Skip to content
Archive

Post

Back to deliverables

Q qinbafrank
qinbafrank
@qinbafrank

看到有说英伟达Rubin上的HBM要减配了,实际情况是什么?应该不是已经开始交付并逐步放量的标准Vera Rubin NVL72减配,而是英伟达计划27年下半年推出的rubin升级版Rubin Ultra版配置规格还没确定下来,多家机构认为相比英伟达之前在GTC上公布的Rubin Ultra激进配置来说,可能最终规格要降低。详细聊聊: 1、标准版Vera Rubin现在的进度 标准Vera Rubin(NVL72)已经开始交付,6月1日前后,Dell率先向CoreWeave交付首批基于Vera Rubin NVL72的系统,CoreWeave完成行业首次完整bring-up和验证。 到7月,已有数十家客户收到测试机架/初期出货,包括CoreWeave、Microsoft、OpenAI、Anthropic、SpaceX AI、Google Cloud、Oracle、Nebius等。部分已在客户数据中心运行。 到秋季会开始更大规模交付,整体进度比Blackwell更顺利,无架组装时间大幅缩短(约5分钟级)。之前老黄和英伟达官方明确否认标准Vera Rubin有重大延期,路线图保持完整。 2、原计划27年下半年推出的Rubin Ultra 版(相当于是标准版的升级)的进展 1)原GTC2026公布的激进配置(目标2027年) (一)计算Die:4个接近光罩极限尺寸(near-reticle-sized)的计算芯粒,集成在单一封装内(相对标准Rubin的2个Die翻倍)。 (二)HBM配置:16个HBM4E堆栈。内存容量:单封装约1 TB HBM4E(业界首款TB级AI加速器)。 (三)采用更先进的封装(CoWoS-L相关),配套新的Kyber机架(垂直托盘、默认液冷),可支持NVL144等高密度配置(单机架最多144个GPU封装),整体算力和带宽目标显著高于标准Vera Rubin。 原来计划配置,单封装性能大幅跃升,强调极致密度和内存容量,用于超大规模模型。 2)6月底 SemiAnalysis等报告称因TSMC CoWoS-L封装基板翘曲(warpage)、光罩尺寸限制、良率/制造执行难题,原四Die + 16 HBM4E设计被取消。TSMC下一代CoPoS(Chip-on-Panel-on-Substrate)量产预计要到2028年底-2029年,赶不上2027节点。 可能的调整是四Die方案取消,转向双Die设计(与标准Rubin相同)+ 8个HBM4E堆栈,单封装计算规模和内存带宽大致减半。 容量示例:约384 GB/GPU(例如8个12-Hi、每堆栈48 GB的HBM4E),仍高于标准Rubin的288 GB HBM4,但远低于原1 TB目标。 带宽方面,HBM4E单堆栈速率更高(可达更高pin rate,如16 Gbps级别),但总堆栈数减半,整体仍不及原计划。 这属于相对原计划的明显缩水,但通过机架级扩展(更多封装)可部分补偿系统级性能。 3)7月底 TrendForce最近报告认为Rubin Ultra的HBM规格仍未最终锁定,因供应紧张、价格上涨,NVIDIA优先保障出货量、I/O速度和整体规模。英伟达可能考虑更低规格选项,包括HBM4E 8hi(8层堆栈,容量更低)等。 TrendForce报告提到有四种可能的HBM配置,核心是容量与供应确定性之间的权衡。最终规格预计在2026年下半年验证后确定。英伟达的目标仍是2027年出货,但更务实地平衡性能、成本和可制造性。 总结 1)标准Vera Rubin已经开始量产交付、计划秋季更大规模交付,节奏和规格应该是保持不变。 2)而原计划27年下半年推出的Rubin Ultra 版其配置规格相比原今年GTC大会上宣布激进版(4 Die + 16 HBM4E ≈ 1 TB)可能要调整至双Die + 8 HBM4E ≈ 384GB左右(标准Rubin的288 GB HBM4),但现在还没最终确定具体规格。

@

关于SemiAnalysis说“英伟达KyberNVL144机架架构出现大规模延迟”这篇推文,可能很多人都不理Rubin,Rubin Ultra两种芯片型号,Oberon和Kyber这两种机架型号的区别是什么?需要关注的点: 1、Rubin是GPU架构(芯片层面),搭配Vera CPU,2026下半年开始量产出货(6月初台北电脑展已经展示量产机型了)使用Oberon机架(NVL72规模)。 2、Rubin Ultra是Rubin的升级加强版(更多内存、更高性能配置),计划在2027年推出,使用Kyber机架。Kyber机架是专门为Rubin Ultra设计的新一代高密度机架,目标是把单机架GPU数量从72提升到144,通过更先进的 NVLink 实现更大的“scale-up 域”(把更多 GPU 像一个大 GPU 一样紧密连接)。 Semi说的延迟就是这个Kyber机架因为PCB中板的可制造性可能出现延迟。 3、在英伟达的规划里,Rubin Ultra NVL144(Kyber机柜架构)本身就要到27年才会推出,量产可能也就到27年下半年去了。毕竟rubin的NVL 72今年才半年才能量产嘛。其实不用过多关注kyber的进度,其实一切按照达子的预期节奏推动,Kyber量产落地也要到明年下半年去了。 4、然后SemiAnalysis说一个原本计划在27年下半年才能量产落地的Kyber NVL144机架架构要落后延迟到28年。大家都再说英伟达要延迟了🤣 重点还是关注今年Rubin NV72(Oberon 机架)量产落地的节奏。

