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
GB300 NVL72
22 atoms · 跨 10 天 · 首见 2026-04-17 · 最近 2026-07-02
三色: 🟦 fact 13 · 🟥 take 9 · stance ▲5/▼2/◆4
来源: X 22
时态: fresh:22
标签: 好数字:18 · 好观点:3 · 好思考:1
预测 (forecast) (9)
- 2026-06-16 · train 500B model on single rack in 90 days
you could probably train a 500B model on a single GB300 NVL72 rack in 90 days if you optimized everything
tags: 好数字 · → daily
- 2026-06-16 · 30% MFU training time one year assuming 20B active 30T tokens
assuming 20B active, 30T tokens, 6ND for flops, 360 pflop/s for a GB300 NVL72, and 30% mfu, this takes a whole year
tags: 好数字 · → daily
- 2026-06-16 · 60% MFU wrt bf16 peak still half a year training
even at 60% mfu wrt bf16 peak that's still half a year
tags: 好数字 · → daily
- 2026-06-16 · 2030 rack cost 5x of today
the rack will probably cost 5x of what it costs today but whatever
tags: 好数字 · → daily
- 2026-06-10 · total revenue at 75k racks
75,000 Racks GB300 NVL72 ~$450B
tags: 好数字 · → daily
value: qty=~$450B
- 2026-06-10 · total revenue at 85k racks
85,000 Racks GB300 NVL72 ~$510B
tags: 好数字 · → daily
value: qty=~$510B
- 2026-06-16 · 2030 single rack 5e25-1e26 flop model possible
in 2030 a 5e25-1e26 flop model should be possible on a single rack
tags: 好观点 · → daily
- 2026-06-16 · 200-300B model on single rack would be sick by 2030
_well then even if it's just a 200-300B model that would still be pretty sick on a single rack
and in 2030 a 1T model is probably possible on one rack_
tags: 好观点 · → daily
- 2026-06-16 · 2030 should have 100T models
if we don't have 100T models then something went wrong
tags: 好观点 · → daily
事实 (fact) (13)
- 2026-06-12 · 每GPU代币吞吐量
GB300 NVL72 最高吞吐量: 约 12.2k tok/s/gpu
tags: 好数字 · → daily
value: qty=~12.2k tok/s/gpu · date=2026-06-12
- 2026-06-12 · 最大交互性
GB300 NVL72 最大交互性: 约 245 tok/s/user
tags: 好数字 · → daily
value: qty=~245 tok/s/user · date=2026-06-12
- 2026-07-02 · 训练扩展能力
Turns out with a little Speculators+Mooncake, I'm able to scale training on GB300 NVL72
tags: 好思考 · → daily
- 2026-06-27 · performance relative to best H100 config
the most optimized GB300 NVL72 hits about 17x the best H100 config in FP8, 32x in FP4
tags: 好数字 · → daily
value: qty=17x in FP8, 32x in FP4
- 2026-06-22 · 每百万Token成本
GB300 NVL72 (Dynamo SGLang) 2026-05-29 成本: 0.21 美元/百万Token
tags: 好数字 · → daily
value: qty=$0.21 · date=2026-05-29
- 2026-06-22 · 每百万Token成本
GB300 NVL72 每百万 Token 成本: $0.42
tags: 好数字 · → daily
value: qty=$0.42 · date=2026-06-18
- 2026-06-12 · 每GPU吞吐量
🚀New record on GB300 NVL72: SGLang exceeds 12K tok/s per GPU on DeepSeek V4 Pro 1.6T (FP4, 8K/1K)
tags: 好数字 · → daily
value: qty=>12K tok/s/gpu · date=2026-06-12
- 2026-06-10 · revenue per rack
GB300 NVL72 ~$6M
tags: 好数字 · → daily
value: qty=~$6M
- 2026-06-09 · fine-tuning usage
i used a full gb300nvl72 for ft (took ~6 weeks) and now i am hosting it on the same cluster
tags: 好数字 · → daily
value: qty=1
- 2026-06-09 · fine-tuning duration
i used a full gb300nvl72 for ft (took ~6 weeks)
tags: 好数字 · → daily
value: qty=6 weeks
- 2026-05-31 · 单机架 PCB 成本
对象：GB300 NVL72；本周披露或引用数据：单机架 PCB 成本约 3.5 万美元
tags: 好数字 · → daily
value: qty=约3.5万美元 · date=本周
- 2026-05-17 · BBU配置
GB300 NVL72：配21700 3.0Ah、21700 4.0Ah，非全极耳电芯，约5000-6000颗/柜
tags: 好数字 · → daily
value: qty=5000-6000颗/柜
- 2026-04-17 · 聚合带宽
GB300 NVL72 聚合带宽: 130 TB/s
tags: 好数字 · → daily
value: qty=130 TB/s · date=2026-04-17
时间轴 (近 20)
- 2026-07-02 ·
fact · 训练扩展能力 · → daily
- 2026-06-27 ·
fact · performance relative to best H100 config · → daily
- 2026-06-22 ·
fact · 每百万Token成本 · → daily
- 2026-06-22 ·
fact · 每百万Token成本 · → daily
- 2026-06-16 ·
forecast · train 500B model on single rack in 90 days · → daily
- 2026-06-16 ·
forecast · 2030 single rack 5e25-1e26 flop model possible · → daily
- 2026-06-16 ·
forecast · 30% MFU training time one year assuming 20B active 30T token · → daily
- 2026-06-16 ·
forecast · 60% MFU wrt bf16 peak still half a year training · → daily
- 2026-06-16 ·
forecast · 200-300B model on single rack would be sick by 2030 · → daily
- 2026-06-16 ·
forecast · 2030 rack cost 5x of today · → daily
- 2026-06-16 ·
forecast · 2030 should have 100T models · → daily
- 2026-06-12 ·
fact · 每GPU吞吐量 · → daily
- 2026-06-12 ·
fact · 每GPU代币吞吐量 · → daily
- 2026-06-12 ·
fact · 最大交互性 · → daily
- 2026-06-10 ·
fact · revenue per rack · → daily
- 2026-06-10 ·
forecast · total revenue at 75k racks · → daily
- 2026-06-10 ·
forecast · total revenue at 85k racks · → daily
- 2026-06-09 ·
fact · fine-tuning usage · → daily
- 2026-06-09 ·
fact · fine-tuning duration · → daily
- 2026-05-31 ·
fact · 单机架 PCB 成本 · → daily
← 实体目录 · 系统日志