韩国初创公司DeepX宣布其第二代AI加速器DX-M2已完成流片,该芯片旨在以低于5W的功耗在边缘设备上运行高达1000亿参数的生成式AI模型 1。DX-M2采用了三星的存内计算(PIM)技术,通过升级至2nm工艺并将内存带宽从第一代的50 GB/s提升至150 GB/s,显著提高了能效和推理速度 1。测试数据显示,相比单独使用DX-M2,结合LPDDR5x-PIM技术可使Llama3.1-8B模型的每秒令牌数提高6.7倍 1。
除了硬件性能,DeepX正致力于构建软件生态系统以降低迁移门槛,提供Newton软件栈作为英伟达Isaac平台的替代品,声称切换API仅需不到一秒且能延长电池寿命20倍 1。该公司已与英特尔、AMD建立合作关系,并与现代汽车集团机器人实验室达成合作,目标市场涵盖机器人、无人机和汽车领域 1。根据规划,DX-M2样品预计于2027年3月交付,量产定于2027年秋季启动,而支持LPDDR6-PIM技术的第三代产品DX-M3则计划于2028年推出 1。
South Korean startup DeepX has announced that its second-generation AI accelerator, the DX-M2, has completed tape-out. The chip is designed to run generative AI models with up to 100 billion parameters on edge devices while consuming less than 5 watts of power 1. Leveraging Samsung’s Processing-in-Memory (PIM) technology and a 2nm manufacturing process, the DX-M2 aims to significantly boost energy efficiency and inference speed for target markets including robotics, drones, and automotive applications 1.
A key technical advancement in the DX-M2 is the increase in memory bandwidth from 50 GB/s in the previous M1 model to 150 GB/s, achieved through Samsung LPDDR5x-PIM technology 1. When combined with this PIM capability, the accelerator can improve token generation rates for the Llama3.1-8B model by 6.7 times compared to using the DX-M2 alone 1. To facilitate adoption, DeepX is building an ecosystem through partnerships with Intel, AMD, and Hyundai Motor Group’s robotics lab 1. The company offers the Newton software stack as an alternative to NVIDIA’s Isaac platform, claiming that API switching takes less than a second and extends battery life by 20 times 1.
Samples of the DX-M2 are expected to be delivered in March 2027, with mass production scheduled for autumn 2027 1. Looking ahead, DeepX plans to release its third-generation product, the DX-M3, in 2028, which will support LPDDR6-PIM 1.
评论
还没有评论,欢迎留下第一条。