Video Transcoding Server: CPU, GPU, and What to Check Before You Rent

A video transcoding server takes one incoming video and turns it into several versions, at different resolutions and bitrates, fast enough to keep pace with your audience. In short, renting one comes down to three decisions: CPU or GPU, how many streams it must hold in real time, and which details the listing leaves out.

This guide explains what transcoding asks of a server, then covers each decision with real numbers. Then it shows why an AI training GPU has no video encoder, how to measure capacity on your own footage, and what to check before you pay. It also names the Swify plans that fit, and where GPUs are quoted to order.

What Does a Video Transcoding Server Do?

A video transcoding server decodes a source video and encodes it again as one or more other versions, each with its own resolution, bitrate or codec. Platforms do this constantly. For example, YouTube tells broadcasters that it automatically transcodes every live stream into many output formats, so viewers on any device and network can watch.

Each version is a rendition, and the set is an adaptive bitrate ladder. In practice, the player switches between its rungs as the viewer’s connection changes.

The work is heavy because transcoding decodes the input stream, scales it, and re-encodes it in each target format, as NVIDIA’s engineers describe. In addition, a live ladder multiplies it: four rungs mean four encodes for every incoming stream. A server that falls behind real time drops viewers, so speed matters more than average load.

Flow diagram of a live transcoding pipeline: one ingest stream is decoded, scaled once per rung, encoded once per rung, and sent to an origin or CDN.

Table of three transcoding jobs, live ladder, high-volume video on demand, and archive masters, with what each needs most and whether a GPU encoder or a CPU with a slow preset suits it better.

CPU or GPU: Which Video Transcoding Server Do You Need?

A CPU server suits offline encoding and maximum compression, while a GPU server suits live channels, where throughput matters most. Both follow one trade, because time spent on each frame buys bitrate saved.

When Does CPU Transcoding Win?

CPU transcoding wins whenever you can spend time to save bits. Software encoders such as x264, x265 and SVT-AV1 search harder at slow presets, so they reach better compression. Fora Soft’s encoder benchmark cites an analysis by Jan Ozer. At equal quality, x265’s slow preset cut the bitrate of a whole ladder by about 26% against its medium preset.

Furthermore, the same benchmark makes the economic point. Encoding cost is paid once, while the bandwidth saving is harvested on every view. A title watched a million times earns back a slow encode. By contrast, a clip watched a few hundred times does not.

Therefore, a CPU server is the natural home for video libraries and archives. Swify’s standard plans are CPU servers, with 10 to 40 physical cores, and the dual-socket Dedicated 7 and 8 reach 40.

When Does GPU Transcoding Win?

GPU transcoding wins when the encode must keep pace with a live source. In Fora Soft’s house benchmark on one reference machine, NVIDIA’s hardware H.264 encoder, NVENC, ran at 600 frames per second at 1080p. The x264 software encoder managed 95 at its medium preset and 12 at veryslow.

The hardware encoder also needed only about 5% more bitrate than x264 medium for the same quality. A fast CPU preset did worse: x264 veryfast ran at 240 frames per second but needed about 18% more. To match that throughput on a CPU, you would need a fast preset, and that costs bitrate. In this benchmark, the GPU was both faster and more efficient than x264 veryfast.

Horizontal bar chart of encoding speed at 1080p: NVENC H.264 on a GPU at 600 frames per second with plus 5 percent bitrate, x264 veryfast at 240 with plus 18 percent, x264 medium at 95 as the baseline, and x264 veryslow at 12 with minus 10 percent.

Even a 5% penalty adds up. At 6 Mbps per rung, it is 0.3 Mbps per viewer. Across 1,000 viewers, that is 300 Mbps of extra delivery, which is why libraries favour slow CPU presets.

Moreover, AV1 changes the maths again. NVIDIA reports that its Ada-generation hardware AV1 encoder saves about 40% bitrate against its own H.264 encoder at 1080p60. According to NVIDIA’s support matrix, hardware AV1 encoding needs an Ada or newer card.

Why Is an AI Training GPU Not a Transcoding GPU?

An AI training GPU is not a transcoding GPU, because NVIDIA’s support matrix lists no NVENC encoder on the H100, the A100 or the B200. A server sold for AI training can therefore be useless for encoding video.

However, only some cards carry encoder hardware, and the number of units differs. Inference cards such as the L4 and the L40S do carry it. A GeForce card also caps concurrent sessions, while most professional and data centre cards do not.

Table of NVIDIA cards showing NVENC units, concurrent session limit and AV1 encode support: GeForce RTX 4090 with two units and 12 sessions, L4 with two, L40S with three, Tesla T4 with one and no AV1, and H100 and A100 with no encoder at all.

