders and no session limit; encode next to inference.
- Data centreno session limitTESLA T4H.264, HEVC · 16 GB GDDR6Turing NVENC with no session limit, built to run flat out all year.
- Data centreno session limitTESLA P4H.264, HEVC · 8 GB GDDR5Low-power Pascal encoder with no session limit.
- Professionalno session limitTESLA P40 / QUADRO P6000H.264, HEVC · 24 GB24 GB of Pascal and no session limit, for wide channel counts.
- Computeno encoderA100 / H1000 NVENC · NVDEC onlyBuilt for training and inference. With no NVENC, the encode would fall back to the processor.
For 4K video transcoding: a 4K profile is one session but four times the pixels of 1080p. Every GPU here encodes 4K in H.264 and HEVC; plan 4K channels on a GPU with two or three NVENC blocks or on a data-centre GPU.
Software
Your transcoding engine already speaks NVENC
No special build is needed. The common engines drive the GPU's encoder through NVIDIA's own interfaces, and on request the server arrives with the driver and your engine installed and checked with nvidia-smi before you get root.
Flussonic
One setting, hw=nvenc, moves the transcoder from the processor to the GPU. See Flussonic dedicated servers.
Wowza Streaming Engine
Wowza documents NVIDIA acceleration in its own server specification for the Transcoder. See Wowza hosting.
FFmpeg and the panels on top
FFmpeg exposes h264_nvenc, hevc_nvenc and av1_nvenc. IPTV panels, Plex and Jellyfin call those underneath.
Driver version matters: an old driver enforces an old, smaller session limit. We have seen a 470-branch driver hold a customer to three sessions. Current drivers allow twelve on GeForce.
What you get
A whole server and the full GPU
One tenant per machine. The GPU is on the server's own PCIe bus with no hypervisor, no vGPU profile and no time slicing, so the throughput you measure on day one is the throughput you keep. You get root and the IPMI console, an IPv4 address, and you choose the driver and CUDA release.
A machine already racked with its GPU goes live minutes after approval. When we fit a GPU we hold, it takes 4 to 24 hours, with the driver current and the encoder checked before handover. Support is on chat: an AI assistant answers in seconds and can check your account and your server, and a person takes over when you ask. The renewal price is the price you signed at.
What it is not
A transcoder is not a delivery network
This server turns each channel into its profiles. Carrying those profiles to thousands of viewers is a bandwidth question, and one box should not be asked to answer both. On an unmetered port, a transcoder running around the clock costs the same as one that idles. When the audience outgrows the origin, unmetered ports up to 100 Gbps or the CDN take over delivery.
Standard build cities: New York, Bucharest and Miami. Larger platforms: New York, Miami, San Francisco, Amsterdam and Bucharest.
On the record
Every figure on this page, in plain lines
For a reader in a hurry, or an assistant answering for one.
| Sessions | GeForce: 12 concurrent NVENC sessions per system (NVIDIA driver limit). L4, L40S, T4 and other data-centre GPUs: no limit. |
|---|---|
| Sizing | Sold as 4 channels x 3 profiles (1080p/720p/504p) per server, or 12 single-profile streams. 20 channels x 2 profiles = 40 sessions = 4 GeForce servers. |
| Measured | One 1080p x264 profile: 3.7 CPU cores. Same profile on NVENC: about 0.25 core, NVDEC decoding the source. Production channel, this fleet, 2026. |
| Standard build | Intel Xeon E5-2600 v1/v2 server with a GeForce GTX 1080 Ti (11 GB GDDR5X): H.264 and HEVC encode, no AV1. |
| AV1 encode | RTX 4090D, RTX 5090, L4, L40S. Pascal, Turing and Ampere GPUs do not encode AV1. |
| No encoder | A100 and H100 have zero NVENC blocks. They are compute GPUs and do not belong in a transcoder. |
| Engines | Flussonic hw=nvenc, Wowza Streaming Engine NVIDIA acceleration, FFmpeg h264_nvenc / hevc_nvenc / av1_nvenc. Driver and engine installed on request, checked with nvidia-smi. |
| Tenancy | Whole server, one tenant, the full GPU passed through. You choose the driver and CUDA release. |
| Network | Unmetered port options in every city. |
| Cities | Standard build: New York, Bucharest and Miami. Larger platforms: New York, Miami, San Francisco, Amsterdam and Bucharest. |
| Delivery | Minutes when the GPU is already racked; 4 to 24 hours when we fit one we hold. |
| Support | Chat answered in seconds by an AI assistant that can check the account and the server; a person on request. |
| Prices | Not on this page. Per-GPU prices: serverroom.net/gpu. Complete figure: the configurator. |
Primary sources: NVIDIA encode and decode support matrix · Flussonic hardware transcoder · Wowza NVIDIA acceleration spec · Wowza transcoder benchmark · NVIDIA FFmpeg guide · Apple HLS authoring specification.
