SemiAnalysis报道称，Meta正在开发一款定制的AMD MI400系列芯片，该芯片尺寸是标准MI455X的一半，并针对推荐系统工作负载和内存带宽效率进行了优化。这款定制芯片使用的HBM（高带宽内存）比标准版本少得多，因此不太适合LLM推理和训练。该报道还批评了Meta更广泛的基础设施战略，提到了过去GB200 NVL72 Ariel出现的问题，并暗示定制芯片设计可能会阻碍Meta出租计算资源的能力，而这一策略的灵感来源于埃隆·马斯克的做法。 AI
影响 Meta的定制芯片设计可能会限制其LLM能力，从而影响其在AI开发方面的竞争力。
排序理由 该集群包含来自SemiAnalysis的多个推文，讨论Meta的定制芯片设计和基础设施战略，提供分析和批评，而不是主要公告。
AI 生成摘要 · Google Gemini · 来自 7 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
@elonmusk We think that the MI455X will be a great chip as long as @AnushElangovan invests enough in software and automated testing capabilities to fix AMD's long history of poor software quality, but Meta's overengineering leads to less flexibility in the overall strategy that A…
Another issue that Meta's custom half-size package design will cause is that it will be harder for Zuck to rent them out, following his strategy of copying @elonmusk's neocloud strategy. 6/7🧵
Another example of Meta's overengineering obsession is GB200 NVL72 Ariel, which caused massive infrastructure issues with its cross-rack NVLink ACC cables due to signal integrity problems, as Meta was obsessed with having the Grace CPU at a 1:1 ratio with the GPU. This decision h…
This is yet another instance of the recsys infrastructure strategy team at Meta wanting to look like they “add value” by doing weird micro-optimizations that hurt other important divisions at Meta. 4/7🧵
But the issue is that Meta’s custom MI400-series SKU is not as optimized for LLM inference and training. The decision was made before TBD Lab was formed or could have its say. Given the significant decreases in compute and HBM in its custom SKU, it will be less attractive to
Compared to a normal MI455X package, it will use six HBM4 8i stacks instead of 12 HBM4 12Hi stacks. The reasoning is that Meta’s recsys infrastructure strategy wanted to have a CPU compute-to-GPU compute ratio tuned for recsys and to optimize memory $/BW. 2/7🧵 https://t.co/jtfrdP…
ALERT🚨🚨: META's CUSTOM AMD MI400-series chip will be half the size of a normal MI455X chip. It is "optimized" for recsys workloads and $/Memory Bandwidth. It will use ~144GB of HBM instead of 432GB. We break it down below👇️ 1/7🧵 https://t.co/hiz53VYqkv