MatX 籌資5億美元挑戰Nvidia

🔑 增強重點摘要

• •MatX's latest round was led by Jane Street and Situational Awareness, with participation from Marvell Technology, NFDG, Spark Capital, and Stripe co-founders Patrick and John Collison[1].
• •The startup blends high-bandwidth memory for AI training with static random access memory for faster inference in a single chip design[1].
• •MatX plans to finalize chip design in 2026, ship products in 2027, and manufacture with TSMC using funding to secure capacity[1].
• •Previous funding included a $25M seed round in late 2023 led by Nat Friedman and Daniel Gross, followed by an $80M Series A in November 2024 valuing the company at over $300M[2][3].

🛠️ 技術深入

• •Combines high-bandwidth memory (used by Nvidia and Google for training) with static random access memory (for faster inference) in one product[1].
• •Targets AI workloads with at least 7 billion parameters, ideally 20 billion or more[3].
• •Prioritizes cost-efficiency over latency but promises <0.01 second latency per token for 70B parameter models[2].
• •Features advanced interconnect for scaling to clusters of hundreds of thousands of chips supporting millions of simultaneous users[2].
• •Aims for 10x better performance than Nvidia GPUs in LLM training and inference[3].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線