Tech Explainer: What is PCIe?

It’s hard to imagine modern PCs and servers without the speed and flexibility of PCIe. Thanks to this protocol, users can quickly and affordably add functionality in place. But how does it work? Find out here.

Peripheral Component Interconnect Express, better known as PCIe, is a high-speed data-transfer system built into a computer’s motherboard.

PCIe appears on the surface of the motherboard as a collection of empty expansion slots. Its game-changing facility is realized when users fill those slots with expansion cards.

The modern incarnation of these cards is often hungry for bandwidth. It’s that bandwidth that sets PCIe apart from other data-transfer protocols such as USB-C and Wi-Fi.

Compared with those protocols, PCIe much more closely matches the incredible speed of internal components. That includes the lightning-fast data sharing between a server’s CPU and RAM.

Given this attribute, it’s hard to imagine modern PCs and servers without the speed and flexibility of PCIe. Thanks to this protocol, users can quickly and affordably add functionality in place. Without it, we’d be replacing our machines far more often.

A Little History

PCIe was originally conceived by Intel back in 2001. By 2004, just three years later, the protocol had replaced its PCI predecessor. At the time, PCIe’s primary use was enabling users to add discreet GPUs, sound cards and network-expansion devices.

Today, more than 20 years later, those use cases are still valid. But the most exciting use of enterprise-class PCIe expansion these days is extending AI-workload functionality.

For this use case, the protocol enables IT admins to add high-performance AI accelerators to PCIe-equipped servers. This is a must-have when it comes to tackling today’s toughest AI workloads, including AI training, inferencing, deep learning and agentic AI.

Two Innovations

What makes PCIe so special? Two innovations in particular stand out.

The first is called serial over parallel. This is PCIe’s ability to increase the speed of data transfer by sending a single bit of data at a time, extremely quickly, over just two pairs of wires—one pair for sending, the other for receiving.

These wires are actually microscopic metal channels known as traces. Because they’re so small, traces are etched directly onto a printed circuit board. There, they carry data throughout the system, terminating in pins embedded in a PCIe slot connector.

PCIe’s serial-over-parallel method may sound counterintuitive, especially when you compare it with the original PCI. That older technology sent multiple bits simultaneously across 32 individual traces. That sounds like a better methodology—until, that is, you factor in PCI’s Achilles’ heel: cross talk.

Cross talk is the crippling interference caused when those multiple bits traveled over many traces simultaneously. The interference got so bad, it effectively capped PCI’s bandwidth at 133 MB/sec., a fraction of PCIe’s throughput.

The effect of the new PCIe design was like replacing a dangerous multilane highway with a high-speed rail system. Bits of data no longer crashed into each other. The norm then became extremely fast, extremely safe travel.

The second ground-breaking innovation was introduced in the latest revision, PCIe 6.0. It’s called Pulse Amplitude Modulation, 4-level.

PAM4 signaling bests previous PCIe designs by increasing the number of bits per cycle. PCIe generations 1.0 through 5.0 were stuck encoding just 1 bit per cycle with a simple binary voltage on/off signal. With PAM4 came the use of 4 distinct voltage levels.

The resulting throughput in PCIe 6.0 is 8 GB/sec. per lane. That’s double PCIe 5.0’s maximum per-lane throughput of 4 GB/sec.

What’s Up with Lanes?

PCIe bandwidth is often discussed in terms of lanes. For instance, we can say PCIe 5.0 is capable of 4 GB/sec. per lane throughput.

However, it’s sometimes necessary to combine additional lanes. This is needed to accommodate high-performance expansion cards that require more throughput.

In such cases, manufacturers design motherboards with x4, x8 and x16 expansion slots. These numbers denote how many lanes are combined in a given PCIe expansion slot.

For example, a PCIe 5.0 x16 slot offers 16 single lanes banded together. This produces a combined 64 GB/sec. of bandwidth (calculated as 16 lanes x 4 GB/sec.).

AMD’s New PCIe Powerhouse

One great example of an AI accelerator that demands maximum available bandwidth is the new AMD Instinct MI350P PCIe card.

This GPU-on-a-card offers hardcore AI-processing performance with 8,192 stream processors, 128 compute units, and 512 matrix cores operating at up to 2.2 GHz.

To support its 600-watt maximum typical board power (TBP), this expansion card requires two PCIe 5.0 x16 slots and sufficient power.

This AMD accelerator also raises an important question: What happens when AMD’s next-generation AI accelerator meets a couple of next-generation PCIe 6.0 slots?

The answer is part of the magic of the PCIe system. Each new PCIe revision maintains backward compatibility. That means a PCIe 5.0 card like the AMD instinct MI350P will work optimally in either a PCIe 5.0 or 6.0 slot.

What’s Ahead

It’s a given that each new PCIe revision will offer more bandwidth than its predecessor. In an age of rapidly shrinking silicon and sci-fi-come-to-life tech, that hardly qualifies as news anymore.

What’s really interesting is the question of what it will take to achieve specs worthy of the AI-powered future we imagine for ourselves.

Here’s a hint: Copper will only get you so far.

To imagine the theoretical magic of future PCIe revisions, consider the transition from Alexander Graham Bell’s 19th-century copper-wire telephone network.

As fiber optics slowly took over, we discovered that light as a transmission medium was far faster and more efficient than pushing electrical currents over metal wires. We gave our data the speed of light.

To achieve the promise of nearly limitless bandwidth, PCIe design engineers may need to make the same leap. If they succeed, their work could light the way to something truly amazing.

Do More: