PFN Targets More Than Double Memory Bandwidth for MN-Core L2000
Preferred Networks (PFN) announced that it will continue working with SEMIFIVE and Samsung Electronics on the MN-Core L2000 AI inference accelerator, using a vertically stacked logic-chip and DRAM design. PFN says the L2000 targets more than twice the memory bandwidth of the L1000 series, while chip design is planned for completion in the first quarter of 2027; the bandwidth target is a vendor claim rather than an independently measured result.
The announcement identifies PFN’s vertically stacked logic-and-DRAM architecture and higher-memory-bandwidth target for a planned AI inference accelerator, providing a concrete update on the L2000 development roadmap.
IT之家 11 月 17 日消息,日本人工智能初创企业 Preferred Networks (PFN) 当地时间今日宣布,其在 AI 推理加速器 MN-Core L2000 上将延续与芯片设计服务企业 SEMIFIVE、芯片合同制造服务商三星电子的合作,继续打造垂直堆叠逻辑芯片与 DRAM 内存的产品。

PFN 的 MN-Core 系列初代产品 L1000 单芯片峰值 FP8 算力 200 TFLOPS,集成带宽超 10TB/s 的 32GB 片上内存。而此次官宣的二代产品 L2000 目标实现 L1000 系列 2 倍以上的内存带宽,进一步提升 AI 推理速率,改进 AI 工作负载能效。
PFN 计划 2027 年正式发布 MN-Core L1000,提供单芯、四芯盒子和 16 芯工作站解决方案。MN-Core L2000 系列也将于 2027Q1 完成芯片设计。
#Preferred Networks#SEMIFIVE#Samsung Electronics#MN-Core L2000#MN-Core L1000