Media server

This page describes performance characteristic for Smelter server running in different configurations on different hardware.

You can learn what affects performance of the specific element of the processing pipeline (decoding, rendering, encoding). What happens when they are GPU accelerated, and when they are not.

Rendering

Rendering process is implemented using wgpu crate (implementation of WebGPU standard). It supports running on top of different API e.g. Vulkan, Metal, DirectX, or OpenGL. Smelter supports Vulkan for Linux and Metal for macOS.

GPU

Rendering on a GPU it is always very cheap compared to decoding/encoding. It is true even on older hardware or integrated GPUs.

CPU

Rendering on a CPU is using LLVMpipe project to emulate GPU code on CPU. It is very inefficient in most cases, but you can make some steps to minimize the difference. For example:

• Enable CPU optimize mode (worse quality of color blending).
• Performance difference is smaller for lower resolutions.
• Simple layouts (no alpha blending, rounded corner, borders, box shadows).

Decoding/Encoding

Decoders/Encoders in Smelter come in 2 variants:

• Software based, that run on CPU. Supports H264, VP8, and VP9.
• Hardware based, that run on GPU using Vulkan extensions. Currently, it is limited to H264 only, but more codecs will be supported in the future.

Bottlenecks

For CPU decoding/encoding, actual decoding and encoding processes is a primary source of performance bottlenecks. However, it is not the only. You still need to upload raw frames to GPU for rendering, and download rendered output frames from GPU to encode them. Even on powerful CPU that upload/download will be a bottleneck at certain point.

You won’t have that problem on GPU. Vulkan decoder is producing frames as GPU textures, and encoder takes GPU textures as an input. The only transfer between RAM and VRAM (CPU memory and GPU memory) is passing encoded data which is a lot smaller.

Price-to-performance

GPU processing will almost always be a better option. However, there are certain exceptions where CPU decoding/encoding might be better.

• Software decoders can produce higher quality for a specific bitrate. However, it requires so much processing power that is often not feasible for real time use cases.
• Some cards might have limitations specific to encoding/decoding:
  • Older AMD cards have very limited custom hardware for decoding. For example, RX 570 can handle only 3-4 streams 1080p.
  • Nvidia limits the amount of concurrent encoder sessions to 8 on their consumer cards.

Price estimation

This is example pricing that shows cost of running the server in different variants.

Values below were calculated based on AWS pricing (pricing from Dec 4th 2025). We are using g4dn and c5 instance family and calculate result based 3-year hardware commitment pricing.

Rendering scenario assumes

• 2 input streams 1080p and one output stream 1080p.
• All inputs and outputs use H264.
• Hardware encoding is compared to fast and veryfast preset.

  This is conservative comparison, in reality Nvidia encoder might produce better quality, on par with even medium or slow.

Decoding	Rendering	Encoding	Price (per hour of stream)
GPU	GPU	GPU	0.028$/hour
GPU	GPU	CPU	preset: fast - 0.117$/hour preset: veryfast - 0.065$/hour
CPU	GPU	CPU	preset: fast - 0.156$/hour preset: veryfast - 0.104$/hour
CPU	CPU	CPU	preset: fast - 0.285$/hour preset: veryfast - 0.285$/hour

This example shows specific case, but you will see similar trend in other scenarios. Of course there will be some predictable differences, for example:

• If you have more inputs per output, then GPU decoding will be more important, and encoding not as much.
• Encoding quality and performance will differ between different GPU vendors and card models.
• If you have complex custom shaders, CPU rendering will be even more expensive.

Caution

Result for CPU only run are a best case scenario for that approach. Benchmark for that case were executed in different mode that is more optimized for CPU.

See raw benchmark data here for more specific cases.