In the graphics card market, NVIDIA is far ahead of AMD in both game cards and acceleration cards. Even the ampere architecture graphics card A100 released by NVIDIA two years ago has more than twice the performance of the insight mi250 released by AMD at the end of last year, with obvious advantages.
NVIDIA recently announced the comparison between A100 accelerator card and AMD mi250 accelerator card in its official blog. The test mainly focuses on the popular applications such as lammps, namd, openmm, loomacs and amber in the data center.
Under the same 4-channel GPU configuration, the performance ofa100 graphics card is at least 10% higher than that of other cards, and more than 40% higher. In the highest amber applications, the performance advantage of A100 is 110%, twice as high
In the case of a single card, the A100 can lead the mi250 by up to 90%.
If the energy efficiency is compared, NVIDIA is more confident. Compared with AMD mi250, thea100 has a maximum energy efficiency advantage of 180%, which is very important for the data center
You should know that A100 is an acceleration card released by NVIDIA in May 2022. It is manufactured by TSMC with 7Nm process, integrates up to 54.2 billion transistors, has a core area of 826 square mm, has 6912 fp32 CUDA cores, has a maximum acceleration frequency of 1.4GHz, and is equipped with 512 bit 40Gb hbm2 video memory, with a bandwidth of 1.6tb/s.
Compared with the V100 core of the previous generation volt architecture, its int8 reasoning and fp32 training performance are improved by 20 times, and fp64 computing performance is improved by 2.5 times, which is the largest progress in history.
As for the mi250 accelerator card, this is the accelerator card released by AMD last November. It is upgraded to a new cdna2 computing architecture, combined with the upgraded 6nm FinFET process, 58billion transistors, and uses the 2.5D EFB bridging technology. It is the industry's first multi die integrated packaging (MCM), with two cores integrated internally.
The insight mi250 has 208 computing units, 13312 stream processor cores, a 128GB hbm2e with 8192 bit width, a frequency of 1.6GHz, a peak bandwidth of 3276.8gb/s, and supports full chip ECC.
Amd said that the double precision performance of the insight mi200 series is up to 4.9 times higher than that of the competitive products and up to 4 times higher than that of the previous generation
Coincidentally, a few days ago, Microsoft also first selected AMD's mi200 series accelerator card to provide AI acceleration for cloud computing. AMD's 5x friend business has very good and powerful performance. NVIDIA immediately released the test without the latest H100 accelerator card. The A100 accelerator card two years ago was compared with the mi250 accelerator card.
