In March, AMD announced a new CDNA architecture for high-performance computing, AI artificial intelligence, rather than gaming and graphics applications, which are all given to rDNA architecture. This is somewhat similar to NVIDIA’s thinking, such as the latter’s Volt architecture, Turing architecture, dedicated to computing and gaming.
More than three months after the architecture was announced, when will the product be born? AMD CTO Mark Papermaster today publicly confirmed that the first product of the CDNA architecture, the Radeon Instinct MI100, will be released as scheduled in the second half of this year for the high-performance computing market.
He didn’t give further details, but the name Radeon Instinct MI100 has been heard many times, revealing that its core frequency is 1091-1334MHz, with 32GB HBM2 memory, optional Samsung orSK Hynix, frequency 1GHz, and a full-card power consumption of just 200W.
Of course, this should be only the specifications of engineering sample cards, the frequency is relatively conservative, but also can also see that the energy efficiency ratio will be very impressive.
According to AMD’s previous presentation, the CDNA architecture is optimized for GPU computing, focusing on compute/zhang operation performance to accelerate machine learning, high-performance computing, and flexible design performance through the Infinity Fabric interconnect bus, supporting enhanced enterprise-class RAS features, security, virtualization technologies, and a higher energy efficiency ratio, reducing enterprise TCO costs.
On the roadmap, the first generation of CDNA architecture products uses the 7nm process and uses the second generation Infinity Fabric interconnect bus architecture, which is now used in the 7nm Zen 2 architecture, allowing efficient collaboration with Zen architecture dragons.
The second generation of CDNA 2 is not yet specific, or the more advanced node, but will technically upgrade to the third generation of Infinity Fabric and expand its support for 10 billion compute (Exascale), with products expected to be available in 2022.
CNDA architecture computing card came out in the second half of the year, see the node just in sync with the Zen 3 architecture of the third generation of dragons, can be called a perfect match, but also will be NVIDIA ampere architecture of the new products to compete.