Friday, June 26, 2026
EN·DarkSubscribe
AI Infrastructure · News & Analysis
HomeChips & HardwareReport
Chips & Hardware · Report

AMD's upcoming RDNA 5 and UDNA GPU architectures may ship in multiple configurations with 96, 40, 24, and 12 compute unit options.

Offers product segmentation across performance tiers for datacenter, professional, and consumer markets, challenging Nvidia's GPU pricing ladder.
Trade pressSlicast · August 27, 2025 · Global · Source: wccftech.com
importance 55

Kepler_L2, a leaker with a track record of accuracy on AMD and Intel information, has posted block diagrams of at least four AMD RDNA 5 / UDNA SKUs to the Anandtech Forums. The top-tier die features 8 Shader Arrays with two shader engines each, totaling 16 shader engines, with 6 compute units per shader engine for a combined total of 96 Compute Units. Each Shader Engine has its own RB (Render Backend) unit connected to a central SoC block containing the Graphics Command Processor, Graphics Engine, HWS, and L2 cache. The die includes 16 Unified Memory controllers at 32-bit each, providing a maximum bus size of 512-bit, and could pack up to 128 MB of Infinity Cache if AMD maintains its current configuration.

The next variant in the stack features 40 Compute Units across 8 Shader Engines arranged in 4 Shader Arrays, with 5 CUs per Shader Engine. This configuration includes 6 memory controllers, yielding a 192-bit bus interface, and could support up to 48 MB of total Infinity Cache. The entry-level configurations scale from 24 down to 12 CUs: the 24 CU die contains four shader engines/arrays with 6 compute units per engine and 8 memory controllers (either 16-bit or 32-bit); the smallest die features just two shader arrays with 6 compute units total and 4 memory controllers (either 64-bit or 16-bit). These smaller variants might include 32 MB and 16 MB of Infinity Cache respectively.

Kepler also noted that one RDNA 5 Compute Unit is equivalent to a single RDNA 4 WGP (Work Group Processor), meaning the actual core count would be twice the reported CU count, since RDNA 4 WGPs contained two Compute Units each. Additionally, the Instinct lineup's CDNA 5 architecture is set to introduce increases in local cache, rising from 32KB L0 and 160 KB LDS cache on CDNA 4 to 448 KB of Shared L0/LDS cache in the MI400 series. While these are datacenter parts, the unified UDNA designation suggests these architectural improvements may extend to consumer-tier chips.

According to separate reporting from Chiphell forum member ZhangZhonghao, the RDNA 4 lineup may feature four configurations: a very large top-tier die, followed by mid-tier, small-tier, and tiny-tier SKUs. AMD's next-generation gaming lineup is expected to enter mass production by Q2 2026, coinciding with NVIDIA's "SUPER" series release and Intel's BMG-G31 "Big Battlemage" lineup launch. Both rumors should be treated with appropriate caution pending official AMD announcements.

Read the original
AMD's upcoming RDNA 5 and UDNA GPU… · Slicast