Eight things you never knew about PCI Express
To most of us, PCI Express just means that big old slot where your graphics card sits, but this ingenious interconnect has hidden depths
Peripheral Component Interconnect Express, abbreviated to PCI Express, PCIe or PCI-e, is a high-speed serial computer expansion bus standard that was developed as a standardised interface for motherboard components. These include graphics, memory and storage.
While it’s been a staple in modern computing, and somewhat taken for granted, there are, actually, depths to the standard you may not know. We outline eight things about PCI Express that you may not know, including how it works, and the future of the standard.
PCI Express is old enough to buy a round of drinks
We may think of PCI Express as a modern standard, but it was first introduced way back in 2002. To put into context just how long ago that was, the cross-industry group that developed it included Compaq and a company that still went by the full name of Hewlett-Packard, as well as Dell, IBM, Intel and Microsoft.
The tech was originally called 3GIO, short for “third-generation I/O”. That reflects its role as a successor to the venerable PCI bus and its short-lived follow-up PCI-X. While the name emphasises its heritage, however, PCI Express works in a fundamentally different way from its predecessors. Those older interfaces used parallel data transfer, while PCI Express switched to a serial approach.
This is because, although a serial connection might not transfer so many bits per cycle as a parallel one, it can often reliably run at much higher frequencies (hard disks were going through a similar transition at around this time, moving from the parallel IDE standard to Serial ATA, or SATA for short). With its new architecture, the PCI Express 1.0 interface was able to support bidirectional data transfers at up to 4GB/sec, around four times as fast as the PCI-X hardware that was common at the time.
PCI Express has phenomenal cross-generation capability
The first PCI Express interface was a big hit, quickly supplanting PCI-X in both consumer and server hardware. But the people wanted more: in 2007, the cross-industry PCI-SIG (“Special Interest Group”) introduced the PCI-E 2.0 standard, which doubled the top speed to 8GB/sec in each direction.
Three years later came PCI-E 3.0, which made use of a new encoding system to nearly double the effective top speed again, from 8GB/sec to 15.75GB/sec. We’re now seeing PCI-E 4.0 systems that push the throughput up to 31.5GB/sec.
These generational leaps have underpinned two decades of performance improvements right across computing technology. The really impressive part is that, through them all, PCI-E (or PCIe) has retained full forward and backward compatibility. You can plug a PCI-E 1.0 card in a PCI-E 4.0 slot without any tweaks. You can, alternatively, install a brand-new PCI-E 4.0 card into a computer from 15 years ago, and it’ll work without a murmur – although the slower slot will prevent it from operating at its full capacity.
Slot size correlates with speed… well, mostly
Almost all motherboards offer a mixture of long and short PCI-E slots. A full-length 89mm slot can support the fastest data rates, by aggregating together 16 “data lanes” using 82 electrical contacts between the card and socket. The shortest 25mm slots are identical from a software point of view, but have only 18 contacts and support only a single data lane. These are referred to as PCI-E x1 connectors, while the full-sized slots are called PCI-E x16. There are also standard sizes for x4 and x8 variants.
This complicates the PCI Express standard: a PCI-E card that’s designed for an x16 slot won’t physically fit into an x1 or x4 receptacle. What you might not realise is that PCI-E does work the other way round – so an x1 card will slot happily into an x16 socket and work perfectly well.
The same applies to PCI-E x4 and x8 cards, such as USB expansion cards or hard disk controllers. If your motherboard doesn’t have the right slot, you can just use a larger one. Indeed, some motherboards offer open-ended slots, allowing you to insert (say) an x8 card into an x4 slot. This, too, will work, but the limited bandwidth could cause operational issues.
There’s one other potential catch here. Rather than offering slots in the x4 and x8 formats, manufacturers sometimes provide x16-sized slots that only support x4 or x8 connections. If you suspect a graphics card or other device isn’t running as smoothly as it should, investigate whether that PCI-E slot is really what it appears to be, or a low-speed impostor.
You don’t need to turn off your PC to exchange PCI Express cards
This comes with a caveat, but let’s start with the good bit: PCI Express is hot-pluggable. In other words, you can shove a card into the slot while Windows is running, and it should simply start working. This isn’t merely an anecdotal observation: it’s a deliberate part of the spec. The pins on the card that pick up power from the motherboard are shorter than the ones that carry data, so they don’t make contact until everything else is already lined up and connected.
While the connector is designed to cope with this type of ad-hoc connection (and disconnection), however, that doesn’t guarantee everything else in the system will. If you’re using a PCI-E drive controller, abruptly unplugging it without cleanly dismounting the drives could lead to data being lost, or could even cause applications or the OS to crash when a volume that’s in use disappears.
Note, too, that most graphics cards have additional power connectors that provide current directly from the main power supply. That means that the power will stay on while you’re yanking the card from its slot, which could lead to glitches and crashes. We haven’t heard of anyone damaging their hardware by trying to hotplug a powered graphics card, but we’d suggest you don’t make a habit of it.
