机械硬盘知识库 · 买之前先弄懂这几件事

两种记录方式的本质区别

CMR（常规磁记录）的磁道彼此独立、互不重叠，写入任何一条磁道都不影响邻居。SMR（叠瓦式磁记录）把磁道像屋顶瓦片一样部分重叠，提高了面密度、降低了成本，但代价是：要改写一条磁道，就得连带重写它后面被压住的那几条。

为什么在阵列里会出事

SMR 盘内部有一段「介质缓存」来吸收突发写入，平时看不出来。但一旦持续覆盖写超过缓存容量（重建整列、批量同步、备份校验都会触发），盘就必须边搬边写，性能断崖式下跌，严重时会被 RAID 控制器判定为超时并踢出阵列。这不是盘坏了，是特性使然。

怎么快速分辨

第一看型号：企业级近线盘的 CMR 型号通常带 NM / NE / NT 等后缀，消费级 SMR 常见于某些「红盘」「蓝盘」细分级。第二看规格表里的「持续传输速率」，SMR 会写得含糊或用顺序读速代替。第三最可靠 —— 直接问供货方要这一批的 CMR/SMR 属性，并写进报价单。

我们的做法

看板上的企业级型号均为 CMR。我们不做「同容量模糊替换」，如果某批拆机盘是 SMR，会在报价单上单独标注并说明适用场景（冷归档、单盘外置、顺序写入为主），不会混进阵列用盘里。

The core difference

In CMR (conventional magnetic recording) tracks are independent and do not overlap, so writing one track never disturbs its neighbour. SMR (shingled magnetic recording) partially overlaps tracks like roof tiles, raising areal density and lowering cost — but to rewrite one track you must also rewrite the shingled tracks sitting on top of it.

Why it breaks in arrays

An SMR drive contains a media cache that absorbs burst writes, so short workloads look normal. But once sustained overwrite exceeds that cache — a full array rebuild, a bulk resync, a backup verification — the drive has to move and write data simultaneously. Performance cliffs, and in bad cases the RAID controller times the drive out and drops it from the array. The drive is not faulty; the behaviour is by design.

How to tell quickly

Start with the model number: enterprise nearline CMR parts usually carry NM / NE / NT suffixes, while SMR shows up inside certain consumer "Red"/"Blue" sub-tiers. Then read the sustained transfer rate in the datasheet — SMR listings tend to be vague or quote sequential read instead. The only fully reliable route is to ask the supplier for the CMR/SMR attribute for the actual batch and have it written onto the quotation.

What we do

Every enterprise model on our board is CMR. We do not do fuzzy same-capacity substitution: if a batch of datacenter pulls is SMR, it is flagged separately on the quotation with its intended use case (cold archive, single external enclosure, sequential-write workloads) and never mixed into array-grade stock.

先弄清楚编号规则

SMART 属性有两种写法：十进制（5、197、198、199）和十六进制（05、C5、C6、C7）。它们是同一组东西。CrystalDiskInfo 一类的工具默认显示十六进制，所以行业里习惯说「05 / C5 / C6 / C7」，换个工具就变成 5 / 197 / 198 / 199。

05 重分配扇区计数（Reallocated Sector Count）

盘发现某个扇区读写异常，就把它的数据搬到备用的保留区，并把原扇区标记为坏。这个数字就是「已经搬走了多少个」。关键不是值大不大，而是「增量」：一个长期稳定在 8 的盘通常没事，一个上个月 0、这个月 40 的盘在快速劣化。

C5 当前待映射扇区计数（Current Pending Sector）

读写时出错、但还没能成功重映射的扇区数量。这是最该警惕的一项 —— 它代表「已知有问题、还没处理掉」的扇区。理想值是 0。持续大于 0 且不下降，说明盘面在恶化。

C6 无法校正的扇区计数（Offline Uncorrectable）

离机自检时无法纠正的扇区。与 C5 不同的是它已经确认不可修复。任何非 0 值都应当视为拒收理由之一。

C7 接口 CRC 错误计数（UDMA CRC Error Count）

这条不看盘面，看链路：数据在盘与主机之间传输出错（接触不良、线材劣质、背板供电不稳）会累加它。它本身不代表盘坏，但如果伴随读写速率异常或掉盘，就要先查线材、背板与供电，再怀疑盘。反过来说，一块只有 C7 偏高、其余全绿的盘，往往是可以用的。

