How breaking a single chip into pieces unlocked massive GPU power

Science · Published 1 November 2026 Swipe more like this →

In short: For years, high-performance graphics cards relied on giant, singular silicon dies. AMD broke this mold with RDNA 3, moving to a modular chiplet design. By splitting the hardware into separate pieces, they achieved better manufacturing yields and allowed different parts of the chip to be built on different production lines.

AMD's Senior Vice President Sam Naffziger initiated the transition towards consumer GPU chiplet designs during the final months of 2017. The primary motivation was economic and physical: manufacturing one massive monolithic die is expensive and prone to defects. By using smaller dies, more units can fit on a single wafer, maximizing yields. Furthermore, it allows for functional optimization; logic components can be built on advanced 5 nm nodes, while cache and SRAM—which do not scale as linearly in density or power—can be fabricated on the cheaper, more mature N6 node.

RDNA 3 utilizes two distinct types of chiplets: Graphics Compute Dies (GCD) and Memory Cache Dies (MCD). To connect these pieces without the high cost of a silicon interposer, AMD uses an integrated fan-out packaging technique. This creates a silicon bridge providing nearly 10x the bandwidth density compared to the organic substrates used in previous Ryzen or Epyc processors. These interconnects achieve a cumulative bandwidth of 5.3 TB/s, enabling the high-speed data flow required for modern gaming and AI.

This architecture also enables unique hardware features. Each MCD contains 16 MB of L3 cache and two physical 32-bit GDDR6 memory interfaces. For example, the Radeon RX 7900 XTX uses six of these MCDs to create a massive 384-bit memory bus. Additionally, RDNA 3 introduced a dedicated media engine for AV1 encoding and a specialized scheduler for AI acceleration. This allows the GPU to handle matrix instructions on normal compute cores, significantly boosting inference performance for tasks like generating images with Stable Diffusion.

The evolution continued in 2024 with RDNA 3.5, which targets mobile and handheld devices. Instead of raw power, it focuses on efficiency by doubling per-cycle texel output from texture units to compensate for lower clock speeds. This ensures that low-power laptops can maintain high performance without draining batteries rapidly.

General information, not professional advice. Spotted an error? Flag the card or see corrections and complaints.

Intellect Plus is a visual educational discovery platform for exploring fascinating facts and ideas. Brain Food, Not Brain Rot. What is Intellect Plus?

Related

• Why AMD Abandoned Its Own Factories to Become a Global Semiconductor Powerhouse
• Intel's architectural pivot from monolithic silicon to disaggregated chiplets
• Intel’s pivot to chiplets marks a major shift in processor architecture
• The font that powers half of Google's web views
• How high-end GPUs fight the invisible interference of cosmic radiation
• How a backported microarchitecture redefined the limits of 14nm silicon
• The software that dominated 95% of the professional publishing market
• The tiny annotations that bridge the gap between logographs and speech
• The fortress of privacy faces a paradox of trust
• The browser that prioritizes one-handed navigation and ad-blocking over defaults
• The engineering required to sustain 1600 nits of brightness indefinitely
• The internet suite that refused to let go of its past

More in Found by OrchestratorBot · More in Science · Latest · A–Z · All topics