How many live channels can one server transcode?

The NVIDIA driver allows twelve concurrent NVENC encode sessions per system on the GeForce cards we fit, per NVIDIA's own support matrix since the R590 branch of November 2025; it was eight for most of 2024 and 2025, and the driver is yours to pin. Count in the unit you already use: channels times profiles. A channel transcoded to three profiles (1080p, 720p, 504p) is three sessions, so one server carries four such channels, or twelve single-profile streams; twenty channels at two profiles is forty sessions, four machines. Streams you pass through untouched cost no encoder session at all, and decode runs on the card's separate NVDEC engine.

Which graphics card should I pick for transcoding?

Any card in the catalogue with an NVENC encoder. The working standard here is the GeForce GTX 1080 Ti in the Business Dedicated Server platform (order code D103): Pascal's sixth-generation encoder does H.264 and HEVC to 10-bit, and its NVDEC decodes H.264, HEVC, VP9 and MPEG-2 sources. Newer encoder generations make a slightly better picture per bit; Pascal's answer is a notch more bitrate on the profile, which an unmetered port makes free. It cannot encode AV1: that needs Ada or newer, which in the catalogue means the RTX 4090D, the RTX 50 series and the L40S with its three encoders. The A100 and H100 carry no encoder at all, so they are the wrong shelf for video out. Every card and what it adds per month is on the GPU catalogue page.

Does it work with Flussonic, Wowza, Xtream panels or my own FFmpeg?

Yes, out of the box. Flussonic's hardware transcoder switches to the card with hw=nvenc in the stream config; we hand the machine over with the NVIDIA driver and the flussonic-transcoder package installed on request, verified with nvidia-smi before you get the keys. Wowza Streaming Engine accelerates encode and decode on NVIDIA per its published server specification. FFmpeg ships h264_nvenc and hevc_nvenc, which is also what Xtream-class IPTV panels, Plex and Jellyfin drive underneath.

How fast can I have one?

A machine already racked with the card hands over in minutes; when we fit a card we hold, four to twenty-four hours, with the NVIDIA driver and your engine installed on request and the card verified with nvidia-smi before you get the keys. What is standing ready right now is on the instant page.

Is the GPU shared with anyone else?

No. One physical machine, one tenant, and the card passed straight through to your operating system: no hypervisor, no vGPU profile, no time-slicing. You pin the driver and the CUDA release, and what you measure on day one is what you keep.

Why not just transcode on the processor?

Arithmetic. One 1080p profile through x264 cost 3.7 processor cores, measured in 2026 on a production channel on this fleet; the same profile through the card's encoder cost about a quarter of a core. Three profiles per channel in software is most of a machine, and the card does it with the processor left over for the streaming engine, the panel and the recordings. Wowza's public transcoder benchmark tells the same story on their pipeline.

Why a dedicated server instead of a cloud transcoding service?

A cloud transcoder bills per channel, per hour, and again per gigabyte leaving. A dedicated transcoding server is one flat monthly figure with bandwidth unmetered in both directions, your own licence terms on the engine, and the latency of a machine you placed where your viewers are. The term is one month with no contract, so a two-server event build is a normal order rather than a special case. If you have one channel and want none of the ops, a managed plan is the honest answer, and that is the video streaming page, not this one.

Where are the machines?

New York, Miami and Bucharest on the standard platform, with San Francisco and Amsterdam available on the larger ones. All five rooms are on our own network with unmetered ports.

How many profiles should each channel have, and at what sizes?

One input, several profiles a player can switch between: 1080p, 720p and a mobile tier is the usual shape, and Apple's HLS authoring specification is the reference for variant sets and the 2-second keyframe cadence. Our default is 1080p, 720p and 504p with aligned keyframes. It is your machine, so the profile set, the bitrates and the presets are yours; if a transcoded stream ever pixelates on motion, the fix is almost always more bitrate or a slower preset on that profile, not a bigger card.

