One GPU at home vs. a datacenter node

Both sides receive the identical stream of requests. Each lit tile is one sequence resident in KV cache. Watch which side runs out of tiles.

Self-hosted

0GPU tok/s

Resident batch

0 / 0 sequences0%

VRAM

Per request0 tok/s

Queue0

Wait for 1st token—

Full response—

GPU utilization0%

Timed out0

Datacenter node

0GPU tok/s

Resident batch

0 / 0 sequences0%

VRAM (per GPU)

Per request0 tok/s

Queue0

Wait for 1st token—

Full response—

GPU utilization0%

Timed out0

Throughput—node / home

Response—home / node

Paused

Model

Home

Node

Prompt 512

Output 400

Ctx 2048

Shape

Load 2.0 req/s

Try

View

queued prefill (reading your prompt) decoding (writing tokens) free KV slot timed out 1 spark = 60 tokens