七小服 Seven Small Services @MayaGPU

Data centre maintenance & spare parts · Servers, storage, network, GPU & AI compute · 80+ cities, 35+ countries · 全球数据中心维保与备件 sevensmallservices.com Hong Kong SAR Joined October 2022

Tweets

297
Followers

46
Following

585
Likes

268

七小服 Seven Small Services @MayaGPU

11 hours ago

IBM z17 vs a typical server: · Processor units installed ahead, activated as needed, upgraded concurrently · RAIM memory recovers from DIMM failures without business impact · Chip-level water loops, system air cooling: no facility water Machine Teardown #11. Per IBM Redbooks.

0 0 0 10 0

View Details

七小服 Seven Small Services @MayaGPU

11 hours ago

IBM z17 大型机和普通服务器哪里不一样？ · 处理器多装少开，按需激活，不停机升级 · RAIM 内存：坏了自动恢复，不影响业务 · 芯片水冷、整机风冷，机房不用接冷却水机型拆解第 11 期，资料来自 IBM 官方红皮书。

0 0 0 5 0

View Details

七小服 Seven Small Services @MayaGPU

20 hours ago

How long does an SSD last? Sandisk: 5 to 10 years under normal use, often longer. It depends on how much you write. · TBW: total terabytes the drive can write · DWPD: full drive writes per day; the data center metric · Life (years) = TBW ÷ (TB written per day × 365). 600 TBW at 20 GB a day is ~82 years, in theory · Update firmware, enable TRIM, leave 10–20% free Our view: PCs get replaced before the flash wears out. AI data centers write far more, so track SMART wear and replace early. Seven Small Services: post-warranty maintenance and spare parts for servers and storage, 24×7 with 4-hour response. Per Sandisk. Industry explainer, not investment advice.

0 0 0 7 0

View Details

七小服 Seven Small Services @MayaGPU

20 hours ago

一块 SSD 能用多久？闪迪：正常使用一般 5 到 10 年，很多更久。寿命主要看写了多少数据。 · TBW：一共能写多少 TB · DWPD：保修期内每天能写满几遍，数据中心看这个 · 自己算：寿命（年）= TBW ÷（每天写入 TB × 365）。600TBW、每天 20GB，理论约 82 年 · 建议：更新固件、开 TRIM、留 10%–20% 空间我们的看法：家用电脑一般等不到写坏就换机了；AI 机房存检查点、缓存，写入量大得多，更要盯 SMART 剩余寿命，快到线的提前换。七小服做服务器和存储过保维保与备件，7×24×4 响应。能修的修，该换的换。据闪迪官网。行业知识分享，不构成投资建议。

0 0 0 61 0

View Details

七小服 Seven Small Services @MayaGPU

20 hours ago

Why does Google's TPU rely on Broadcom? Google sets the goals, Broadcom builds the chip, TSMC makes it. · Partners since TPU v1 in 2015; the 8th gen splits into 8t (training) and 8i (inference) · Google: what it computes, how chips connect, software. Broadcom: high-speed I/O, physical design, packaging · Broadcom's April SEC filing: supply assurance to 2031, incl. networking for next-gen racks · No longer exclusive: MediaTek works on 8i; Marvell on TPU-adjacent chips Our view: custom silicon is a supply chain, not one company. Seven Small Services: post-warranty maintenance for servers and GPUs, 24×7 with 4-hour response. Per filings and reports. Industry explainer, not investment advice.

0 0 0 31 0

View Details

七小服 Seven Small Services @MayaGPU

20 hours ago

谷歌的 TPU，为什么离不开博通？谷歌出想法，博通出手艺，台积电来生产。 · 2015 年第一代 TPU 起合作，11 年八代，今年第八代第一次分成训练 8t 和推理 8i · 分工：谷歌定算什么、怎么连、配什么软件；博通做高速接口、物理设计、封装 · 博通 4 月 SEC 文件：供应保障协议签到 2031 年，含下一代 AI 机柜的网络芯片 · 不再独家：8i 由联发科参与，Marvell 也做 TPU 周边芯片我们的看法：「自研」从来不是一家公司单干。芯片之外，TPU 机房里的服务器、网络和存储一样要维保。七小服做服务器和 GPU 过保维保，7×24×4 响应。据博通文件和公开报道。行业知识分享，不构成投资建议。