看报告的正确姿势

把这四个值和「原始值（RAW）」一起看，不要只看「当前值 / 最差值 / 阈值」那一列 —— 归化值在多数厂商那里一旦出问题就直接钉死在阈值之上，看不出严重程度。另外务必要求看序列号级的原始日志，而不是一张截图汇总。

Sort out the numbering first

SMART attributes are written two ways: decimal (5, 197, 198, 199) and hexadecimal (05, C5, C6, C7). They are the same set. Tools like CrystalDiskInfo default to hex, which is why the industry says "05 / C5 / C6 / C7" while other tools show 5 / 197 / 198 / 199.

05 Reallocated Sector Count

When the drive finds a sector it cannot read or write reliably, it moves the data to spare reserve area and marks the original bad. This number counts how many have been moved. What matters is not the absolute figure but the delta: a drive stable at 8 for a year is normally fine, while a drive at 0 last month and 40 this month is degrading fast.

C5 Current Pending Sector

Sectors that errored on access but have not yet been successfully remapped — pending work the drive knows about but has not resolved. Ideal value is 0. A persistently non-zero value that does not fall means the media is deteriorating.

C6 Offline Uncorrectable

Sectors that failed correction during an offline self-scan. Unlike C5, these are confirmed unrecoverable. Any non-zero value should be treated as grounds for rejection.

C7 UDMA CRC Error Count

This one is not about the media but the link: transmission errors between drive and host (poor contact, cheap cable, unstable backplane power) increment it. It does not by itself mean a bad drive, but if it comes with abnormal throughput or dropouts, check cable, backplane and power before blaming the drive. Conversely, a drive whose only anomaly is a high C7 with everything else green is usually serviceable.

How to read the report properly

Read these four alongside their raw values, and do not rely on the normalised "current / worst / threshold" columns alone — once a problem appears, most vendors pin the normalised value just above threshold, hiding the severity. And always ask for the serial-level raw log rather than a summarised screenshot.

先换算成「年」再看

20,000 小时 ≈ 连续运行 2.3 年；35,000 小时 ≈ 4 年；43,800 小时就是整整 5 年不间断运行。企业级盘的设计工况本来就是 7×24 连续通电，所以两三年通电时长在数据中心里属于「刚过磨合」。

真正决定寿命的是工作负载量

比通电时长更重要的是每年写了多少 TB。企业级近线盘通常标 550TB/年，监控盘标 180–550TB/年，NAS 盘标 180TB/年，桌面盘只标 55TB/年。一块通电 40,000 小时但年均写入 80TB 的冷数据盘，实际压力可能比一块通电 15,000 小时、年均写入 500TB 的盘小得多。

失效曲线不是线性的

硬盘遵循典型的浴盆曲线：早期偶发批次缺陷，中期进入长期平缓的随机失效区（企业级盘宣称 AFR 约 0.35%，大规模实测常在 1–2%），最后才是磨损耗尽期。也就是说，通电 3 万小时的盘并不比通电 1 万小时的盘「危险三倍」，它更可能正处在最平稳的那一段。

我们的分级口径

我们把通电时长作为分级依据之一而不是唯一依据，且与 SMART 一起看：A 级 < 20,000 小时且 SMART 全绿无坏道，走生产环境；B 级 < 35,000 小时且 SMART 全绿，走监控、备份与冷数据；C 级读写正常但有历史记录，报价单上明确标注。分级依据会随货附在报告里，可以核对。

Convert to years first

20,000 hours is about 2.3 years of continuous operation; 35,000 hours about 4 years; 43,800 hours is exactly five years of uninterrupted running. Enterprise drives are designed for 7×24 duty, so two to three years of power-on time is barely past break-in in datacenter terms.

Workload is what really drives wear

More important than hours is terabytes written per year. Enterprise nearline drives are typically rated for 550TB/year, surveillance drives 180–550TB/year, NAS drives 180TB/year, and desktop drives only 55TB/year. A cold-data drive with 40,000 hours but 80TB/year written may be under far less stress than one with 15,000 hours and 500TB/year.

