30-minute podcast pass

Transcribe an episode, then generate a synthetic read of the summary. The transcription half of the job most creators pay for monthly.

The pipeline

Transcribe 30 minutes of audio, whisper-large-v3
Write show notes from the transcript, qwen3-32b
Voice a 90-second intro, kokoro-tts

Fastest card where every stage is a real measurement: NVIDIA H100 80GB HBM3, 34 s for the whole job.

Every GPU, slowest job to fastest

GPU	Total time	Compute	Energy	Basis
NVIDIA H100 80GB HBM3	34 s	21 s	1.05 Wh	all 3 stages measured
NVIDIA H200	46 s	22 s	0.85 Wh	all 3 stages measured
NVIDIA A100 40GB SXM4	59 s	37 s	2.13 Wh	all 3 stages measured
NVIDIA A100 80GB SXM4	67 s	43 s	1.81 Wh	all 3 stages measured
NVIDIA L40S	75 s	33 s	1.89 Wh	all 3 stages measured
NVIDIA A10G	77 s	58 s	2.03 Wh	all 3 stages measured
NVIDIA L4	2 min	1.5 min	1.6 Wh	all 3 stages measured

Cards that can't run this pipeline

Every one of these fails on the same kind of wall, a stage that will not fit in VRAM.

NVIDIA T4, blocked at "Write show notes from the transcript" (Qwen3 32B)

How these numbers are built

Each stage time is the quantity of work divided by that card's measured throughput for that model, from our own bench. The pipeline is assumed to run batched, every image, then every clip, so each model loads once. Model load time is added where we recorded it; our text-generation runs don't carry a load measurement yet, so pipelines with a language-model stage are slightly optimistic. Nothing here is a single timed run of the whole pipeline, and we don't present it as one.