AMD Next-Gen Radeon GPUs to Deliver Massive AI Performance Leap With AT2 Specs

AMD is developing its next-generation RDNA 5 graphics architecture to include dedicated AI acceleration units, according to hardware leaker Kepler_L2 speaking on the AnandTech forums. The upcoming architecture, succeeding the current Radeon RX 9000 series powered by RDNA 4, aims to double artificial intelligence performance per compute unit compared to previous hardware while introducing support for FP4 numeric formats.

RDNA 5 AI Architecture and Matrix Cores

RDNA 5 shifts away from shared blocks by introducing standalone AI accelerators, likely utilizing matrix cores derived from the CDNA 5 architecture featured in Instinct MI400 accelerators, as reported by Kepler_L2. Leaker Kepler_L2 stated that gaming-focused RDNA 5 matrix cores will deliver half the performance of CDNA 5 implementations, internally designated as Gfx1250. This design yields twice the AI workload performance of RDNA 4, known internally as Gfx1200.

AT2 GPU Specifications and Memory Bandwidth

While standard RDNA 5 desktop graphics cards are not anticipated until 2028, a specific GPU designated as AT2 is slated to arrive earlier, potentially within the coming year, according to recent hardware leaks. AT2 serves as an approximate successor to the Navi 48 chip found in the Radeon RX 9070 XT. Instead of a 256-bit bus utilizing GDDR6 memory, the AT2 processor adopts a 192-bit bus paired with GDDR7 memory operating at speeds up to 36.0 GHz, delivering a memory bandwidth of up to 864 GB/s, based on disclosures from hardware leaks.

Did you know? The AT2 chip is planned with 70 compute units, providing 4,480 shaders alongside a 192-bit memory interface, according to information shared by Kepler_L2 on AnandTech. This unusual count results in an odd number of 35 WorkGroup Processors, each containing a pair of compute units.

Compute Units and Performance Projections

Initial estimates regarding the AT2 compute unit count varied widely, with Moore’s Law is Dead pointing to 64 units while older reports suggested 80. Kepler_L2 confirmed a physical configuration of 70 compute units. This layout grants roughly 9.375 percent more compute units than the fully active Navi 48 chip in the Radeon RX 9070 XT. Target clock speeds for the platform range between 3.1 GHz and 3.4 GHz, compared to the 2.97 GHz clock speed of Navi 48. When factoring in higher clock speeds and increased compute unit counts, hypothetical raw performance gains over the Radeon RX 9070 XT range from 14.2 percent to 25.2 percent, though actual gaming performance will depend heavily on driver optimizations and architectural nuances.

High-End RDNA 5 and Future Roadmap

Beyond mainstream hardware like the AT2 processor, AMD previously outlined plans for a larger high-end processor designated as AT0 in earlier architectural leaks. The AT0 design features a 512-bit memory bus utilizing GDDR7 memory alongside approximately 184 compute units encompassing 11,776 shaders, mirroring the class of high-end Nvidia offerings such as the GeForce RTX 5090. Company plans remain unconfirmed regarding whether AT0 will launch as a consumer gaming graphics card or be reserved strictly for enterprise server deployments and cloud gaming streaming instances, according to industry disclosures.

Frequently Asked Questions

When will RDNA 5 graphics cards be released?

Mainstream RDNA 5 graphics cards are anticipated around 2028, though the AT2 GPU variant may arrive earlier, according to hardware roadmaps.

What memory type does the AT2 GPU use?

The AT2 GPU uses GDDR7 memory over a 192-bit memory bus, achieving up to 864 GB/s of memory bandwidth, according to leaks reported by Kepler_L2.

How many compute units does the AT2 chip have?

The AT2 chip features 70 compute units, translating to 4,480 shaders distributed across 35 WorkGroup Processors.

Stay Updated on Hardware Leaks

What are your thoughts on AMD’s upcoming RDNA 5 architecture and the shift toward dedicated matrix cores? Join the discussion in the comments below, share this article with fellow hardware enthusiasts, or subscribe to our newsletter for the latest GPU updates.

Leave a Comment