The failure curve is not linear

Hard drives follow the classic bathtub curve: occasional early batch defects, then a long flat region of random failure in the middle (vendors claim an AFR around 0.35% for enterprise parts; large-scale field data often shows 1–2%), and only finally the wear-out phase. A drive at 30,000 hours is therefore not "three times riskier" than one at 10,000 — it is more likely in the flattest part of the curve.

How we grade

We use power-on hours as one input to grading rather than the only one, and always read it together with SMART: Grade A is under 20,000 hours with all-green SMART and no bad sectors, for production environments; Grade B is under 35,000 hours with all-green SMART, for surveillance, backup and cold data; Grade C reads and writes normally but carries history and is flagged on the quotation. The grading basis ships with the report so you can check it.

企业级（Nearline / Enterprise）

面向 7×24 高负载与多盘阵列。7200 转、工作负载额定 550TB/年、MTBF 通常标 2.5M 小时、不可恢复读错误率（URE）1E-15、满盘振动补偿做得最足。它是阵列里最稳的一类，也是我们看板上的主力。

监控级（Surveillance）

固件为「多路顺序写、几乎不读」优化，支持 ATA 流命令，写入抖动小，能容忍写入中断。额定 180–550TB/年。它的短板是随机读写与小文件场景：拿它当系统盘或虚拟机盘会明显难受。

NAS 级

介于前两者之间，重点是多盘同箱的振动容忍与 RAID 友好的错误恢复策略（不会因为一个读错误就长时间卡死导致掉阵列）。额定通常 180TB/年。家用与小团队 NAS 足够，但大规模阵列仍建议上企业级。

桌面级

8×5 工况、额定 55TB/年、URE 常为 1E-14。它便宜，但不该被放进 7×24 的阵列 —— 错误恢复超时策略与振动补偿都不是为多盘环境设计的。

一句选购口诀

阵列与生产环境只用企业级；纯写入的监控归档可以走监控级；家用 NAS 用 NAS 级；桌面级只做单盘外置与冷备份，且永远不要用它去重建大容量阵列。

Enterprise (nearline)

Built for 7×24 high load in multi-drive arrays: 7200 RPM, 550TB/year workload rating, MTBF typically quoted at 2.5M hours, unrecoverable read error rate (URE) of 1E-15, and the strongest full-array vibration compensation. It is the most stable class in an array and the mainstay of our board.

Surveillance

Firmware tuned for many sequential write streams and almost no reads, with ATA streaming command support, low write jitter and tolerance for interrupted writes. Rated 180–550TB/year. Its weakness is random I/O and small files: using one as a system or VM disk is noticeably unpleasant.

NAS

Sits between the two, with an emphasis on vibration tolerance in multi-bay enclosures and RAID-friendly error recovery (so a single read error does not stall long enough to drop the array). Typically rated 180TB/year. Fine for home and small-team NAS; large arrays should still use enterprise drives.

Desktop

8×5 duty cycle, 55TB/year rating, URE usually 1E-14. It is cheap, but does not belong in a 7×24 array — neither its error-recovery timeout policy nor its vibration compensation is designed for multi-drive environments.

One-line rule

Enterprise only for arrays and production; surveillance for write-only surveillance archives; NAS-class for home NAS; desktop only for single external enclosures and cold backups — and never to rebuild a high-capacity array.

先算清楚重建要读多少

RAID5 重建一块盘，需要读取同组其余所有盘的全部数据才能通过校验算回丢失的内容。5 块 16TB 的阵列坏一块，重建期间要读 4×16 = 64TB。RAID6 同理，但能容忍两个错误。

再看 URE 的概率

URE（不可恢复读错误率）表示每读多少位会出现一个无法纠正的错误。企业级盘常见 1E-15，即每 10¹⁵ 位约一个错误；桌面级常见 1E-14。16TB = 1.28×10¹⁴ 位，换算下来：用 1E-15 的盘读满一块 16TB，撞错概率约 12%；用 1E-14 的盘，约 72%。乘上 64TB 的重建读取量，单靠 URE 一项，RAID5 的重建成功率就已经不乐观。

还有重建窗口本身的压力

重建是长时间满负荷顺序读，剩余几块盘在这段时间里承受的压力远高于日常。如果同批次盘有共同的早期缺陷（同厂同批次是常态），这段时间正好是它们最容易一起出事的时候 —— 行业里叫「第二块盘故障」。

