← 返回列表

@zephyr_z9: "HBF can only target a narrow set of workloads, like read-heavy, high-capacity data that benefits from being close to the GPU." HBF can be ...

@zephyr_z9 3 信息等级 3 1 噪音/剔除；2 较弱；3 普通事实；4 重要行业动态；5 极重大事件。该分数是信息显著性，不是投资建议。 发布：2026-05-24T05:48 抓取：2026-05-24 12:57
🔗 原文链接
AI 数据中心 算力
摘要

推特讨论HBF技术主要适用于读取密集型高容量数据，并指出随着模型规模增长，HBF可能有用。同时提到英伟达当前策略是扩大scale-up域规模至144/576/1152，将权重存储在大域中，并通过STX将KV缓存卸载到SSD。

客观事实
  • HBF主要针对窄范围工作负载，如读取密集型高容量数据
  • 英伟达策略是扩大scale-up域至144/576/1152并存储权重
  • 英伟达通过STX将KV缓存卸载到SSD
英伟达

原文

"HBF can only target a narrow set of workloads, like read-heavy, high-capacity data that benefits from being close to the GPU."

HBF can be useful if model sizes grow to 50T-100T parameters
But as far as I can tell, Nvidia's current strategy is to increase the size of the scale-up domain to 144/576/1152 to store the weights in a big single domain and offload the KV cache to SSDs thru STX

likes: 144 | retweets: 13 | replies: 12 | views: 46871