算力商业
www.aibase.com ·
韩国 FuriosaAI 公布第三代推理加速器：算力与内存带宽均提升 32 倍
韩国 AI 芯片初创公司 FuriosaAI 公布第三代 AI 推理加速器技术细节，该产品与 Broadcom 合作开发，采用两颗 2nm 计算 die、一颗独立 I/O die、12 堆 48GB HBM4(E) 内存，机架内使用 SUE 全互联拓扑并支持 PCIe Gen7。官方称芯片物理尺寸约为第二代 RGND 的 8 倍，算力达 32 PFLOPS（32 倍），HBM 容量 576GB（12 倍），HBM 带宽 48TB/s（32 倍）。上述为厂商公布参数，尚无第三方实测或量产信息。
AI推荐理由 FuriosaAI 公布第三代推理加速器技术细节，含与 Broadcom 合作、2nm 计算 die、12 堆 HBM4(E)、32 PFLOPS、576GB 容量、48TB/s 带宽及 PCIe Gen7 等具体参数，属可核验的硬件规格信息，对 AI 芯片与推理算力方向有参考价值；但来源为 AIbase 转述，非官方首发，且未提供实测性能、功耗、量产时间与价格，参数为厂商宣称，需保留不确定性。
72
正文
South Korean FuriosaAI Unveils Third-Generation AI Inference Accelerator: 32x Improvement in Computing Power and Memory Bandwidth
AIbase基地
Published inAI News · 2 min read · Sep 22, 2026
9
Korean AI semiconductor startup FuriosaAI announced the technical details of its third-generation AI inference accelerator yesterday.
The accelerator was developed in collaboration with Broadcom, featuring two "mask limit-level" 2nm computing dies, one independent I/O die, 12 HBM4(E) memory stacks each with 48GB capacity, and using a SUE full interconnection topology within the rack, supporting PCIe Gen7.
FuriosaAI stated that the physical size of the chip is about eight times that of the second-generation chip (RGND); the computing power can reach 32 PFLOPS, 32 times that of RGND; the HBM memory capacity reaches 576GB, 12 times that of RGND; and the HBM memory bandwidth reaches 48TB/s, also 32 times that of RGND.
AI Inference AcceleratorFuriosaAI2nm Compute DieHBM4(E)
This article is from AIbase Daily
Scan to view
Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications.
—— Created by the AIbase Daily Team
© Copyright AIbase Base 2024, Click to View Source -