AMD’s next-generation RDNA 4 graphics architecture will see the company focusing on the performance segment of the market. Unlike its predecessor architectures, RDNA 4 will not include successors to the enthusiast-segment chips, such as “Navi 21” and “Navi 31.” Instead, it emphasizes improving performance and efficiency in high-volume segments, akin to the original RDNA-powered Radeon RX 5000 series.
Navi 48 and Navi 44: The Core Chips
The RDNA 4 generation introduces two main chips that have sparked considerable interest: “Navi 48” and “Navi 44.” Among these, “Navi 48” is the faster chip, designed to power the top SKUs of this generation. “Navi 44,” on the other hand, is expected to serve as the mid-tier chip.
Memory Interface and Speed
According to reliable sources such as Kepler_L2 and VideoCardz, the top “Navi 48” silicon features a 256-bit wide GDDR6 memory interface. There will be no upgrade to GDDR7, but the top SKU based on this chip, “Navi 48 XTX,” will offer a memory speed of 20 Gbps, resulting in 640 GB/s of memory bandwidth.
The next-best SKU, codenamed “Navi 48 XT,” will have a slightly lower 18 Gbps memory speed on the same bus-width, providing 576 GB/s of memory bandwidth. The “Navi 44” chip, with a 192-bit wide memory bus, will feature a top SKU with 19 Gbps speed, equating to 456 GB/s of bandwidth.
Infinity Cache and Architecture
The RDNA 4 chips also come with impressive Infinity Cache sizes. The “Navi 48” models, including both “Navi 48 XTX” and “Navi 48 XT,” will feature 64 MB. In comparison, the top “Navi 44” SKU will come with 48 MB of Infinity Cache.
Notably, the RDNA 4 family will adhere to traditional monolithic silicon designs, moving away from the chiplet disaggregation approach used in the “Navi 31” and “Navi 32.”
Ray Tracing and Performance Improvements
Rumors from Moore’s Law Is Dead suggest that AMD’s RDNA 4 design focus will prioritize acing performance, performance-per-watt, and the cost-efficiency of ray tracing. To improve ray tracing performance, AMD may integrate not one but two ray accelerators per compute unit and rely more on fixed-function acceleration rather than programmable shaders.
Overall, RDNA 4 is poised to provide substantial improvements in performance and efficiency, targeting the segments where NVIDIA currently achieves high volumes, if not margins.
