Learn

PCIe: the bus that connects a GPU to the rest of the machine

PCIe (Peripheral Component Interconnect Express) is the standard high-speed connection that plugs expansion cards, most often a discrete graphics processor (GPU), into the rest of a computer. Data crossing it is fast but still slower than memory the chip reaches directly, which matters when a model's data has to move back and forth.

At a glance

What it is
The bus connecting expansion cards, like a GPU, to the processor
What rides it
Most often a discrete graphics card and its memory
Why it matters
Moving data across it costs time the GPU is not computing
On a DGX Spark
The GPU shares one memory pool, so model data does not cross PCIe to reach it
Flow

Where PCIe sits in a discrete-GPU machine

On a desktop, data travels across the PCIe link between the main processor's memory and the graphics card. The link is fast, but every crossing is time the GPU spends waiting instead of computing.

1
Processor and system memory where data starts, on the main board
2
PCIe link the bus the data crosses to reach the card
3
Discrete GPU and its VRAM where the model runs, on the far end

What is PCIe?

PCIe (Peripheral Component Interconnect Express) is the bus, the shared high-speed connection, that you plug expansion cards into. The most important one for local AI is a discrete graphics card: it sits in a PCIe slot and talks to the processor and system memory over that link. The link is fast, but it is still a link, and anything that travels across it takes a measurable amount of time that the GPU is not spending on actual computing.

That cost is usually invisible until your work involves a lot of back-and-forth. When a model and its data do not fit in the card’s own memory and have to be shuttled in and out over PCIe, the bus becomes the thing you are waiting on rather than the processor.

Why does it barely matter on a DGX Spark?

A DGX Spark does not put the GPU on the far end of a PCIe link. The graphics processor and the main processor share one pool of memory on the same package, so the model’s data is already where the GPU can reach it. There is no discrete card to feed across a bus. That is the whole point of a unified-memory design: it removes the hop that, on a desktop, you spend effort avoiding. The trade is that the memory is fixed at purchase, not a card you can swap out later for a bigger one.

Discrete GPU over PCIe

  • GPU is a separate card on the PCIe bus
  • Model data crosses PCIe to reach the GPU's VRAM
  • You can swap or add cards
  • Each crossing of the bus costs latency

Unified memory (DGX Spark)

  • GPU and processor share one pool of memory
  • Model data does not cross PCIe to be used
  • Memory is fixed, not a card you swap
  • No bus hop between processor memory and GPU

Related terms

Go deeper

← All terms Reviewed: June 2026