NVENC density: 25+ Full HD channels on one consumer GPU

"How many channels per GPU?" is the question that decides your capital cost, your power bill and how many boxes sit in the rack — and the one most encoder vendors answer with a shrug. So we measured it. In production, on a single RTX 5070 Ti — a consumer card — we run 25+ Full HD (1080p) channels comfortably, and a stress test pushed it past 30. Here is why that is real, and why the datasheet would have told you it was impossible.

The session cap is a policy, not a limit

Ask about GPU encoding density and someone will quote NVENC's concurrent-session cap — the number of simultaneous encode sessions a GeForce driver allows. It has crept up over the years, and professional cards (Tesla, RTX A-series) remove it entirely. But here is what operators miss: that cap is a driver licensing policy, not a property of the silicon.

NVENC is a dedicated hardware block on the die, separate from the CUDA shader cores that do graphics and compute. Encoding a stream barely touches the shaders. So the real ceiling on how many channels a card can encode is the throughput of that encode block — resolution × frame rate × channel count — not a number on a spec sheet, and not the GPU's compute load.

A session cap is a line in a driver's licence. Throughput is a property of the chip. Only one of them tells you how many channels you can actually run.

What we measured

The instrumented statmux run we published the VMAF for was a 14-session MPTS — 13 services plus a playout channel — sampled over three minutes. The revealing figure in that run was not the session count. It was the load: the GPU averaged 33% — a third of capacity. The encode block was working; the card was mostly idle. That headroom is exactly why the same RTX 5070 Ti carries 25+ Full HD channels in production, with a stress test taking it past 30 before anything strained.

Decode rides the same card: NVDEC, the hardware decode block, takes the incoming feeds, and when the same GPU decodes and encodes a stream the frames never leave GPU memory — zero-copy, no round trip to system RAM. That is a large part of why the density is what it is: the pipeline avoids the bottleneck that sinks naïve GPU transcoding.

Density is capital cost

Twenty-five-plus Full HD channels on one consumer card is not a benchmark boast — it is a line in the budget. The alternative to running that line-up on a €-few-thousand GeForce is a data-centre encode card — a Tesla A16, an RTX 6000 Pro — at several times the price, plus the power, cooling, rack space and spares that come with it. For the same channels.

Multiply that across a real plant and encoder density is one of the largest hardware decisions you make. A platform that needs three cards to do what another does on one is not a detail — it is your capital plan.

What actually limits density

If you are sizing this yourself, the real constraints, in order:

• NVENC throughput — resolution × frame rate × channel count against the encode block's capacity. Full HD is cheap; 4K is not. This is the true ceiling.
• Zero-copy path — decode and encode on the same GPU so frames stay in VRAM. Break that and the PCIe/host round trip becomes the limit long before the encoder does.
• VRAM — enough for every concurrent decode + encode context.
• CPU — filters, muxing and packaging still run on cores; NUMA-affine allocation matters at scale.
• The session cap — a real gotcha on consumer drivers, and a legitimate thing to check when you evaluate a platform, precisely because it is policy and not throughput.

What to demand

Never accept a spec-sheet session number as a density answer. Ask for a measured channel count on the actual card, at your resolution and codec — the sustainable production figure, not a stress-test peak — and ideally on a running system you can watch, not a slide. The gap between the policy number and the measured one is the whole point, and it is one of the questions worth asking every vendor.

Our numbers come off the running encoder, with the VMAF embedded live on the statmux page. Want the density figure for your resolution and channel mix? Tell us the line-up.