以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
Vera Rubin NVL8
7 atoms · 跨 2 天 · 首见 2026-05-31 · 最近 2026-06-01
三色: 🟦 fact 6 · 🟥 take 1 · stance ▲4/▼1/◆2
来源: X 7
时态: fresh:7
标签: 好数字:5 · 好思考:2
反驳 (refute) (2)
- 2026-05-31 · memory bandwidth vs single H100
48x memory bandwidth?!
tags: 好数字 · → daily
value: qty=48x
- 2026-05-31 · NVLink bandwidth vs single H100
31x NVlink bandwidth?!
tags: 好数字 · → daily
value: qty=31x
预测 (forecast) (1)
- 2026-05-31 · 真实加速性能远低于理论峰值
It'll take serious engineering to get speedups even close to 6.5x on any workload!
tags: 好思考 · → daily
事实 (fact) (4)
- 2026-06-01 · 整体系统内存带宽相对于单个H100的倍数
the 48x memory BW... claims seem to come from comparing the full NVL8 system vs. a single H100
tags: 好数字 · → daily
value: qty=48x
- 2026-06-01 · 整体NVLink带宽相对于单个H100的倍数
the 31x NVLink claims seem to come from comparing the full NVL8 system vs. a single H100
tags: 好数字 · → daily
value: qty=31x
- 2026-05-31 · memory capacity vs single H100
3.5x more memory
tags: 好数字 · → daily
value: qty=3.5x
- 2026-05-31 · inter-GPU intra-node memory read latency vs H100 local memory
it's actually faster to transfer and read memory from another GPU inside a node of Rubin NVL8 than it is to just read the memory inside an H100
tags: 好思考 · → daily
时间轴 (近 20)
- 2026-06-01 ·
fact · 整体系统内存带宽相对于单个H100的倍数 · → daily
- 2026-06-01 ·
fact · 整体NVLink带宽相对于单个H100的倍数 · → daily
- 2026-05-31 ·
refute · memory bandwidth vs single H100 · → daily
- 2026-05-31 ·
refute · NVLink bandwidth vs single H100 · → daily
- 2026-05-31 ·
fact · memory capacity vs single H100 · → daily
- 2026-05-31 ·
fact · inter-GPU intra-node memory read latency vs H100 local memor · → daily
- 2026-05-31 ·
forecast · 真实加速性能远低于理论峰值 · → daily
← 实体目录 · 系统日志