Yinpu Launches AISSD to Cut LLM Inference Costs via Storage-Compute Tradeoff
创业邦 · wechat · 2026-08-28
Startup Yinpu proposes a "storage-for-compute" approach with its AISSD (AI Solid State Drive), aiming to reduce the cost of scaling AI Agents by offloading warm/cold data and intermediate results to lower-cost SSDs, alleviating pressure on expensive VRAM and RAM.
Core Points & Background:
- As model parameters approach trillions and Agents rise, soaring memory costs make pure GPU/RAM scaling inefficient.
- The bottleneck has shifted from compute to storage; the focus is now on efficient data relay and reduced waiting between storage hierarchies.
Products & Tech Roadmap:
- AISSD: Compatible with existing computers, it uses firmware and scheduling algorithms to optimize model data handling (weights, KV Cache, MoE experts), enabling larger models on consumer PCs.
- KLEENE Smart Control: Real-time monitoring and dynamic adjustment of voltage/frequency to maximize hardware performance.
- Chips: Developing near-memory compute coprocessors and inference ASICs to hardcode dense, repetitive compute tasks and resolve I/O bottlenecks.
Team & Philosophy: Founded by Tsinghua PhDs and former Huawei 2012 Labs staff, the company uses a "software-defined hardware" approach, designing chips based on actual inference workloads.
More from Embodied
- ZED Camera priced at $500? DIY alternative costs just $150 — _William_F_ · 2026-08-28
- Asimov 1 humanoid robot launched with sim-to-real capability — freelerobot · 2026-08-28
- Cleave Library: Visual GUI for Approximate Convex Decomposition — kevin_zakka · 2026-08-28
- Zero-WAM: In-Context World-Action Modeling for Robotic Task Generalization — Robbyant-Research · 2026-08-28
- Isaac 0.5 robot plays chess: perception, decision, actuation — code_star · 2026-08-28
- Forecast: 1 Billion Optimus Robots by 2036, Best-Selling Product Ever — tetsuoai · 2026-08-28