· 58K Views

26 Reposts 1 Quotes 145 Likes 88 Bookmarks
replies reposts likes
17 replies collected of 75 X reports
超级幸运星~ @luckybibiw1p ·

@qinbafrank Sacrificing flagship performance for mass production certainty may protect the stock price in the short term, but in the long term, it lowers the ceiling. The market buys NVDA for "infinite possibilities," not "manufacturability."

original · zh

@qinbafrank 牺牲旗舰性能换量产确定性,对股价短期是保护,长期是上限变低了。市场买NVDA买的是“无限可能”,不是“可制造性”。

2
milly.hl @milly_rock7 ·

@qinbafrank The standard version is on track, but the Ultra's downgrade is still undecided. TrendForce's four HBM solutions are still under evaluation; 384GB is just one possibility, not the final specification.

original · zh

@qinbafrank 标准版进度没问题,但Ultra减配这事其实还没定论 TrendForce那四种HBM方案还在评估,384GB只是其中一个可能,不是最终规格

1
Nibili @NiNibili ·

@qinbafrank It seems the short sellers want to drive the storage market down another notch.

original · zh

@qinbafrank 看来空头想把存储再砸一个台阶

3
SilkDAO @SilkDAO_RWA ·

@qinbafrank The specifications won't be locked until the end of the year. If customers place orders now, should they sign for 1TB or 384G?

original · zh

@qinbafrank 规格年底才锁,客户现在下单是按1TB签还是按384G签?

leige @leige2017 ·

@qinbafrank It's not that HBM demand is decreasing, but rather that the increase in single-card usage for Ultra may not be as significant as expected.

original · zh

@qinbafrank 不是HBM需求减配,而是Ultra单卡增量可能没预期那么猛。

BTCD @BTCD_BTC ·

@qinbafrank The 1TB version has been reduced to 384GB; the reduction is real, but the increased capacity is also real.

original · zh

@qinbafrank 1TB砍到384GB,缩水是真,放量也是真。

Hideinsea @hideinsea ·

@qinbafrank Semi's latest report says it's down to 192G x.com/hideinsea/status/2083899…

original · zh

@qinbafrank Semi最新的报告写降到192G https://x.com/hideinsea/status/2083899301357641897/photo/1

4 11
youdaddy @dongmengbayue ·

@qinbafrank They're all ghost stories.

original · zh

@qinbafrank 都是鬼故事

@qinbafrank 很明顯儲存行情還沒結束

Brucewh8 (🫰~ 🐈🦜🦜) @brucewh8 ·

@qinbafrank If the computing power demand remains unchanged, reducing the configuration of a single rack will increase the number of racks required, but it will not have a significant impact on the supply chain.

original · zh

@qinbafrank 如果算力需求不变的话,单机柜减配会使得机柜需求数量上升,其实对于供应链并没有太大影响。

Ck Zria @ZriaCk ·

@qinbafrank Rubin lite

rock @vincent_wong_ ·

@qinbafrank Whether or not to reduce specifications depends on HBM's production capacity in the next few years; the memory/logic ratio is growing too fast.

original · zh

@qinbafrank 减配与否取决于hbm产量 后几年是真不够用 memory/logic ratio增长太快了

JIN 金晖 @huijin253999 ·

@qinbafrank Following this logic, storage is essentially nonexistent. The core HBM logic is shaken, and the most alarming factor is the 33% drop in demand. Furthermore, this logic takes effect in the second half of 2027; the specification downgrade means that the new version is virtually indistinguishable from the old one, which will also impact Nvidia. If this trend continues, hardware will continue to suffer.

original · zh

@qinbafrank 按这个逻辑来的话,存储就无了。最核心的 hbm 逻辑都动摇了,而且是最可怕的需求下滑(33%), 并且 这个逻辑是 27 年下半年生效,规格下降导致达子和旧版没什么区别,这也会对英伟达造成影响,按这样发酵,硬件还得继续跪🧎‍♂️

4 1
奶牛叔 @WWTLitee · 68K

@qinbafrank Rubin Ultra's reduced specifications are quite significant.

original · zh

@qinbafrank Rubin Ultra减配规格落差有点大

MoonJune @mujun5566 ·

@qinbafrank I don't like long articles written with AI.

original · zh

@qinbafrank 不喜欢用ai写的长篇大论

유정 @PQTonX · 34K

@qinbafrank The downgrade is a compromise based on reality, but the standard version is sufficient.

original · zh

@qinbafrank 减配是现实妥协,但标准版已够用

3 replies whose parent comment X withheld

Hideinsea @hideinsea ·

@qinbafrank If this logic holds true, is it possible that Rubin will also have to reduce the HBM component later on, because the price inversion would make Ultra unsellable?

original · zh

@qinbafrank 如果是这个逻辑 那有没有可能Rubin后面也要减配hbm 因为倒挂会导致Ultra卖不出去

1 1
qinbafrank @qinbafrank · 151K

@hideinsea vera Rubin has already started shipping, so it's a bit unreasonable to reduce the features any further. Just keep selling the standard version and wait for the next generation Feynman.

original · zh

@hideinsea vera rubin都已经开始交货了、再减配时不时有点说不过去呐。那就继续卖标准版就好了、等再下一代feynman了。

3
Xiaobaok @xiaob123 ·

@qinbafrank @hideinsea I think the number of GPUs and CPUs needs to be much higher, and there should be different types of HBM configurations.

original · zh

@qinbafrank @hideinsea 我倒觉得GPU CPU数量要上很多,而且HBM配置应该有不同种

These were collected in full; the comment they answer was not returned by X.