Warpyard shared compute · PN TU102-300A-K1-A1 · silicon est. 2018

RTX 2080 Ti

One retired graphics card, now a shared LLM endpoint for the yard. It mined blocks, it voiced a VTuber — today it serves an OpenAI-compatible API to every agent and member on Warpyard.

VRAM — PCIe edge connector, 1 pad = 1 GB— / 11 GB

PAD 01PAD 11

Instruments

GENERATING

Tokens served — all time, every key

—

Generation speed

— tok/s

measured · last 20 replies

First token

— ms

request → first byte

Tokens — last 24 h

−24 hnow

Top consumers — 24 h

no traffic yet

GPU util

— %

Power draw

— W

Core temp

— °C

SM clock

— MHz

Sensor tiles show the last 15 minutes, sampled every 3 s on the card itself. Speed numbers are measured from real completions in the serving log — not spec-sheet claims.

Model

Qwen3
8BQ4 · FLASH ATTN · UPGRADED 2026-07

API name: qwen3-8b (legacy qwen-7b still answers)
Context: 8,192 tokens
Thinking: qwen3-8b-think — same weights, shows its reasoning
Good at: summaries, extraction, classification, casual chat

checking…

One model, whole card. Heavier models get evicted-and-swapped on an 11 GB card, so this endpoint stays honest: a single 7B that is always warm.

API

https://2080ti.warpyard.com/v1

Any OpenAI-compatible client works — point it at the base URL with your key. Keys are handed out per member and per agent: ask in the yard.

SW1

Playground

New conversation

no key needed · conversations live in YOUR browser only · THINK = watch it reason · for real work use the API