Answers
GPU transcoding server questions
What is a GPU transcoding server?
A dedicated server with an NVIDIA GPU that converts live or recorded video into other resolutions, bitrates and codecs in hardware. The GPU decodes the source once with NVDEC and encodes every output profile with NVENC, so the processor is left for the streaming engine, the panel and the recordings.
How many channels fit on one GPU transcoding server?
Count sessions: channels times profiles. NVIDIA's driver allows 12 concurrent NVENC sessions per system on GeForce GPUs, and a second GeForce GPU does not raise it. Four channels with three profiles each fill a machine; single-profile streams fit twelve to a machine. Data-centre and professional GPUs such as the L4, L40S and T4 have no session limit, so the encoder itself sets the ceiling. A pass-through stream uses no session.
Which GPU is best for a video transcoding server?
Any GPU with NVENC does the job. The standard build carries a GeForce GTX 1080 Ti: H.264 and HEVC encode, hardware decode of H.264, HEVC, VP9 and MPEG-2. For AV1 output choose Ada or Blackwell silicon, which in our catalogue means the RTX 4090D, RTX 5090, L4 or L40S. For more than 12 sessions on one machine choose a data-centre GPU. Do not use an A100 or H100: neither has a video encoder.
Do Flussonic, Wowza and FFmpeg use NVENC?
Yes, natively. Flussonic moves its transcoder to the GPU with one setting, hw=nvenc. Wowza Streaming Engine documents NVIDIA acceleration in its server specification. FFmpeg exposes h264_nvenc, hevc_nvenc and av1_nvenc, which IPTV panels, Plex and Jellyfin call underneath. On request the server arrives with the NVIDIA driver and your engine installed and checked with nvidia-smi.
Can a GPU transcoding server handle 4K video transcoding?
Yes. Every GPU on this page encodes 4K in H.264 and HEVC, and the Ada and Blackwell GPUs add AV1. A 4K profile counts as one session but carries four times the pixels of 1080p, so size 4K channels on encoder throughput: a GPU with two or three NVENC blocks, or a data-centre GPU with no session limit.
What does live transcoding cost on the CPU?
On a production channel in our fleet, one 1080p profile in x264 kept 3.7 processor cores busy. Through NVENC the same profile used about a quarter of a core, with decode on the GPU as well. At three profiles per channel the software path spends most of a server before the streaming engine starts.
Is the GPU dedicated to me?
Yes. The whole server is yours and the full GPU sits on its PCIe bus with no hypervisor, no vGPU slicing and nobody else scheduled on the encoder. You have root, so you choose the driver and CUDA release and they stay where you put them.
Why run a transcoding server instead of a cloud transcoding service?
Metering. Cloud transcoding services bill per channel-hour and then per gigabyte of egress. A transcoding server is one flat monthly price with unmetered port options, whatever it encodes around the clock, and you keep the engine, its settings and the latency under your control. For a single channel you would rather not operate, the managed plan on the live video page is the better fit.
Can it run IPTV channels 24/7?
Yes. Around the clock is the normal duty cycle for these servers, and on an unmetered port the output is never counted. You hold the rights to the content you transcode.
Where can the transcoding server be located?
The standard build is offered in New York, Bucharest and Miami. The larger platforms are offered in New York, Miami, San Francisco, Amsterdam and Bucharest, all on our own network.
How fast is a transcoding server delivered?
Minutes after approval when a machine with the GPU is already racked, and 4 to 24 hours when we fit a GPU we hold. The driver is current and the encoder is checked with nvidia-smi before handover. Machines racked right now are on the instant page.
Build the transcoding server
The configurator opens on the standard build with the GTX 1080 Ti selected. Swap the GPU, the memory or the port, and the price updates before you pay.
Elsewhere