AMD releases Instinct MI250X benchmarks result
A few days ago, AMD announced the launch of the Instinct MI200 series of computing cards based on the CDNA 2 architecture. This is the first GPU in an MCM multi-chip package and the first GPU accelerator. What’s more unexpected is that AMD now also announced the comparison test data with the competitor Nvidia’s A100 computing card. Although it is not uncommon for hardware companies to show performance advantages at some events, there are not many detailed actual comparison test data published on the official website, which shows that AMD is very confident in the performance of this product.
AMD Instinct MI200 series computing cards are equipped with GPU code-named Aldebaran, with 58 billion transistors, manufactured using TSMC’s 6nm process, with 128GB of HBM2e memory (total bandwidth 3.2 TB/s), and TDP of 560W. This series has two products, Instinct MI250 and Instinct MI250X, with 208 CUs (13,312 stream processors) and 220 CUs (14080 stream processors) respectively.
The small chips in the GPU are supplemented by XGMI internal interconnection design, all have a VCN 2.6 controller, and use 2.5D Elevated Fanout Bridge (EFB) packaging technology. In addition, the two small chips in the GPU are connected through AMD Infinity Fabric technology, providing up to 100 GB/s of bidirectional bandwidth. This belongs to the third-generation Infinity Fabric technology, which can manage up to 8 Infinity Fabric links, connect the Instinct MI200 series computing card to the third-generation EPYC processor to achieve CPU and GPU memory consistency and maximize system throughput. Compared with the previous generation of Instinct MI100 based on the cDNA architecture, the Instinct MI200 series of computing cards based on the CDNA 2 architecture has a very large performance improvement.
Obviously, AMD Instinct MI200 series computing cards are mainly designed for HPC and AI workloads. AMD’s official comparison test involves processing algebra, physics, cosmology, molecular dynamics, and particle interactions. Many of them are widely used and industry-recognized tests, such as LAMMPS and OpenMM.
However, it’s worth noting that AMD lacks projects on AI benchmarking. On paper, AMD Instinct MI200 series computing cards will have higher performance in traditional HPC and matrix workloads, while NVIDIA should still have some advantages in AI.
From the results of these benchmark tests, AMD Instinct MI250X has an absolute performance advantage compared to NVIDIA A100, which is between 1.4 times and 3.05 times that of the other party. It can be said that Instinct MI250X is the most powerful HPC computing card in the world. Considering that AMD Instinct MI250X was launched about a year and a half later than NVIDIA A100, coupled with the adoption of a new architecture and the use of more ALUs, it is only natural to have such test results.
In a few months, Intel will launch Ponte Vecchio. This Intel technology master should have quite a powerful performance.
