tech

Modern CPUs and GPUs from AMD, Intel, and others increasingly ship as a bundle of smaller dies inside one package.

Why do chip designers split a big chip into several smaller "chiplets"?

Q: Why do chip designers split a big chip into several smaller "chiplets"?

To improve yield and mix process nodes. Smaller dies have fewer defects each, so more of them pass testing and overall yield rises, and each chiplet can use the process node that suits it best. This modular approach lets one package combine cutting-edge logic with cheaper, mature support chips.

To improve yield and mix process nodes ✓
To make the chip run at a lower clock speed
To remove the need for any packaging

To improve yield and mix process nodes
Smaller dies have fewer defects each, so more of them pass testing and overall yield rises, and each chiplet can use the process node that suits it best. This modular approach lets one package combine cutting-edge logic with cheaper, mature support chips.

Chiplets improve yield because smaller dies have fewer defects each, so a higher share pass quality testing compared with one giant monolithic chip where a single flaw ruins the whole thing; splitting the design also lets each piece be made on whichever process node best suits its job, mixing cutting-edge and cheaper mature silicon in one package.

Chiplets don't inherently make a chip run at a lower clock speed, that depends on the design and workload, not whether the silicon is split into pieces. They also don't remove the need for packaging, if anything, chiplets require more advanced packaging to connect the separate dies together.

This modular approach mirrors how the auto industry assembles cars from parts made by many suppliers rather than casting one giant piece, letting chipmakers mix and match components for cost and performance in ways a single monolithic design never could.

▶ Quration Play

tech

What makes a gate-all-around transistor different?What is the main goal of the US CHIPS and Science Act?As a pure-play foundry, what does TSMC actually do?What is the name of Google's in-house chip designed for this purpose?What approach are data centers increasingly using to keep these dense AI racks from overheating?Which company is the sole manufacturer of these EUV lithography machines?What is this phishing-resistant, passwordless credential stored on your device called?What is this security model called?What is this branch of cryptography commonly called?What is this type of attack, which reaches its target through a trusted supplier or dependency, called?What term describes this scam that uses AI to clone a real person's face or voice?Which kind of file is being phased out here?

Quration — Quration Play