AMD´s next-gen UDNA: four die sizes, one potential 96-cu flagship

Leaked block diagrams attributed to hardware leaker Kepler_L2 outline a modular AMD UDNA graphics family spanning four die sizes and a possible 96 compute unit flagship, with detailed shader array, engine, and memory controller layouts.

Leaked block diagrams from well-known hardware leaker Kepler_L2 sketch a modular next-generation AMD UDNA graphics family that could come in four die sizes, with the largest configuration rumored to top out at 96 compute units. The images depict repeating shader arrays and shader engines, with each engine containing multiple compute units and its own render backend. Those shader engines appear to feed into a central system-on-chip block that houses the graphics command processor, the graphics engine, hardware schedulers, and a shared L2 cache. On the memory side, the flagship diagram shows sixteen unified memory controllers, each listed as 32 bits wide, which would total a 512-bit external interface, and the leak also mentions a potentially much larger on-die Infinity Cache for the top-tier die.

Mapping the diagrams to specific configurations produces numbers consistent with the 96 compute unit claim. The flagship layout is described as eight shader arrays with two shader engines per array, yielding sixteen shader engines. At six compute units per engine, that configuration reaches the 96 compute unit figure. A mid-tier design reduces the scale to four shader arrays and eight shader engines with five compute units per engine, which works out to about 40 compute units and an estimated six memory controllers for a 192-bit memory bus. The leak further shows smaller designs with roughly 24 compute units and 12 compute units built from smaller arrays and fewer memory controllers.

One unusual detail is the 24 compute unit design being shown with as many as eight memory controllers, suggesting the controller widths could vary depending on whether the part uses a traditional gddr-style interface or an lpddr5x-like arrangement. Overall, the diagrams emphasize a highly modular architecture that can scale shader arrays, shader engines, compute unit counts, and memory controller counts across multiple die sizes. The details come from a rumor and leaked diagrams and should be treated as unconfirmed until AMD provides official specifications.

62

Impact Score

NVIDIA Blackwell leads MLPerf Training 6.0

NVIDIA’s latest MLPerf Training 6.0 results put Blackwell across every benchmark in the suite, including new MoE workloads. Partner systems from Microsoft Azure and CoreWeave highlighted large-cluster runs on Llama 3.1 405B and DeepSeek-V3 671B.

HPE and NVIDIA expand AI Factory for agentic systems

HPE and NVIDIA are adding agent tooling, confidential computing and updated accelerated systems across the HPE AI Factory portfolio. The expansion targets production deployments that need governance, secure data handling and integrated networking.

Coherent expands Texas optics plant for AI data centers

Coherent’s Sherman expansion will increase production of indium phosphide wafers used in lasers, transceivers and optical modules for large AI data centers. NVIDIA CEO Jensen Huang framed the site as part of a broader push to rebuild U.S. semiconductor manufacturing.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.