QUALITY
Data quality & coverage
Live build-time stats · transparency · spot contribution opportunities.
Total entities
all types
94
Avg evidence
per hardware
1.9
China case coverage
cards with cases
8/14
Measured / total
evidence tier
22/97
Evidence tier distribution
📄 54
✅ 22
⚠️ 21
📄 Vendor-claimed 56% ✅ Measured 23% ⚠️ Estimated 22%
Goal: increase the measured (✅) share over time, reducing dependency on vendor-claimed numbers. Each measured evidence requires contributor attestation.
Hardware coverage gaps (21)
No deployment cases yet — high-priority contribution opportunity.
- AMD Instinct MI300A US
- Apple M4 Max Neural Engine US
- AWS Inferentia 2 US
- 壁仞 BR100 CN
- 寒武纪 MLU370-X8 CN
- Cerebras WSE-3 US
- 燧原 云燧 T21 CN
- Etched Sohu US
- Google TPU v5p US
- Groq LPU (TSP v1) US
- 昇腾 950 CN
- 海光 DCU Z100 CN
- Intel Gaudi 2 US
- NVIDIA B200 SXM 180GB US
- NVIDIA B300 SXM 288GB US
- NVIDIA GB200 NVL72 US
- … and 5 more
Models
No operator decomposition (0)
Calculator Tier 1 Roofline can't run without these.
No cases (6)
- MiniMax M2.7 minimax
- Mistral Small 4 mistral
- GLM-5 Reasoning zhipu
- Mistral Large 3 mistral
- AlphaFold 3 google
- GraphCast google
Single-evidence hardware (4)
Suggest adding corroboration from third-party benchmarks, MLPerf, or HotChips.
Engine adoption
| Engine | Supported hardware | Cases |
|---|---|---|
| vLLM | 32 | 15 |
| LMDeploy | 6 | 2 |
| MindIE | 3 | 2 |
| SGLang | 9 | 2 |
| TensorRT-LLM (Dynamo) | 9 | 1 |
| HanGuangAI | 1 | 0 |
| MoRI | 2 | 0 |
China hardware priority gap (6)
Highest-impact contribution opportunity — help Chinese accelerator ecosystem get transparent measured data.
Generated at build time. Run pnpm audit:data for the full audit.