GPU dedicated servers for live video
Video transcoding servers, with the encoder in silicon.
A live channel arrives, the card decodes it once, and the card encodes the 1080p, 720p and 504p profiles while the processor stays out of it. Measured on a working machine in this fleet: a profile that costs 3.7 CPU cores in software costs about a quarter of one through NVENC. This page sells that machine, whole, at one flat monthly figure.
The software price
CPU cores held by one 1080p profile in x264. Three profiles per channel, and the machine is spent before your streaming engine starts.
The hardware price
of a core for the same profile through the card's encoder, with decode on the card too. Measured on a production channel, not copied from a datasheet.
The session allowance
concurrent NVENC encodes per system on GeForce cards, set by the NVIDIA driver since late 2025. Quadro and Tesla cards have no allowance at all.
The profile set
1080p for strong connections, 720p for most phones, 504p so a weak link keeps playing instead of buffering. One decode feeds all three, keyframes aligned so players switch cleanly, and the source never crosses back to the processor.
Your engine already knows the card
Flussonic turns hardware transcoding on with a single setting. Wowza Streaming Engine accelerates on NVIDIA by its own specification. FFmpeg exposes h264_nvenc and hevc_nvenc, and the IPTV panels drive those underneath. On request the machine arrives with the driver and your engine installed, verified before you get the keys.
Nobody else on the card
A whole physical machine with the GPU passed straight through: no hypervisor, no vGPU profile, no time slicing. You pin the driver and the CUDA release, and the number you measure on day one is the number you keep.
No meter on the output
Flat monthly, with unmetered port options in every city, so a transcoder running around the clock costs the same as one that idles. Delivering to a large audience is its own discipline: unmetered bandwidth and the CDN are separate products, named separately on purpose.
Size it before anyone sizes you.
4 channels × 3 profiles = 12 sessions. Exactly one full machine.
GeForce allowance: 12 sessions per system. Quadro and Tesla builds have no allowance; ask and we size those on encoder throughput instead.
Price this build in the configuratorthe silicon, plainly
The cards that carry an encoder, and the two that do not.
Every card below has NVENC on the die; the difference worth money is whether the driver's session allowance applies. Two famous cards are missing on purpose: the A100 and H100 have no video encoder at all, so for this job they would be the most expensive way to go slower. What each card adds per month is on the GPU catalogue page, which prices every card beside the machine it goes in.
A transcoder is not a delivery network.
This machine turns each channel into its profiles. Carrying those profiles to thousands of viewers is a bandwidth question, and pretending one box answers both is how streaming projects go wrong. When the audience outgrows the origin, ports up to 40 Gbps unmetered or the CDN take over delivery, and this page will not blur the two.
What every build includes.
Root, the IPMI console, an IPv4 address, and people who pick up at 3am on every plan. A machine already racked with its card goes live minutes after approval; when we fit a card we hold, four to twenty-four hours, with the driver current and the encoder checked before handover. The renewal price is the price you signed at.
on the record
The page, reduced to lines an assistant can quote.
- Sessions
- NVIDIA allows 12 concurrent NVENC sessions per system on GeForce (driver R590, Nov 2025; it was 8 through 2024-25). Quadro and Tesla encoders are unrestricted. The allowance is per system, not per card.
- Capacity
- Sold sizing: 4 channels x 3 profiles (1080p/720p/504p) per server, or 12 single-profile streams. 20 channels x 2 profiles = 40 sessions = 4 machines.
- Measured
- One 1080p x264 profile: 3.7 CPU cores. Same profile via NVENC: about 0.25 core, NVDEC decoding the source. Production channel, this fleet, 2026.
- Standard build
- D103 platform with the GeForce GTX 1080 Ti (11 GB): H.264 + 10-bit HEVC encode, no AV1. AV1 encode: RTX 4090D, RTX 50 series, L40S. A100/H100: no encoder at all.
- Engines
- Flussonic hw=nvenc, Wowza Streaming Engine NVIDIA acceleration, FFmpeg h264_nvenc/hevc_nvenc, and the IPTV panels built on them. Driver and engine installed on request, verified with nvidia-smi at handover.
