Confirm NVENC and the Twitch minted-key publish on a GPU host

[rainmanjam]

Two EXPERIMENTAL labels name specific unverified claims. A GPU host deletes both — and the labels were written so that confirming them removes a precise sentence rather than requiring a judgement call.

What is unproven, exactly:

• No NVENC, QSV, VA-API or AMF encode has been observed. All twelve encoder profiles were read off real FFmpeg option tables in a container; only VideoToolbox has run on silicon. TestEveryConfiguredEncoderOpensWithItsOwnFlags and scripts/acceptance-encoders.sh exercise this immediately on a GPU box.
• No broadcast has been published through a minted Twitch key. The negotiation itself already runs against ingest.twitch.tv on every test run and succeeds — Twitch accepts a declared inventory, grants a VOD audio track and mints a 314-character key. Everything after Negotiate returns has only been driven by an httptest server.

Needs: an NVIDIA GPU with NVENC — A100 and H100 have none; RTX 4090, L4, A10 or T4 do. Plus a real Twitch key and a GPU declared in settings.multitrack.gpus (Twitch validates the declared inventory, not the hardware).

Risk worth knowing: Twitch was measured refusing an Intel iGPU and unrecognised vendor IDs by name. A datacenter card may be refused as unsupported — which would itself be a finding, since no document currently says the feature is consumer-GPU-only in practice.

[rainmanjam]

Which GPUs actually have NVENC

Checked against NVIDIA's Video Encode and Decode GPU Support Matrix rather than assumed. The pattern is the inverse of price — NVIDIA strips the encoder from compute-tier silicon.

Cannot run this test — 0 NVENC engines

GPU	Architecture
H100 (SXM/PCIe)	Hopper
B200	Blackwell
A100	Ampere GA100
A30	Ampere

These are the cards a GPU host offers first and charges most for. All four are useless here.

Any of these works

GPU	Architecture	NVENC
L4	Ada Lovelace	1
L40 / L40S	Ada Lovelace	1
A10 / A40	Ampere	1
A16	Ampere	4
T4	Turing	1
V100	Volta	1
RTX 6000 Ada	Ada Lovelace	1
RTX A4000 / A5000 / A6000	Ampere	1
RTX 3050–4090 (consumer)	Ampere / Ada	1–2

Recommended: L4. Ada-generation NVENC (same as a 4090), cheap, and datacenter-standard so it is widely available. A10 is the Ampere fallback.

One caveat for the second half of this issue

Twitch validates the declared GPU inventory and was measured refusing unrecognised vendor IDs by name. A datacenter card may be refused as unsupported even where the encoder itself works — which would split this issue cleanly:

• NVENC confirmed → deletes the hardware EXPERIMENTAL label
• Enhanced Broadcasting still blocked → and we would have learned something no document currently says: that the feature is consumer-GPU-only in practice

[rainmanjam]

Correction / addition: the RTX 50 series (Blackwell consumer)

My earlier comment recommended an L4 and did not mention the 50 series. Checked against the same NVIDIA matrix:

GPU	Gen	NVENC engines
RTX 5090	9th	3
RTX 5080	9th	2
RTX 5070 Ti	9th	2
RTX 5070 / 5060 Ti / 5060 / 5050	9th	1
RTX PRO 6000 Blackwell	9th	— (workstation)

They also add 4:2:2 H.264/HEVC encoding, which Ada does not have — visible as the NO → YES flips between the two rows:

RTX 4090   Ada Lovelace  8th Gen  D 1 2 12  YES NO  YES YES YES NO  …
RTX 5090   Blackwell     9th Gen  D 1 3 12  YES YES YES YES YES YES …

The rule this establishes

B200 is Blackwell with 0 NVENC. RTX 5090 is Blackwell with 3. Same architecture, opposite answer — so it is not the generation that decides, it is the tier. NVIDIA strips the encoder from compute silicon (A100, A30, H100, B200) and keeps it everywhere aimed at graphics or video.

Better heuristic than any list: if the card is sold for training LLMs, assume no NVENC.

Revised recommendation for this issue

An RTX 5090 or 4090 beats an L4, for two reasons beyond raw capability:

1. Three encoder engines on the 5090 means the rendition ladder's three tiers can each hold their own engine — closer to what the ladder is actually claiming.
2. Twitch validates the declared inventory and was measured refusing unrecognised vendor IDs by name. A consumer GeForce is far more likely to be accepted for Enhanced Broadcasting than a datacenter L4 — which matters, because the EB publish is the half of this issue that a datacenter card might block on its own.

L4/A10 remain fine for the NVENC half alone if a consumer card is not available.