Get ready for our third and final round of graphics architectural deep dives. Two months ago, we spilled the beans on Nvidia’s Ada Lovelace architecture. Last month, we did the same for Intel’s Arc Alchemist. Last to the party, but perhaps the most interesting in its approach in the realm of GPU design is AMD’s RDNA 3, premiering with the Radeon RX 7900 XTX and Radeon 7900 XT.
Thanks to its use of chiplets, the RDNA 3 architecture fundamentally changes several of the key design elements. AMD pioneered chiplets in the consumer realm with its Zen 2 (Ryzen 3000) CPUs. That has enabled AMD to deliver more cores and better performance, giving Intel some much-needed competition. Now it’s looking to do the same with GPUs and graphics cards.
It’s not an easy transition to make and, of course, GPUs are fundamentally quite different from CPUs. Whereas CPUs typically come with a few dozen general-purpose cores at most—up to 96 with the latest AMD Genoa EPYC processors—GPUs already have thousands of specialized shader ‘cores’, each of which needs access to the large data sets involved with graphics work, such as the textures and the geometry.
So, it’s time to don your swimsuit as we dive into the latest RDNA 3 architecture to see what makes it tick.
THE GPU CHIPLET ERA BEGINS
For now, AMD has only revealed one specific design for RDNA 3, the Navi 31 GPU. It consists of two separate pieces of silicon, the Graphics Compute Die (GCD) and the Memory Cache Die (MCD). The names already provide a good indication of how AMD chose to split up the normal monolithic chunk of silicon that makes up a GPU into multiple chiplets.
The GCD houses the Compute Units (CUs) that provide the raw number