The PCI Express interface provides power as well as data
We tend to focus on the data-transfer capabilities of the PCI-E interface, but an x16 slot can also supply a hefty 75W of power to a connected card. That’s not enough to drive a top graphics card (the Nvidia RTX 3090 FE can suck up over 350W when pushed hard), but it should be enough for most other hardware types, even including storage arrays. Smaller slots have a lesser power budget, but an x1 card can still draw as much as 25W.
For hungrier peripherals, two types of external power connector are normally provided – a six-pin plug that’s rated for up to 75W, and an eight-pin variant that can provide up to 150W. These aren’t strictly part of the PCI-E standard – they’re defined in the ATX 2 standard for power supplies, as revised in 2003 – but they are often called PCI Express power connectors, as they were specifically designed to work with PCI-E cards.
They also embody something of the PCI-E philosophy: just as you can mix and match cards and slots, you can do the same with these plugs and sockets. An eight-pin plug will work fine in a six-pin socket (with two pins sticking out at the end), and a six-pin power connector will fit into the left-hand side of an eight-pin socket – although the hardware may not be happy with the reduced power.
PCI-E connectors come in different shapes and sizes
When you think of PCI Express, you probably picture a standard motherboard slot, but there’s no reason why a PCI-E connection needs to be limited to that format. Almost all laptops use PCI-E as an internal interconnect, even if they have no user-accessible slots, and there are plenty of variant PCI-E connectors designed for specific roles.
One familiar example of that is NVM Express (NVMe). SSDs that use the NVMe standard are usually supplied in the form of a 22 x 80mm M.2 module, with a 75-pin connector along the short edge. This is to all intents and purposes a mini PCI-E x4 card, so in a PCI-E 3.0 system it can transfer data at close to 4GB/sec. (The M.2 format can also connect a drive to a computer’s SATA controller, but there’s little call for this capability since NVMe is so much faster.)
Another flavour of PCI Express doesn’t use a slot, but a USB socket. Thunderbolt can carry both regular USB connections and PCI Express data. A Thunderbolt 3 cable supports four PCI-E 3.0 lanes of data, making it fast enough to attach high-speed storage and external graphics cards.
PCI-E 4.0 is a big deal for next-generation storage
Generation 4 of the PCI Express standard was finalised in 2017, and became widely available in 2019 as part of AMD’s Zen 2 (Ryzen 3000) platform. Intel followed last year with its 11th-generation mobile chips, and finally upgraded its desktop processors to PCI-E 4.0 in March 2021. The upshot is, almost any computer bought today will support PCI-E 4.0.
There’s nothing shocking about PCI-E 4.0; like previous updates, it doubles the bandwidth of a data lane, from 985MB/sec to 1.97GB/sec, so an x16 connection can now transfer nearly 32GB per second. At the same time, it retains backward compatibility with earlier PCI-E versions, so you might never even notice which connector your computer has.
The move is significant though, for one big reason. Today’s fastest NVMe SSDs are capable of speeds well in excess of 4GB/sec, but the still-common PCI-E 3.0 x4 interface isn’t. To achieve the full performance potential of a modern SSD, you need a drive and a motherboard that both use PCI-E 4.0. The fastest PCI-E 4.0 drives already offer read speeds approaching 6GB/sec, and future models will surely go all the way up to the maximum of 8GB/sec.
PCI-E 5.0 is already here – and so is its successor
If even 8GB/sec seems too slow, don’t worry: while PCI-E 4.0 is establishing itself in the mainstream, the PCI-SIG has been finalising the 5.0 spec, with sample hardware in production. You won’t be amazed to learn that this upgrade doubles the bandwidth again to just under 4GB/sec for a single lane or 63GB/sec over an x16 link.
Such speeds are overkill right now for consumers, but AI and Big Data aficionados are salivating at the thought of fetching and processing data at such immense rates. PCI-E 5.0 could be helpful for enterprises and data centres too, enabling them to host a 400GbE network card in a single PCI-E 5.0 x16 slot.
Believe it or not, the spec for PCI-E 6.0 has already been published. Using a new encoding scheme, this manages once again to double the maximum data rate to 126GB/sec, yet will still be able to interoperate with older cards and slots. It looks certain, therefore, that PCI Express will still be evolving and thriving at least 25 years after its inception, delivering an astonishing 31.5x increase in performance, while retaining full backward compatibility – an incredible testimony to an underappreciated technology.
2022 State of the multi-cloud report
What are the biggest multi-cloud motivations for decision-makers, and what are the leading challengesFree Download
The Total Economic Impact™ of IBM robotic process automation
Cost savings and business benefits enabled by robotic process automationFree Download
Multi-cloud data integration for data leaders
A holistic data-fabric approach to multi-cloud integrationFree Download
MLOps and trustworthy AI for data leaders
A data fabric approach to MLOps and trustworthy AIFree Download