2 0 2 132 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

When does an AI factory pay off? NVIDIA's Oct 1 blog gives three tests: productive, durable, fungible. · Cost: about $60M per MW, so ~$60B for 1 GW (our math) · Productive: Vera Rubin NVL72 delivers 30x+ throughput per MW vs. GB300 (SemiAnalysis, cited by NVIDIA) · Durable: six-year-old A100s still worth about a quarter of list; Microsoft's V100s ran 8.4 years vs. 6 on the books · Fungible: training, inference, data processing, simulation Our view: a useful framework, but discount the numbers; NVIDIA sells the GPUs. Long service lives come from steady maintenance. Seven Small Services: post-warranty maintenance for servers and GPUs, 24×7 with 4-hour response. Per NVIDIA. Industry explainer, not investment advice.

0 0 0 25 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

一座 AI 工厂，怎么才算赚钱？英伟达 10 月 1 日发文，给了三个标准：能产、耐用、通用。 · 成本：每兆瓦约 6000 万美元，1 吉瓦约 600 亿（我们按单价推算） · 能产：英伟达引用的数据，Vera Rubin NVL72 每兆瓦吞吐是上一代的 30 倍以上 · 耐用：A100 用了六年还值原价约四分之一；微软的 V100 实际跑了 8.4 年，账面折旧按 6 年 · 通用：训练、推理、数据处理、仿真都能跑我们的看法：三个标准可以当看 AI 机房的框架，数字要打折扣，毕竟是卖 GPU 的人写的。机器能用这么久，背后是一直有人修、有备件。七小服做服务器和 GPU 过保维保，7×24×4 响应。能修的修，该换的换。据英伟达官方博客。行业知识分享，不构成投资建议。

0 0 0 59 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

@AYi_AInotes 👍

0 0 0 95 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

@firesidealpha 👍

0 0 0 907 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

@firesidealpha 👍

0 0 0 395 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

@elonmusk 👍

0 0 1 17 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

@elonmusk 👍

0 0 2 13 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

@Arm @Supermicro @OpenComputePrj 👍

0 0 0 19 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

@MarvellTech 👍

0 0 0 4 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

Why is Microsoft betting on its own inference chip? On Oct 12, Microsoft said it will scale Maia to multi-gigawatt deployments from 2027, as the primary inference system for its products. · Maia 200 already runs Copilot and Teams AI features in the US, Europe and Australia · Maia 300: 8 HBM4 stacks, 288GB, 18TB/s, 950 W, compute on TSMC N2P · 96 chips and 167 kW per rack; up to 7x Maia 200 throughput · Expected 40% more tokens per MW than GPU systems; volume in 2027 Our view: inference is a tokens-per-watt game. Caveat: the 40% is an expectation and the GPU baseline is not named. Seven Small Services: post-warranty maintenance for servers and GPUs, 24×7 with 4-hour response. Per Microsoft. Industry explainer, not investment advice.

0 0 0 19 0

View Details

七小服 Seven Small Services @MayaGPU

2 days ago

为什么微软要让自己的芯片扛起 AI 推理？ 10 月 12 日，微软宣布 2027 年起把自研 Maia 芯片扩到「多吉瓦」规模，成为自家产品和服务的主要推理系统。 · Maia 200 已在美国、欧洲、澳大利亚部署，跑 Copilot、Teams 里的 AI 功能 · Maia 300：8 摞 HBM4、288GB、18TB/s，950 瓦，计算芯粒用台积电 N2P · 一柜 96 颗、167 千瓦，微软说整柜吞吐最多是 Maia 200 的 7 倍 · 每兆瓦 token 预计比主流 GPU 系统多 40%，2027 年量产部署我们的看法：推理拼的是每度电出多少 token。要留意：40% 是「预计」，对比的 GPU 没点名。七小服做服务器和 GPU 过保维保，7×24×4 响应。能修的修，该换的换。据微软官方博客。行业知识分享，不构成投资建议。