Memory is a product constraint.
What the compact host handled
SD1.5 completed its four image-quality prompts at a 5.14-second median; two mixed-load images took a 7.95-second median. Adding image work raised queueing for other tasks. These are small workloads, not steady-state capacity estimates.
The shared SDXL attempt exhausted host RAM and its worker was killed. Nine failed requests remain in the evidence. A later isolated SDXL retest cannot be used to claim safe shared capacity.
Audio needs its own evaluation
Whisper WER was 38.0% on the twelve Swahili clips and 19.1% on twelve Nigerian English clips. Regional acoustic supplements were tiny convenience samples, with differing recording conditions; they do not support ranking accents. Kokoro generated English on CPU with a 2.33-second median across twelve requests.
Feasible is not production-ready
Wan 2.1 1.3B produced a 17-frame, 448×256 video at eight frames per second in 258.11 seconds, after 260.98 seconds of readiness work. It required an isolated run and heavy swap. This was a feasibility result, not a usable interactive video service.
Four corrected durable media jobs survived an API restart and idempotent replays. Earlier harness failures are retained; successful queue recovery does not make all co-resident models safe.
Results, in context.
| Workload | Result | Boundary |
|---|---|---|
| SD1.5 quality · 4 images | 5.14s median | 512px, warm loopback |
| SD1.5 mixed · 2 images | 7.95s median | Shared host |
| SDXL shared | 9 failures | Host RAM exhaustion |
| Kokoro EN · 12 generations | 2.33s median | CPU on shared host |
| Wan 1.3B · 1 video | 258.11s generation | Isolated; not production capacity |
Failures and isolated retests are intentionally not merged into one success average.
How we ran it.
- Cape Town g4dn.xlarge: one T4, four vCPUs, 16 GiB RAM. Gemma 4 E4B, Qwen retrieval and Whisper shared the host. SD1.5 and SDXL were tested one at a time alongside that core.
- Images used four matched prompts, seed 42, 25 steps and FP16. SD1.5 ran at 512px; SDXL used offload/slicing/tiling at 512px and 1024px. Core/media mixes used four concurrent clients on loopback.
- Whisper evaluated twelve Swahili FLEURS and twelve Nigerian English recordings, with an isolated supplement of six Kenyan, six South African and three Ghanaian clips from an EdAcc derivative. Kokoro used twelve synthetic English generations. Wan video ran separately with core workers stopped.
The limits matter.
- Loopback excludes customer networking. Small image sets were inspected qualitatively, with no automated aesthetic score.
- Isolated SDXL/video retests do not establish co-residency. GPU sampling may miss brief peaks. Speech samples may overlap model training; no general accent accuracy estimate.
- Synthetic billing rates tested arithmetic only. They are not customer pricing.
Evidence & sources.
The public extract contains the methods, selected results and caveats in this note. It is a curated summary, not the full raw run. Fingerprints identify the retained source artifacts; gated audio and private operational records are not included.
Download results extract- Google FLEURS · dataset and attribution
- Kokoro · model card
- Nigerian English dataset · Benjamin Ogbonna
- EdAcc · Edinburgh CSTR
Source artifact fingerprints · SHA-256
bdc82eae08967518ad702b964d2d6e3cdad41871317a42fffbc735f4b5f80bbb313ed32995503ae058061262e97da7261b2581786a867b2bcb72479c48e9a994Suggested citation: Embiro Labs Research. “How much media fits on one T4?” Published 12 Oct 2026. embiro.ai/research/shared-media-on-one-t4
Report a correction