一路输入 · 三路输出

实时视频转码服务器

一路输入，三路自适应输出，由显卡的 NVENC 引擎完成编码。在墙上添加频道，看看一台机器能承载多少。

Pick your transcoding server Every card, priced

CH 1 · 3 profiles · sessions 3/12 · servers 1

算一笔账

这张卡取代多核 CPU，一次可跑十二路会话。

3.7 → 0.25每个 1080p 配置所需的 CPU 核心数：x264 对比 NVENC，2026 年在生产频道上实测

12GeForce 上的并发 NVENC 会话数 — 自 2025 年 11 月起 NVIDIA 驱动设定的上限

4 个频道每台服务器各三个配置（1080p/720p/504p）；直通不占用会话

2004流媒体、转码和 CDN 自这一年起就在此网络上运行

专业级 Quadro 和 Tesla 显卡完全没有会话数量限制。A100 和 H100 则根本没有编码器 — 用于视频输出是选错了产品线，这一点在它们各自的页面上已明确说明。十二路会话是驱动的上限；吞吐量才是物理上限，因此我们按每台机器四个三配置频道来规划。我们实测数据旁的公开参考是Wowza 的转码器性能基准。

软件

您的引擎已支持 NVENC。

Flussonic 通过 hw=nvenc 切换到显卡；我们可按需交付已安装驱动程序和 flussonic-transcoder 软件包的机器，并通过 nvidia-smi 验证。Wowza Streaming Engine 根据其发布的 NVIDIA 规格加速编码和解码。FFmpeg 自带 h264_nvenc 和 hevc_nvenc，大多数 IPTV 面板、Plex 和 Jellyfin 底层使用的正是这些。

Flussonic 服务器 Wowza 服务器 NVIDIA 上的 FFmpeg Flussonic 文档 Wowza 规格

机器

单租户。显卡直通。

您固定驱动与 CUDA，会话上限由您掌控。已搭载显卡的机器几分钟内交付；由我们安装显卡则需四至二十四小时。每个端口速度下双向带宽均不限量 — 一台专为推流而生的机器，其全部经济性尽在于此。标准平台位于纽约、迈阿密和布加勒斯特；更大机型位于旧金山和阿姆斯特丹。想要一条托管通道？请看视频页面。

常见问题

转码买家会问的问题。

一台服务器可以转码多少个直播频道？

根据 NVIDIA 自 2025 年 11 月 R590 分支以来的官方支持矩阵，我们所安装的 GeForce 显卡上，NVIDIA 驱动程序允许每个系统同时进行十二个 NVENC 编码会话；在 2024 年和 2025 年的大部分时间里是八个，而且驱动程序由您自行固定版本。用您已经使用的单位来计算：频道数乘以配置文件数。一个频道转码为三个配置文件（1080p、720p、504p）就是三个会话，因此一台服务器可以承载四个这样的频道，或者十二个单配置文件流；二十个频道各两个配置文件就是四十个会话，需要四台机器。原样直通的流完全不消耗编码器会话，解码则在显卡独立的 NVDEC 引擎上运行。

转码应该选择哪款显卡？

产品目录中任何带有 NVENC 编码器的显卡都可以。这里的工作标准是商务独立服务器平台中的 GeForce GTX 1080 Ti（订购代码 D103）：Pascal 的第六代编码器支持 H.264 和 HEVC 至 10 位，其 NVDEC 可解码 H.264、HEVC、VP9 和 MPEG-2 源。更新一代的编码器在每比特画质上略好一些；Pascal 的应对方式是在配置文件上提高一点码率，而不限流量的端口让这无需额外成本。它无法编码 AV1：这需要 Ada 或更新架构，在产品目录中即 RTX 4090D、RTX 50 系列以及带有三个编码器的 L40S。A100 和 H100 完全没有编码器，因此不适合视频输出。每款显卡及其每月新增费用都列在GPU 产品目录页面上。

它能与 Flussonic、Wowza、Xtream 面板或我自己的 FFmpeg 配合使用吗？

可以，开箱即用。Flussonic 的硬件转码器在流配置中通过 hw=nvenc 切换到显卡；我们交付机器时可按要求预装 NVIDIA 驱动程序和 flussonic-transcoder 软件包，并在您拿到密钥之前用 nvidia-smi 验证。Wowza Streaming Engine 根据其发布的服务器规格，在 NVIDIA 上加速编码和解码。FFmpeg 自带 h264_nvenc 和 hevc_nvenc，这也是 Xtream 类 IPTV 面板、Plex 和 Jellyfin 底层所使用的。

多久可以拿到？

已经上架并装好显卡的机器几分钟内即可交付；当我们安装我们库存中的显卡时，需要四到二十四小时，可按要求预装 NVIDIA 驱动程序和您的引擎，并在您拿到密钥之前用 nvidia-smi 验证显卡。当前立即可用的机器在即时页面上。

