Yogurt's Whisper server GPU selection on Linux
Raw Developer Origin & Technical Request
GitHub Issue
Nov 12, 2026
Hi! On my machine Yogurt's whisper-server runs on the AMD iGPU instead of the RTX 5090, which makes transcription ~40x slower than it could be. The bundled whisper-server sees both cards:
```
ggml_vulkan: 0 = AMD Ryzen 7 7700X 8-Core Processor (RADV RAPHAEL_MENDOCINO) (radv) | uma: 1 | ...
ggml_vulkan: 1 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | ...
```
but [the launcher](github.com/FrigadeHQ/yogurt/... doesn't pass `-dev`, so it takes device 0.
### Numbers
I ran the whisper-server from the 0.3.6 AppImage with the same args Yogurt uses (`-t 4 -bs 5 --vad`, large-v3-turbo q5_0) on the bundled `resources/speed/clip.wav` (15.4 s), one warm-up request then one timed:
| device | time | x clip length |
|---|---|---|
| default (0, iGPU) | 7.79 s | 0.506 |
| `-dev 1` (RTX 5090) | 0.18 s | 0.012 |
| `GGML_VK_VISIBLE_DEVICES=1` | 0.18 s | 0.011 |
The transcript was identical in all three, so it's purely a speed thing.
The speed check doesn't catch this, since it times whatever device whisper landed on:
- Yogurt's own check stored `0.458` for turbo here, just under the [0.5 cutoff](github.com/FrigadeHQ/yogurt/... With a slightly slower iGPU, or just an unlucky run (I got 0.506 above), it would fall back to small or say the computer is too slow, with the 5090 sitting idle
- the [speed key](github.com
Developer Debate & Comments
No active discussions extracted for this entry yet.
Adjacent Repository Pain Points
Other highly discussed features and pain points extracted from FrigadeHQ/yogurt.
Frequently Asked Questions
Market intelligence mapped to Yogurt's Whisper server GPU selection on Linux.
How is Yogurt's Whisper server GPU selection on Linux positioned in the market?
What is the general sentiment around Yogurt's Whisper server GPU selection on Linux?
Which technical concepts are associated with Yogurt's Whisper server GPU selection on Linux?
Engagement Signals
Cross-Market Term Frequency
Counts stored text matches for foundational terms like RTX 5090 and AMD iGPU in analysed public discussions. These counts do not measure adoption.
Stored text matches, not adoption or market-share measurements.
- RTX 5090: 3 records
- AMD iGPU: 2 records
- ggml_vulkan: 2 records
- whisper-server: 1 records
- integrated GPU (iGPU): 1 records
- discrete GPU (dGPU): 1 records
- -dev argument: 1 records
- GGML_VK_VISIBLE_DEVICES: 1 records
- transcription speed: 1 records
- large-v3-turbo q5_0: 1 records
- speed check: 1 records
- 0.5 cutoff: 1 records
SaaS Metrics