Groq reportedly shifts next-gen LPUs to 3D-stacked DRAM
IT之家, cnBeta, Wccftech and 36氪 report that Groq plans to move its next-generation AI inference LPUs from large SRAM caches toward TSV-based 3D-stacked DRAM. The reported shift would prioritize greater memory capacity over SRAM speed, though no official Groq announcement, shipping schedule or final product specifications were provided. Wccftech also reports that current Groq LPX racks entered full production in August and early tests with NVL72 servers recorded 3,400 tokens per second.
If confirmed, Groq’s reported architecture change would materially alter the memory strategy of an AI inference accelerator designed for larger models, but the plan remains based on industry reporting rather than a formal company announcement.
IT之家 10 月 21 日消息，韩媒 ZDNET Korea 当地时间昨日报道称，NVIDIA（英伟达）的下代 Groq LPU AI 推理加速器芯片有望采用基于 TSV（硅通孔）工艺的 3D 堆叠 DRAM 方案。
当前一代的 Groq LPU 采用了大容量 SRAM 的缓存设计，提供快速解码输出。但 SRAM 有着容量较小、制程升级密度收益有限的问题，在大型模型表现和芯片迭代上存在固有局限。
IT之家注意到，同样采用大 SRAM 技术路线的 Cerebras 此前已预告未来的 CS-6 将 3D 垂直集成 DRAM。
相关阅读：
#Groq#NVIDIA#Groq LPX#Groq next-generation LPU#NVIDIA Vera Rubin#NVIDIA NVL72#Cerebras#Cerebras CS-6