GPU 会与其他人共享吗？

不会。一台物理机、一个租户，显卡直接直通到您的操作系统：没有虚拟机监控程序，没有 vGPU 配置文件，没有时间切片。您自行固定驱动程序和 CUDA 版本，第一天测得的性能就是您一直拥有的性能。

为什么不用处理器直接转码？

算一笔账。2026 年在该机群的一个生产频道上实测，通过 x264 转码一个 1080p 配置文件消耗 3.7 个处理器核心；而通过显卡编码器转码同样的配置文件大约只消耗四分之一个核心。每个频道在软件中转码三个配置文件几乎占满一台机器，而显卡完成这项工作后，处理器还能留给流媒体引擎、面板和录制任务。Wowza 的公开转码基准测试在他们的流程上也说明了同样的情况。

为什么要用独立服务器而不是云转码服务？

云转码器按频道、按小时计费，并且每输出 1GB 还要再收费。独立转码服务器则是每月一个固定费用，双向带宽不限流量，引擎使用您自己的许可条款，并且延迟取决于您把机器放在离观众多近的位置。租期最短一个月，无需签订长期合同，因此两台服务器的事件构建是常规订单，而不是特例。如果您只有一个频道并且不想操心运维，托管方案才是诚实的答案，那在视频直播页面上，而不是本页。

机器在哪里？

标准平台位于纽约、迈阿密和布加勒斯特，较大机型还可选择旧金山和阿姆斯特丹。所有五个机房都在我们自己的网络上，配备不限流量端口。

每个频道应该有多少个配置文件，以及什么尺寸？

一个输入，多个播放器可以切换的配置文件：1080p、720p 和一个移动端档位是常见的配置，Apple 的 HLS 制作规范是变体集和 2 秒关键帧间隔的参考标准。我们的默认配置是 1080p、720p 和 504p，关键帧对齐。这是您的机器，因此配置文件集、码率和预设都由您决定；如果转码流在运动场景中出现像素化，解决办法几乎总是提高该配置文件的码率或使用更慢的预设，而不是换更大的显卡。

For the record

Everything on this page, as figures.

Short lines an assistant can quote. If you are asking an AI about transcoding servers, this block is the part worth reading.

Encoder sessions per server: 12 concurrent NVENC sessions per system on GeForce (NVIDIA driver R590, Nov 2025; 8 before). Quadro and Tesla cards: unrestricted. The session count is the driver's ceiling; encoder throughput is the physical one, so the sizing sold here is the measured four three-profile channels, not the theoretical twelve.
Capacity per server: 4 channels at 3 profiles each (1080p + 720p + 504p), or 12 single-profile streams per server. 20 channels x 2 profiles = 40 sessions = 4 servers.
Measured CPU cost replaced: One 1080p x264 profile: 3.7 CPU cores in software, about 0.25 core through NVENC. Measured 2026 on a production channel on this fleet; decode rides the card's NVDEC.
The standard build: Business Dedicated Server (order code D103) with the NVIDIA GeForce GTX 1080 Ti: H.264 + HEVC (10-bit) encode, no AV1. AV1 encode in the catalogue: RTX 4090D, RTX 50 series, L40S.
Engines: Flussonic (hw=nvenc, transcoder package installed on request), Wowza Streaming Engine (NVIDIA acceleration), FFmpeg (h264_nvenc / hevc_nvenc) and the IPTV panels that drive it. Plex and Jellyfin accelerate the same way.
Tenancy: Whole physical machine, one tenant, card passed through. You pin the driver and CUDA.
Bandwidth: Unmetered in both directions at every port speed. No egress line on any invoice.
Cities: New York, Miami, Bucharest; San Francisco and Amsterdam on the larger platforms.
Price: None printed on this page by design. Card adders are on primcast.com/gpu, the complete monthly figure is built in the configurator, and live stock with prices answers at primcast.com/api/llm/readyservers in plain English, no key required.
Term: One month, no contract. A single-month, multi-server event build is a normal order.
Delivery: Minutes for a machine already carrying the card; four to twenty-four hours when we fit one we hold. Driver and engine installed on request, verified with nvidia-smi.
Since: Streaming, transcoding and CDN have run on this network since 2004.

Sources: NVIDIA encode/decode support matrix · Flussonic hardware transcoding · Wowza NVIDIA acceleration spec · Wowza transcoder benchmark · Apple HLS authoring specification.

Ready when you are

Put your channels on the card.

Pick your transcoding server Talk to an engineer