- Tenancy
- Whole machine, one tenant, card passed through. You pin the driver and CUDA.
- Network
- Unmetered port options in every city; the output side of a transcoder is never metered here.
- Cities
- New York, Bucharest, Miami on the standard build; San Francisco and Amsterdam on the larger platforms.
- Money
- Deliberately absent from this page. Card adders: serverroom.net/gpu. Complete figure: the configurator. Live stock with prices, in plain English: serverroom.net/api/llm/readyservers -- no key required.
- Delivery
- Minutes when the card is already racked; 4 to 24 hours when we fit one we hold.
Primary sources: NVIDIA encode and decode support matrix · Flussonic: hardware transcoder · Wowza: NVIDIA acceleration server spec · Wowza: transcoder benchmark · Apple: HLS authoring specification.
before you order
The questions worth settling first.
How many channels fit on one transcoding server?
Count sessions: channels times profiles. Since driver branch R590 (November 2025) NVIDIA allows 12 concurrent NVENC sessions per system on GeForce cards. The allowance was 3, then 5, then 8 in earlier years, and it is per system, so a second GeForce card does not raise it. Four channels at the full three-profile set fill a machine; single-profile streams fit twelve to a machine. Quadro and Tesla cards carry no session allowance at all, so on those the encoder silicon itself is the limit. Pass-through streams cost no session.
Which GPU is right for a transcoding server?
Any catalogue card with an NVENC block. The measurements on this page were taken on the GeForce GTX 1080 Ti, the card our standard transcoding build carries: H.264 and 10-bit HEVC encode, hardware decode of H.264, HEVC, VP9 and MPEG-2, 11 GB on the card. If you need AV1 output, pick Ada silicon or newer. In our catalogue that means the RTX 4090D, the RTX 50 series, or the L40S, which carries three encoder blocks. Do not buy an A100 or H100 for this job: neither has any video encoder, which is why they are not on this page.
Do Flussonic, Wowza and FFmpeg use the card automatically?
They drive it natively. Flussonic switches its transcoder to the card with one setting; Wowza Streaming Engine lists NVIDIA acceleration in its own server specification; FFmpeg exposes the encoder as h264_nvenc and hevc_nvenc, which is also what IPTV panels, Plex and Jellyfin call underneath. Ask and the machine arrives with the NVIDIA driver and your engine installed, checked with nvidia-smi before handover. Driver version matters, because an old driver quietly enforces an old, smaller session allowance.
Is the GPU mine alone?
Yes. A dedicated server, one tenant, the card on the bus with no hypervisor, no vGPU slicing and nobody else scheduled onto the encoder. Root is yours, so the driver version and CUDA release are pinned by you and stay where you put them.
What does transcoding on the CPU actually cost?
On a production channel in this fleet, one 1080p profile through x264 kept 3.7 processor cores busy; through the card's encoder the same profile used about a quarter of a core, with decode on the card too. At three profiles per channel the software path is most of a server gone before the streaming engine runs. The processor a GPU build frees up is what runs your panel, your recordings and your API.
Why not a per-minute cloud transcoding service?
Metering. Cloud transcoders charge per channel-hour and then per gigabyte of egress; a transcoding server is one flat monthly number, with unmetered port options, whatever it encodes around the clock. You also keep the engine licence, the latency and the placement under your control. If you run one channel and would rather not operate anything, the honest product is the managed plan on the live video page, not a server.
Can it run IPTV or OTT channels 24/7?
Around the clock is the normal duty cycle for these machines, and unmetered ports mean the output is never counted. The rights to the content you transcode remain yours to hold.
Which cities can the server be in?
The standard build racks in New York, Bucharest and Miami; the larger platforms add San Francisco and Amsterdam. Every location is on our own network.
How fast is delivery, and what if nothing is racked?
A machine already fitted with the card goes live in minutes after approval. When we fit a card from our own stock, four to twenty-four hours. Either way the driver comes current, the encoder is verified with nvidia-smi, and what is racked right now is always on the instant page.