怎么排架构

几条经验线：RAID5 单盘容量建议不超过 8TB，且阵列不超过 6 块；大容量阵列改用 RAID6 或 RAID10；重建前先做一次全盘校验读，把已有的坏扇区暴露出来；把重建速率限制打开，不要为了快把剩余盘压死；以及 —— 对冷数据与备份池，用认证再生企业级盘 + RAID6，成本比全上新盘低得多，风险并不更高。

Calculate the read volume first

Rebuilding one drive in a RAID 5 set requires reading all remaining members in full to reconstruct the lost content from parity. In a five-drive 16TB array, losing one drive means reading 4 × 16 = 64TB during the rebuild. RAID 6 works the same way but tolerates two failures.

Now apply the URE

URE (unrecoverable read error rate) expresses how many bits must be read before one uncorrectable error appears. Enterprise drives commonly specify 1E-15 — roughly one error per 10¹⁵ bits — while desktop drives specify 1E-14. A 16TB drive is 1.28×10¹⁴ bits, so a full read carries about a 12% chance of hitting one on a 1E-15 drive, and about 72% on a 1E-14 drive. Multiply that by a 64TB rebuild and, on URE alone, RAID 5 rebuild success is already poor.

And the rebuild window itself

A rebuild is a long, saturated sequential read, and the surviving drives are under far more stress than in daily use. If drives from the same batch share an early-life defect — same vendor, same batch is the norm — this window is exactly when they are most likely to fail together, the industry's "second drive failure".

How to design around it

A few rules of thumb: keep RAID 5 member capacity at or below 8TB and set size at six drives or fewer; move high-capacity arrays to RAID 6 or RAID 10; run a full verification read before a rebuild so existing bad sectors surface early; enable rebuild rate limiting instead of pinning the surviving drives for speed; and for cold data and backup pools, certified-reuse enterprise drives plus RAID 6 costs far less than all-new while not being riskier.

两种格式分别是什么

512e（512 emulation）物理扇区是 4096 字节，但对外模拟成 512 字节，兼容性最好。4Kn（4K native）物理与逻辑扇区都是 4096 字节，效率更高、校验更强，但需要操作系统、阵列卡与引导固件都支持。

典型的不兼容场景

老阵列卡与老版本固件、Windows Server 2008 及更早、部分老款引导模式（Legacy BIOS + MBR）、以及一些老 NAS 机型，遇到 4Kn 会直接不识别或无法作为启动盘。反过来，新平台两种都吃，只是 4Kn 的校验开销更小。

怎么查、怎么办

Linux 下看 lsblk -o NAME,LOG-SEC,PHY-SEC，逻辑扇区 4096 就是 4Kn；Windows 下用 fsutil fsinfo ntfsinfo 或阵列卡管理工具看。如果平台只吃 512e，而手头是 4Kn，可以在阵列卡层面做扇区大小转换（部分卡支持），或者换 512e 版本 —— 这也是我们发货前会先问平台型号的原因。

发货前的默认动作

下单时把服务器或阵列卡型号发来，我们会核对扇区格式、托架规格、固件要求与是否需要原厂托架，这三项是「装不上」问题里占绝大多数的原因。

What the two formats are

512e (512 emulation) has a physical sector of 4096 bytes but presents itself as 512 bytes, maximising compatibility. 4Kn (4K native) uses 4096-byte physical and logical sectors — more efficient with stronger integrity checking, but it requires the operating system, RAID card and boot firmware to support it.

Where it typically breaks

Older RAID cards and firmware, Windows Server 2008 and earlier, some legacy boot configurations (Legacy BIOS with MBR) and a number of older NAS models will simply not recognise a 4Kn drive or cannot boot from it. Modern platforms accept both, with 4Kn carrying lower integrity overhead.

How to check and what to do

On Linux, run lsblk -o NAME,LOG-SEC,PHY-SEC — a 4096 logical sector means 4Kn. On Windows, use fsutil fsinfo ntfsinfo or the RAID card utility. If the platform only accepts 512e and the drives are 4Kn, some controllers can translate sector size at the card level, otherwise switch to a 512e part — which is exactly why we ask for the platform model before shipping.

What we do before shipping

Send us the server or RAID card model when you order and we check sector format, bay/tray specification, firmware requirements and whether OEM carriers are needed. Those three items account for the overwhelming majority of "it does not fit" cases.

Clear / Purge / Destroy 是三个不同强度

NIST SP 800-88 Rev.1 把手段分成三档：Clear 用逻辑覆盖防止常规恢复，Purge 使数据在实验室级手段下也无法恢复（ATA Secure Erase、加密擦除、块擦除等），Destroy 是物理销毁。企业盘再流通场景应当做到 Purge。

凭证上必须有这 11 项

我们随货出具的凭证固定包含：证书编号、客户名称/订单号、硬盘品牌、料号、序列号、容量、擦除方法、擦除工具与版本、擦除时间、结果核验、执行人签署。缺任意一项都可以拒收 —— 这是你自己可以掌握的验收标准。

为什么「验证」这一步不能省

执行和验证是两件事。ATA Secure Erase 在部分固件上会静默失败（返回成功但实际没执行），所以标准要求擦除后抽样读取校验。凭证上没有验证记录的，本质上只是一份声明，不是证据。

和 IEEE 2883-2022 的关系

IEEE 2883 是较新的介质净化标准，把 NIST 的分类方式做了更细的工程化定义，并把固态与机械盘的差异写得更清楚。实践中两个标准可以互为参照；我们在报告上同时标注，是因为不同国家的合规审计偏好不同。

你自己可以做的抽查

收到盘后，随机抽 1–2 块做扇区级读取抽样（十六进制查看是否有非随机数据），并核对凭证上的序列号与盘体标签是否一致。这一步花不了多少时间，但能把「凭证是真的」这件事坐实。

Clear, Purge and Destroy are three different strengths

NIST SP 800-88 Rev.1 groups methods into three levels: Clear uses logical overwrite to defeat ordinary recovery; Purge makes data unrecoverable even with laboratory techniques (ATA Secure Erase, cryptographic erase, block erase); Destroy is physical destruction. Drives returning to circulation should meet Purge.

Eleven fields a certificate must carry

Our certificates always contain: certificate number; customer name or order number; drive brand; part number; serial number; capacity; sanitisation method; tool and version used; time of sanitisation; result verification; and operator sign-off. If any one is missing you are entitled to reject the shipment — that is a standard you control yourself.

Why the verification step is not optional

Executing and verifying are separate acts. ATA Secure Erase silently fails on some firmware (it reports success without performing the erase), which is why the standard requires a sampling read-back afterwards. A certificate with no verification record is a declaration, not evidence.

How it relates to IEEE 2883-2022

IEEE 2883 is a newer media-sanitisation standard that gives the NIST taxonomy more precise engineering definitions and treats flash and rotating media differences more explicitly. In practice the two standards cross-reference each other; we cite both because different countries' compliance audits prefer different ones.

A spot check you can run yourself

On receipt, pick one or two drives at random and sample-read sectors (check in hex whether the content is non-random), then cross-check the serial numbers on the certificate against the labels on the drives. It costs very little time and turns "the certificate is real" into something you have actually confirmed.

第一件：服务器 / 阵列卡型号

例如 Dell PowerEdge R740、HPE ProLiant DL380 Gen10、Supermicro CSE-826。厂商白名单（HPE 与 Dell 的固件校验尤其严格）、支持的扇区格式、最大单盘容量都与它直接相关。

第二件：托架类型

同一个型号，配不配原厂托架、是 2.5 吋转 3.5 吋还是原生 3.5 吋、螺丝孔位是否匹配，决定了它能不能插进你的盘位。HPE 与 Dell 的原厂托架与原厂贴牌盘是两件事，报价单上必须分开写。

第三件：接口与协议

SATA 6Gb/s 还是 SAS 12Gb/s、是单端口还是双端口。注意一条规则：SATA 盘可以插在 SAS 背板上用，SAS 盘不能插在纯 SATA 控制上。背板是 SAS 还是 SATA，直接决定你能不能上 SAS 盘。

顺带一提：白牌 / 彩标 / 官换盘

白牌（无厂标贴纸）与彩标（原厂彩色标签）在同一型号上常常价差几百元，差别在溯源与保修凭证而非盘本身；官换盘（R 盘）是原厂换新流出，通电时长通常极短。三种我们都会在报价单上分开标注，让你按需求选，而不是替你决定。

One: server or RAID card model

For example Dell PowerEdge R740, HPE ProLiant DL380 Gen10, Supermicro CSE-826. Vendor whitelists (HPE and Dell firmware checking is especially strict), supported sector formats and maximum per-drive capacity all depend directly on this.

Two: carrier/tray type

For the same drive, whether an OEM carrier is included, whether it is a 2.5-to-3.5 adapter or native 3.5 inch, and whether the screw positions match decide whether it slots into your bay. OEM carriers and OEM-labelled drives are two different things and must be quoted separately.

Three: interface and protocol

SATA 6Gb/s or SAS 12Gb/s, single-port or dual-port. One rule to remember: a SATA drive works on a SAS backplane, but a SAS drive will not work on a SATA-only controller. Whether your backplane is SAS or SATA directly determines whether you can use SAS drives.

Related: white label, colour label and OEM replacement drives

White-label (no vendor sticker) and colour-label (original coloured label) versions of the same model often differ by several hundred RMB, but the difference lies in traceability and warranty evidence rather than the drive itself. Replacement drives (R units) come from vendor exchange programmes and usually show very short power-on time. We label all three separately on the quotation so you choose by requirement rather than us deciding for you.

1. 先核对外箱与序列号范围

外箱标注的序列号范围要与装箱单、擦除凭证一一对上。有出入先拍照留证再开箱，这一步是后续任何索赔的前提。

2. 逐块核对盘体标签与凭证

品牌、料号、序列号、容量四项对齐。同时记下制造年份 —— 如果是发往马来西亚的货，制造日期必须在 5 年以内，这条是硬线。

3. 上电前先做外观与接口检查

看电路板有无水渍、氧化、烧灼痕迹，SATA/SAS 接口针脚有无变形。外观异常直接单独隔离，不要上电，避免扩大损失。

4. 抽 1–2 块做扇区级读取抽查

用十六进制读取前若干 MB，确认不是可识别数据；这既验证擦除凭证，也验证盘能被正常访问。

5. 全批跑一遍 SMART 与短自检

记录 05 / C5 / C6 / C7 的原始值与通电时长，和我们随货报告比对。不一致的单独列出，作为沟通依据。

6. 抽 1–2 块做 24 小时烤机

连续读写 + 温度监控，观察 SMART 是否出现增量。DOA 里相当一部分是「上电即现」的，抽检就能提前筛掉。

做完这六步再批量上架

这六步加起来通常不到一天。批量上架之后才发现问题，责任界定会困难得多 —— 到底是我们发的盘有问题，还是你的机柜供电或背板把盘弄坏了？把证据做在前面，对双方都省事。

1. Check the carton against the serial range

The serial range printed on the outer carton must match the packing list and the sanitisation certificate line by line. If anything is off, photograph it before opening — this step underpins any later claim.

2. Match every drive label against the certificate

Brand, part number, serial number and capacity must align. Also record the year of manufacture: for shipments into Malaysia it must be within five years, and that is a hard line.

3. Inspect the exterior and connectors before applying power

Look for water marks, oxidation or burn traces on the PCB, and for bent pins on the SATA/SAS connector. Isolate anything abnormal rather than powering it up and risking wider damage.

4. Sample-read one or two drives at sector level

Read the first few megabytes in hex to confirm the content is not recognisable data. This verifies both the sanitisation certificate and that the drive is accessible.

5. Run SMART and a short self-test across the batch

Record the raw values of 05 / C5 / C6 / C7 and the power-on hours, and compare them against our accompanying report. List any mismatch separately as the basis for discussion.

6. Burn in one or two drives for 24 hours

Continuous read/write with temperature monitoring, watching for SMART increments. A good share of DOA failures appear immediately on power-up, so sampling filters them out early.

Only then rack the batch

These six steps usually take under a day in total. Finding a problem after the whole batch is in service makes responsibility far harder to establish — was the drive faulty when we shipped it, or did your rack power or backplane damage it? Building the evidence up front saves both sides trouble.

