AMD returns to generalized graphics with the CDNA architecture

With the introduction of Graphics Core Next (GCN) from 2012, AMD invested in a graphics architecture that would form the basis for products on both the consumer side and in the data center. The fact that the architecture was designed to cover two different types of calculations limited its excellence for both tasks.

The answer to the shortcoming with GCN came on the consumer side in the form of Radeon DNA (RDNA) designed specifically for graphics rendering in games. During Financial Analyst Day, AMD is now unveiling the pair horse for data center architecture – Compute DNA (CDNA). As the name suggests, the CDNA is designed to perform computationally intensive tasks such as machine learning and artificial intelligence.

CDNA and RDNA have some common base plates in generalized computation (GPGPU), but otherwise functions specifically designed for graphics rendering have been removed in CDNA. AMD has previously launched specialized products for data centers based on the GCN architecture, such as the Radeon Instinct MI 50/60, but even here the double-bottomed architecture hampered the potential of the cards for the task.

With future graphics cards for data centers, however, AMD does not rely solely on improvements in the CDNA architecture. The communications fabric Infinity Architecture is also at the center of AMD’s ambitions in this area. Infinity Architecture links both processors and graphics cards together, which will provide significantly higher scalability for AMD-based data center solutions.

When CDNA now becomes its own architecture disconnected from the consumer side, the development falls under AMD’s data center manager Forrest Norrod, who outlines a separate product plan for CDNA. With the 2019 GCN-based Vega 20 circuit as a starting point, the first CDNA-based graphics cards will premiere in 2020, together with Infinity Architecture. Here, AMD uses the same improved variant of TSMC’s manufacturing technology 7nm that is used with RDNA 2.

In 2022, this will be followed by CDNA 2, which is manufactured using an unspecified “advanced manufacturing technology”. Here, Infinity Architecture is also expanded to support up to eight interconnected graphics cards (multi-GPUs) whose memory can be synchronized with other graphics cards as well as connected Epyc processors.

Radeon Instinct graphics cards based on CDNA together with Zen 4-based Epyc processors form the backbone of the AMD-based supercomputer El Capitan. The system has a planned capacity of 2 exaflops and launch by the year 2023. El Capitan will be AMD’s second supercomputer on an exascale, where the previously unveiled Frontier system reaches 1.5 exaflops and will be installed in 2021.

Read more about AMD’s graphics architectures: