Sizing NAS Hardware: Bays, HBAs, ECC Memory and Power

A NAS build is decided by constraints, not by picking the fastest parts. The chassis fixes how many drives you can ever hold, the motherboard fixes how many PCIe lanes you can spend, and the power supply has to cope with a load profile that looks nothing like a desktop. Getting the order of those decisions right saves a rebuild later.

This piece is about building one. If you have not yet decided between a build and an off-the-shelf appliance, start with the NAS buying guide and come back once the answer is a build. To turn a capacity target straight into a bay count, a lane budget and a peak power figure, use the drive bay and PCIe lane sizer.

Start with bays, not drives

Drive bay count is the least reversible choice in the build. Adding capacity later means either replacing every drive in a group with larger ones or adding a whole new group, and both are expensive. Buying a chassis with more bays than you need today is far cheaper than replacing the chassis.

Bay count also shapes your redundancy math. Parity based layouts spend a fixed number of drives on redundancy per group, so a small group loses a large fraction of its raw capacity to parity while a large group loses less but takes longer to rebuild and stays vulnerable for that whole window. Wide groups of large drives make rebuilds long. That is the real argument for splitting a large array into several smaller groups.

Check backplane details before ordering. Whether bays are hot swap, whether the backplane is a passive breakout or contains an expander, and what connector standard it uses all determine which cables and controllers will work.

One drive specification interacts directly with this. A rebuild writes continuously to the replacement drive for as long as it lasts, which is a workload that drive-managed shingled disks handle badly; CMR vs SMR NAS drives sets out why every member of a parity group should be CMR before the bay count is even settled.

Controllers and PCIe lanes

Onboard SATA ports run out quickly. The standard answer is a host bus adapter, which presents drives to the operating system directly. This matters for software defined storage, which expects raw disks and their SMART data. A RAID controller that hides drives behind its own abstraction interferes with that, which is why HBAs in plain pass through mode are the norm for software arrays. Many RAID cards can be reflashed into that mode.

PCIe lanes are the scarce resource. Consumer platforms provide relatively few lanes from the CPU, and chipset attached slots share bandwidth upstream. An HBA, a network card, and NVMe drives all compete for the same budget. Server and workstation platforms exist largely because they provide more lanes. Confirm the physical slot size, the electrical lane count, and the lane allocation in the manual, because a slot that is physically long may be electrically narrow.

Budget the network card into that total at the same time rather than afterwards. A dual-port 10 gigabit adapter and an eight-port HBA together will consume most of what a mainstream desktop board offers, and the point at which faster networking is actually worth a slot is covered in NAS slow transfer speeds. Current-generation SAS HBAs and their host interface widths are listed on Broadcom’s host bus adapter pages, which is a more reliable reference than a marketplace listing for a reflashed card.

ECC memory

ECC memory detects and corrects single bit errors in RAM rather than letting corrupted data pass through silently. On a system whose job is storing data for years, that matters more than on a desktop, because bad data written to the array is bad data in every subsequent copy and snapshot. ECC support requires the CPU, chipset, motherboard and modules to all support it. Verify support against the board’s documentation rather than assuming it from the platform.

Power and cooling

Sizing the PSU from steady state draw is a mistake. Spinning drives pull noticeably more current at spin up than during normal operation, and if every drive starts at once the peak is many times the idle figure. Staggered spin up spreads that peak, and enterprise backplanes often support it. Check the twelve volt rail capacity rather than the headline wattage, and leave headroom for future drives.

The datasheets give the numbers to work with. Western Digital’s WD Red Plus product brief lists a 12 V peak of 1.9 A for the 12 TB WD120EFGX against average read/write consumption of 8.8 W and idle consumption of 6.1 W. That peak is roughly 23 W on the 12 V rail per drive at spin up, against under 9 W once running, so eight drives starting together is a fundamentally different load from eight drives working. Size the rail for the first case and the efficiency curve for the second.

Check the figure for the drive you are actually buying rather than for the family. Within the same 12 TB WD Red Plus name, the helium WD120EFBX draws a 1.84 A peak and 6.3 W read/write, so a bay full of one model and a bay full of the other are not the same power budget.

For an existing machine, measure rather than add up datasheet figures. A Kill A Watt P3 P4400 inline at the wall gives real idle, working and startup readings, which is the input a UPS runtime chart needs and a spreadsheet of drive specifications is not.

Cooling gets ignored until drives run hot in a stack with poor airflow. Drives packed together need directed airflow across them, not just a fan somewhere in the case.

Common mistakes

Buying drives before the chassis. Choosing a board without checking lane allocation. Assuming a RAID card will behave like an HBA. Sizing a PSU for idle. Filling every bay immediately, leaving no room to migrate data during a future rebuild.