AMD announces new CDNA GPU architecture: dedicated to data center computing - WritenAreGiven

WritenAreGiven

Technology: Latest News,Review & Leaks

AMD announces new CDNA GPU architecture: dedicated to data center computing

Share This
The AMD CPU architecture has developed smoothly over the past few years. Zen, Zen +, Zen 2, Zen 3, Zen 4 have performed amazingly every generation, and the GPU architecture is relatively dim: GCN has been serving for many years and is quite old. But still can not fully compete with the opponent.

In addition to playing games and rendering graphics, another core task of the GPU is computing. Especially as artificial intelligence, neural networks, and deep learning have increasingly demanding high-performance computing, the GPU must also quickly keep up.
AMD announces new CDNA GPU architecture: dedicated to data center computing

In this respect, NVIDIA is undoubtedly quite successful. From the architecture to the technology, from the product to the market, the situation is good. Especially the architecture has a specific targeted design. The "Volta" used by the Tesla V100 is the most typical.

Although AMD also has a lot of computing card products, especially the Radeon Instinct series, it has achieved quite a lot and even has entered two supercomputers, but the Vega core they use is essentially a gaming GPU.


AMD announces new CDNA GPU architecture: dedicated to data center computing

Today, AMD officially announced a new GPU architecture, "CDNA", which is optimized for GPU computing, thus parting ways with the game's optimized RDNA architecture.

According to the introduction, the CDNA architecture will focus on computing/tensor operating performance, thereby accelerating machine learning, high-performance computing, and can flexibly design performance through the Infinity Fabric interconnect bus, and support enhanced enterprise-grade RAS features, security, and virtualization technologies.

In addition, it will also provide higher energy efficiency ratios, thereby reducing corporate TCO costs.

As for more specific architectural details, AMD will be announced in the future.


AMD announces new CDNA GPU architecture: dedicated to data center computing

Of course, the CDNA architecture has little to do with the RDNA architecture designed specifically for game scenarios, and there is little overlap and similarity. In fact, it has a closer relationship with the GCN architecture.

On the AMD GPU computing architecture roadmap, 7nm GCN is listed as the first-generation data centre GPU, and its products include Radeon Instinct MI50 / MI60.

The CDNA architecture is expected to be born from the end of this year to the beginning of next year. It will continue to use the 7nm process and will use the second-generation Infinity Fabric interconnect bus architecture, which is now applied in the 7nm Zen 2 architecture so that it can be used with the Zen architecture. Xiaolong collaborates efficiently.

The subsequent second-generation CDNA 2 does not have a specific process. It only says that it is a more advanced node. It may be 7nm EUV or 5nm. Technically it will be upgraded to the third-generation Infinity Fabric. And expand to support Exascale calculations (Exascale), it is expected to launch related products in 2022.


AMD announces new CDNA GPU architecture: dedicated to data center computing

For the third generation of Infinity Fabric, AMD also revealed some advanced features, including higher bandwidth and lower latency CPU-GPU interconnect, GPU-GPU memory integration, improved consistency performance and simplified programming, etc.


AMD announces new CDNA GPU architecture: dedicated to data center computing

No comments:

Post a Comment

Like a Reply

Post Bottom Ad