The last few weeks a lot of information has been leaking about the new graphics solutions from AMD. Specifically, there has been a lot of talk about the RNDA2 that will give life to the next generation of AMD graphics. Now supposed data has been leaked from an AMD Radeon Instinct MI100 based on the CDNA architecture. This graph would be superior to the NVIDIA A100 on FP32.
Finally, AMD has decided to create two different architectures, one for gaming and the other for servers and data centers. The company seeks to offer the best possible performance in both segments and to stop experimenting. AdoreTV has leaked information about the next generation of graphics professions from AMD that are based on CDNA, the professional version of RDNA.
[amazon box=»B07FP2Y3DG» title=»AMD EPYC 7551″ template=»list»]AMD Radeon Instinct MI100 CDNA architecture could be better than NVIDIA A100
This AMD CDNA architecture has been developed for the HPC segment, clearly focused on Data Centers. What is sought is to offer great performance, scalability, good efficiency and specific technologies for AI and Deep Learning.
According to the leaked slide, the AMD Radeon Instinct MI100 could outperform the NVIDIA Ampere A100 by 13%. Note that this NVIDIA solution is not complete, since there would be an even more powerful chip that the company has not released. In any case, it will be necessary to finally see the credibility of the data when it is released.
AMD's solution, in addition to being more powerful, would offer a better quality / price ratio. Under FP64, used for AI, Deep Learning, HPC and the like, the NVIDIA A100 is second to none. Everything suggests that the AMD Radeon Instinct MI100 is designed for other markets than the NVIDIA A100.
MI100 1U Specifications
As can be seen, this solution would be based on the second generation of AMD Infinity, with the GPU manufactured in 7nm. It would be focused on Machine Learning and HPC and would use a system with two EPYC processors, being able to choose between the Rome generation and the Milan generation.
Each of the EPYC processors will have two associated Radeon Instinct MI100s using second generation Infinity Fabric. This makes a total of 4 GPUs with a total power of 136TFLOPs in FP32, which is 34TFLOPs for each GPU. Still, it would still fall short of the 156TFLOPs of the NVIDIA Ampere A100s.
The 4 MI100s would offer a bandwidth that would reach 4.9TB / s because they use HBM2E memories. Individually each Instinct MI100 offers 1.225TB / s of bandwidth, each counting with 128GB HMB2E, according to rumors.
These EPYC-based 2020U models will hit the market in December 1. Versions with Intel Xeon processors will arrive by February 2021.
MI100 3U Specifications
2021U server models will arrive in March 3. These models will feature 8 AMD Radeon Instinct MI100 GPUs and two EPYC processors. Each processor will take 4 MI1000 GPUs that will be connected through XGMI with a total bandwidth of 1.2TB / s. This model would offer 272TFLOPs in FP32 and have a consumption of the entire system of 3.000W.
Source: wccftech