为什么用氦气

盘片高速旋转时，空气会在盘片间产生湍流与阻力。换成密度只有空气约七分之一的氦气，湍流和阻力都显著下降，于是可以在同样厚度里叠进更多盘片（这就是 14TB 以上普遍采用氦封的原因），同时功耗和运行噪音也更低。

密封是怎么做的

现代氦封盘用激光焊接的金属外壳，而非传统的橡胶垫圈。厂商标称的泄漏率通常保证 5 年内氦气浓度仍在可工作范围。

怎么看它老没老

部分厂商提供「氦气液位」相关的 SMART 属性（不同厂商编号不同，WD/HGST 上常见 22，Seagate 上另有定义）。更通用的判断方式是看温度与功耗：氦气流失后盘内阻力上升，同工况下温度会缓慢上升、功耗变高，噪音也会变大。所以一盘温明显高于同批次其他盘的氦封盘，值得单独留意。

对二手采购的影响

氦封盘的通电时长与温度历史比空气盘更值得关注。我们的做法是：氦封型号在检测时额外记录温度曲线，同一批次里温度异常的单独分级，不进 A 级。

Why helium

At high rotational speed, air creates turbulence and drag between platters. Replacing it with helium — about one-seventh the density of air — cuts both sharply, so more platters fit into the same height (which is why 14TB and above are generally helium-sealed), with lower power draw and less acoustic noise.

How the seal is made

Modern helium drives use a laser-welded metal enclosure rather than a conventional rubber gasket. Vendors typically guarantee the leak rate keeps helium concentration within the working range for five years.

How to tell if it is aging

Some vendors expose a helium-level SMART attribute (numbering differs by vendor; commonly 22 on WD/HGST, defined differently on Seagate). A more universal signal is temperature and power: as helium escapes, internal drag rises, so under identical conditions temperature creeps up, power draw increases and noise grows. A helium-sealed drive running noticeably hotter than others in its batch deserves separate attention.

What it means for used purchasing

On helium drives, power-on hours and temperature history matter more than on air-filled models. Our approach: helium models get an extra temperature curve recorded during testing, and any unit running hot relative to its batch is graded separately rather than entering Grade A.

产品属性（5 项）

① 确切料号，不接受「同容量代替」；② 制造年份，且能写进合同；③ CMR 还是 SMR；④ 通电时长与工作负载估计（年均 TB 写入）；⑤ 分级口径（A/B/C 的判定标准）。

检测证据（3 项）

⑥ 序列号级 SMART 原始日志，含 05 / C5 / C6 / C7 与通电时长，能提供最近一次全盘扫描的坏道结果；⑦ 满足 NIST SP 800-88 Rev.1 Purge 的擦除凭证，含验证步骤；⑧ 装箱单，列明品牌、型号、序列号、制造年份与检验日期。

交易条款（2 项）

⑨ 质保年限与 DOA 条款写清楚，包括谁承担来回运费；⑩ 贸易术语（DDP 还是 FOB）与随货合规文件清单，尤其是发往东南亚时 —— 目的国是否允许进口「旧电器电子设备」是硬门槛。

为什么这十项能筛人

真做检测的供货方随时能拿出 SMART 原始日志与擦除凭证，因为这些东西本来就是他们流程的产出；转手方通常只能给一张截图或一句「都测过了」。这十项一问，答案的质量差别会非常明显。

Product attributes (5)

(1) The exact part number — "same capacity substitute" is not acceptable. (2) Year of manufacture, stated contractually. (3) CMR or SMR. (4) Power-on hours plus an estimate of workload (annual TB written). (5) The grading basis, i.e. how A/B/C are decided.

Test evidence (3)

(6) Serial-level raw SMART log covering 05 / C5 / C6 / C7 and power-on hours, plus bad-sector results from the most recent full-surface scan. (7) A sanitiser certificate meeting NIST SP 800-88 Rev.1 Purge, including the verification step. (8) A packing list stating brand, model, serial number, year of manufacture and test date.

Commercial terms (2)

(9) Warranty period and DOA terms in writing, including who pays return freight. (10) Trade terms (DDP or FOB) and the shipping-document checklist — especially for Southeast Asia, where whether the destination permits import of "used electrical and electronic equipment" is a hard gate.

Why this list filters suppliers

A supplier who genuinely tests can produce raw SMART logs and sanitiser certificates on demand, because those are outputs of their own process. A reseller usually has a screenshot or a sentence: "everything was tested". Ask these ten and the quality gap is obvious.