Replacing a Discrete GPU with Intel iGPU for Jellyfin Transcoding

I migrated Jellyfin from an AMD RX 570 to the Intel HD Graphics 630 integrated into the host CPU. The result: 4 simultaneous HEVC transcodes at ~10% CPU each, with a ~105-135W power reduction.

Why Switch

The AMD RX 570 was overkill for transcoding. It pulled 50-70W at idle and 120-150W under load — running 24/7 in a media server that mostly handles a few streams at a time. The Intel iGPU built into the i5-7500 has Quick Sync Video, which is purpose-built for exactly this workload.

Hardware Changes

Shut down the host, physically removed the RX 570, and enabled the iGPU in BIOS under Advanced > Graphics > “Internal Graphics.”

After reboot, verified detection:

1
2
3
4
5
6

lspci | grep VGA
# 00:02.0 VGA compatible controller: Intel Corporation HD Graphics 630 (rev 04)

ls -la /dev/dri/
# crw-rw---- 1 root video  226,   1 card1
# crw-rw---- 1 root render 226, 128 renderD128

LXC Configuration

The Jellyfin container already had GPU passthrough configured. The LXC config passes both device nodes with the correct group IDs:

1
2

dev2: /dev/dri/card1,gid=44
dev3: /dev/dri/renderD128,gid=104

The Intel VAAPI driver (i965-va-driver) was already installed in the container. Verified with vainfo:

1
2
3

vainfo
# Driver: Intel i965 driver for Intel(R) Kaby Lake - 2.4.1
# Supports: H.264, HEVC, VP9 encode/decode

Jellyfin Settings

In the Jellyfin encoding config, the key setting is using VAAPI — not QSV. On Linux, QSV uses the VAAPI backend anyway, so VAAPI is the right choice for Intel iGPUs:

1
2

<HardwareAccelerationType>vaapi</HardwareAccelerationType>
<VaapiDevice>/dev/dri/renderD128</VaapiDevice>

Performance Results

Tested with 4 simultaneous transcodes, all requiring HEVC encoding with resolution scaling. All ffmpeg processes showed hardware acceleration active (hevc_vaapi):

The I/O wait was 83% during the test — the bottleneck was disk reads, not encoding. The iGPU had headroom to spare.

Estimated annual savings: 920-1,180 kWh ($100-150/year at $0.11/kWh), plus a cooler, quieter system.

VAAPI vs QSV on Linux

This tripped me up initially. Intel offers two APIs:

• VAAPI: Native Linux video acceleration API
• QSV (Quick Sync Video): Intel’s proprietary API

On Linux, QSV uses the VAAPI backend under the hood. Use VAAPI directly — it’s simpler and avoids an unnecessary abstraction layer.

There are also two driver options for the HD 630:

• i965: Older, stable driver for Gen 4-9.5 Intel GPUs
• iHD: Newer Intel Media Driver for Gen 8+

Both work. The ffmpeg processes chose iHD (driver=iHD), which may have newer codec support.

When iGPU Makes Sense

Intel iGPU is ideal for:

• 24/7 media server transcoding (power efficiency)
• Up to 5-7 simultaneous 1080p transcodes
• Setups where power and heat matter

Keep the discrete GPU for:

• Gaming VMs with GPU passthrough
• Heavy compute (AI inference, rendering)
• More than 7 simultaneous transcodes
• AV1 encoding (newer iGPUs support this)

Outcome

The RX 570 was freed up for gaming VMs on another host where it actually makes sense. The i5-7500’s iGPU handles Jellyfin’s transcoding workload at a fraction of the power draw. Right hardware for the right job.