More NVENC units raise total throughput only when you encode two or more streams in parallel, NVIDIA’s engineers note. A ladder does exactly that. Session caps used to be the catch, but NVIDIA has raised the GeForce cap several times. Its current matrix lists 12. Even so, the encoder’s own throughput can limit you before the cap does.

Are GeForce Cards Allowed in a Data Centre?

The GeForce driver licence says the software is not licensed for datacenter deployment, so ask any provider that offers GeForce cards how it complies. DCD reported the clause in 2018, with an exception for blockchain processing. NVIDIA said it added the provision to discourage misuse in large-scale enterprise environments. Licence terms change, so read the current text.

How Many Streams Can a Video Transcoding Server Hold?

Nobody can give an honest universal number, because capacity depends on the codec, the preset, the resolution and the number of renditions. Instead, you measure it on your own footage. YouTube gives the same advice for live setups: tests should include movement similar to the real stream.

Four steps give you a number you can plan around:

1. Cut two minutes of real content, with the motion your viewers will see.
2. Encode it with the exact encoder and preset you will run in production.
3. Read the speed figure FFmpeg prints at the end. A speed of 1.0x means real time.
4. Keep headroom for ingest, the operating system and spikes.

ffmpeg -i source.mp4 -c:v libx264 -preset medium -b:v 6000k -f null -

ffmpeg -i source.mp4 -c:v h264_nvenc -preset p5 -b:v 6000k -f null -

For example, say FFmpeg reports a speed of 4.0x for your full ladder. That is roughly four live ladders at the limit. Therefore, keeping a quarter of the capacity free leaves three. These numbers are illustrative, so replace them with your own.

How Much Bandwidth Do 20 Live Channels Need?

Twenty live channels need about 280 Mbps inbound and 226 Mbps outbound, which fits inside a 1 Gbps port. YouTube’s current table recommends 14 Mbps for a 1080p30 H.264 ingest, so twenty channels bring in 280 Mbps. Encode each into an illustrative four-rung ladder of 6, 3, 1.5 and 0.8 Mbps, and every channel sends out 11.3 Mbps. Each direction fits inside the 1 Gbps unmetered port that every Swify plan includes.

Bar chart of bandwidth for 20 live channels: 280 Mbps inbound ingest and 226 Mbps outbound ladder, compared with the 1,000 Mbps unmetered port included on every Swify plan.

What to Check Before You Rent a Video Transcoding Server

Check six things, in this order: the encoder hardware, the core count, the storage, the port, the access level and the drive class. A listing that hides any of them is asking you to guess.

Table of six checks before renting a video transcoding server, covering encoder hardware, CPU cores, storage, port, access and drive class, with the question to ask and why each matters.

The last row deserves a second look. Continuous encoding writes a lot, so ask whether the provider fits enterprise SSD storage. Consumer SSD in a Server: When It Works and When It Doesn’t explains how to check a drive’s write rating against your own workload.

Meanwhile, storage needs its own plan for a library. In the custom builder, a 26TB HDD costs €50 a month, which is €1.92 per TB. The builder offers None, RAID 0 and RAID 1, so mirror a pair if the library matters.

Which Swify Plan Fits a Video Transcoding Server?

Start with Dedicated 3 for a first test, because the plan page lists video transcoding among its best-fit workloads. Dedicated 5 offers the same 20 cores for €5 less, with more RAM and a larger drive, on a single socket. Choose Dedicated 7 or 8 when you need 40 cores, and add drives in the custom builder for a library.

Table of four Swify plans for video transcoding, Dedicated 3, 5, 7 and 8, with CPU, physical cores, RAM, storage and monthly price from 180 to 320 euros.

Also, faster ports are priced separately in the builder: 2 Gbps costs €60 a month, 5 Gbps €240, 10 Gbps €540 and 25 Gbps €1,320. GPU servers are not in the standard plans, so tell us the codecs and the channel count first. Plans bill monthly with no long contracts, which means a test on real hardware costs one month.

Video Transcoding Server: The Short Version

In short, a video transcoding server is judged on speed per stream, not on peak specifications. Use a CPU for compression and libraries, a GPU for live density, and measure before you commit.

• Match the hardware to the job: a CPU with slow presets for libraries and archives, a GPU encoder for live ladders.
• Ask for the exact GPU model, because H100, A100 and B200 cards have no NVENC encoder.
• Run your own footage through FFmpeg and read the speed figure before you size anything.

Frequently Asked Questions

What is a video transcoding server?

What is the difference between video encoding and transcoding?

Do I need a GPU for a video transcoding server?

How many CPU cores does a video transcoding server need?

Can I run FFmpeg and an RTMP server as a media server on a dedicated server?

How much bandwidth does a video transcoding server need?