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
Blackwell GPU
8 atoms · 跨 5 天 · 首见 2026-05-08 · 最近 2026-06-05
三色: 🟦 fact 5 · 🟥 take 3 · stance ▲2/▼0/◆1
来源: X 8
时态: fresh:7 · aging:1
标签: 好数字:5 · 好思考:3 · 好观点:1 · 好信源:1
叙事 (narrative) (2)
- 2026-06-05 · rental price assumption reset
Per gpu and assumes 100% utilization.
tags: 好思考 · → daily
value: direction=per gpu and assumes 100% utilization
- 2026-05-08
[老化中] · software maturity timeframe
Software needs like a year to mature, buying VR now means a lot of work. Good time for Blackwells
tags: 好观点 · → daily
value: qty=1 year
预测 (forecast) (2)
- 2026-05-26 · output token capacity
Using 8k:1k input-to-output token requests, we estimate it would be possible to serve ~20B output tok/s
tags: 好数字·好思考 · → daily
value: qty=~20B output tok/s · date=current
- 2026-05-26 · demand exceed supply at long context
global throughput demand could reach 200M to 4B tok/s — likely beyond what today's Blackwell fleet could serve at long context lengths
tags: 好思考 · → daily
事实 (fact) (4)
- 2026-05-19 · thermal design power per chip
The Blackwell GPU generation operates at thermal design power exceeding 1,000 watts per chip
tags: 好数字·好信源 · → daily
value: qty=exceeds 1,000 watts · direction=> 1000W
- 2026-05-26 · prefill phase compute-bound
For 8k:1k requests, we find that the prefill phase is compute-bound
tags: 好数字 · → daily
- 2026-05-26 · decoding phase bandwidth-bound
For 8k:1k requests, we find that... while decoding is bandwidth-bound
tags: 好数字 · → daily
- 2026-05-20 · 算力规模
A year ago, it was 480T.
tags: 好数字 · → daily
value: qty=480T · date=1年前
时间轴 (近 20)
- 2026-06-05 ·
narrative · rental price assumption reset · → daily
- 2026-05-26 ·
forecast · output token capacity · → daily
- 2026-05-26 ·
fact · prefill phase compute-bound · → daily
- 2026-05-26 ·
fact · decoding phase bandwidth-bound · → daily
- 2026-05-26 ·
forecast · demand exceed supply at long context · → daily
- 2026-05-20 ·
fact · 算力规模 · → daily
- 2026-05-19 ·
fact · thermal design power per chip · → daily
- 2026-05-08 ·
narrative · software maturity timeframe · → daily
← 实体目录 · 系统日志