I reverse-engineered the TiinyAI Pocket Lab from marketing photos
3 days ago
- #AI-hardware
- #Kickstarter
- #tech-scam
- TiinyAI Pocket Lab is marketed as a pocket-sized AI supercomputer capable of running 120B parameter models locally at 20 tokens per second for $1,299.
- The device uses a split memory architecture with 32GB on the SoC and 48GB on the dNPU, connected via a slow PCIe Gen4 x4 bus (8 GB/s), creating a significant bottleneck.
- Performance degrades sharply with context length: 16.85 tok/s at 256 tokens, 4.47 tok/s at 65K tokens, and TTFT (Time To First Token) reaches 28 minutes at 64K context.
- The hardware consists of a CIX P1 ARM SoC (12-core, 30 TOPS NPU) and a VeriSilicon VIP9400 dual-die NPU (160 TOPS), both using LPDDR5X memory but split into non-unified pools.
- TiinyAI's software relies on PowerInfer, an open-source project from Shanghai Jiao Tong University (SJTU), which they forked and rebranded as proprietary.
- The company's marketing omits critical details: MoE models (e.g., GPT-OSS-120B activates only 5.1B parameters/token), split memory, and PCIe bottleneck.
- Corporate opacity: No named CEO/CTO, only a GTM director with no prior history, and ties to Hong Kong investors (Gobi Partners) and Chinese semiconductor firms (VeriSilicon).
- Kickstarter raised $1.7M (17,377% of $10K goal), but the $10K target suggests the campaign was for marketing, not funding production.
- The device is a USB-C peripheral requiring a host computer, not standalone, and locks users into TiinyAI's curated model store due to NPU compilation requirements.
- An honest description would highlight: 12-16 tok/s (short context), 6-9 tok/s (8K-32K), MoE models, split memory, and reliance on SJTU's PowerInfer.