Hacker News Reader

Comment by est | Hacker News Reader

GPU would do transcoding, build network packet and copy data via PCI-E, all using hardware, avoid memory copy.

OBS+WebRTC is mostly software doing heavy-lifting.

Imagine if the camera would build WebRTC UDP packets directly and zero-copy to NIC, that would lower latency quite a bit.

View on HN

I wouldn't be surprised to learn that Nvidia is doing exactly that on their cloud: Compressing the video on the GPU using NVENC, building a package around it and then passing it to a NIC under the same PCIe switch (mellanox used to call that peerdirect) and sending it on its way.

The tech is all there, it just requires some arcane knowledge.

"arcane knowledge" is too strong of a phrase. You need someone who is familiar with Nvidia hardware and is willing to write software that only works on Nvidia hardware.

It is arcane as in information how all of this works on their specific hardware is not publicly available, but probably widespread within.

This is premature optimisation. The bus bandwidth and latency needed to get a few Mbps of compressed video to the PC is microscopic. It's completely unnecessary to lock yourself into NVIDIA just to create some UDP packets.

I was talking about Nvidia's Cloud gaming offer (GeForce Now). For them it's certainly not a premature optimization.

Exactly this with „…NVIDIA GPUDirect for Video, IO devices are fully synchronized with the GPU and the CPU to minimize wasting cycles copying data between device drivers“.[1]

1. https://developer.nvidia.com/gpudirectforvideo

Reply on news.ycombinator.com