HOW COMPUTE WORKSBy Michael Hsu

GPU package

What is a GPU compute die?

The compute dies contain the circuits that execute the workload. Memory supplies data; the compute hardware performs operations on it and produces results. A die is one piece of silicon, while a package can contain several. This schematic shows two compute dies to make that distinction visible.

Explore it in the 3D GPU model

What it does

This piece of silicon contains circuits that perform calculations. Memory supplies the data; the compute die works on it and produces results.

Why it sits where it does

The compute dies sit near memory to exchange data through the package. This model shows two dies working together, although GPU designs vary.

Connecting dies is one way to build a larger processor. The connection must move information efficiently enough for the dies to cooperate.

How it works

Many circuits carry out operations in parallel, meaning at the same time. Software organizes the work. Connections between the dies let them cooperate.

Limits and trade-offs

More processing circuits help only when they have enough data, power, and cooling. The software also needs to divide the work effectively.

More compute helps only when software, memory, power, and cooling can keep it useful.

In Blackwell Ultra and MI300X

NVIDIA Blackwell Ultra and AMD MI300X

A lithography exposure field is 26 × 33 mm, about 858 mm², which caps how large one die can be. Chiplets get around that cap and can improve yield and mix process nodes. In exchange, links between dies add energy and latency, assembly adds cost and yield risk, and software has to care where data lives.

NVIDIA Blackwell Ultra

Two reticle-sized dies share a package and communicate over NV-HBI.

Streaming Multiprocessors (SMs) contain CUDA cores, Tensor Cores, registers and shared memory. L2 cache and local memory reuse operands before fetching them again from HBM.

See it in the NVIDIA Blackwell Ultra model

AMD MI300X

Eight XCDs sit above four I/O dies. Two XCDs share each I/O die, making the vertical connections visible in the separated view.

Compute Units contain vector and matrix execution units, registers and local memory. Each XCD has 4 MB of L2, backed by 256 MB of Infinity Cache shared across the package.

See it in the AMD MI300X model

How it is made

  1. A chip designer creates the circuit design.
  2. A chip factory builds circuits on a silicon wafer, which is divided into individual dies.
  3. The dies are tested and assembled with the other parts of the package.

How it developed

  1. Blackwell

    Two dies, one GPU

    NVIDIA describes a two-die design joined by a chip-to-chip interconnect.

    NVIDIA · Blackwell architecture ↗
  2. Hopper → Blackwell

    Splitting the processor across dies

    NVIDIA’s architecture comparison shows one die in Hopper and two in Blackwell. A fast link lets the two Blackwell dies operate as one GPU.

    NVIDIA · Inside Blackwell Ultra, August 2025 ↗
  3. August 2025

    More than just adding circuits

    NVIDIA describes Blackwell Ultra with two compute dies and HBM3E. Its design combines processing, memory, and connections as a system.

    NVIDIA · Inside Blackwell Ultra, August 2025 ↗

Where it is used

Key terms

Die
A piece of silicon containing electronic circuits.
Package
The assembly that holds chips and connects them to a larger system.

Sources

The 3D models are illustrative teaching schematics, not manufacturer CAD. Counts and layouts are simplified where the sources say so.

More explainers