NVIDIA DGX-A100 with AMD EPYC Rome processors, one more blow to the Intel Xeon
This time the NVIDIA Keynote in its GTC 2020 has not been live, it has been very different from what we are used to. The company has uploaded a series of videos starring Jen-Hsun Huang, CEO of NVIDIA. Nor has it been on a stage with a large screen, it has been from the kitchen of his house and has started it by thanking the doctors for their fight against the Coronavirus.
The presentation has nothing to do with gaming and the mainstream, it has been to present their professional solutions. All of these new solutions are based on the new Ampere architecture, which promises great performance. But the presentation has had important and interesting news.
- Powered by geforce rtx 2060
- Integrated with 6GB GDDR6 192-bit memory interface
- Windforce 2x cooling system with alternate fans
- Unique 90mm blade fans
NVIDIA Ampere brings many new features
Perhaps the most surprising is that DGX solutions and the like with Ampere silicones no longer rely on Intel Xeon processors. After many years the DGX are based on EPYC Rome processors from AMD. So both the Ampere architecture, as well as the processors that add management and power, are based on 7nm lithography. But it is that in addition to the number of cores, power and others, AMD has been chosen for efficiency.
Thus the DGX-100 solution that uses Ampere GA100 7nm GPU it also uses 7nm AMD EPYC Rome processors. A new victory for AMD, which improves the position and prestige of its processors. The most powerful HPC on the market for Deep Learning and Artificial Intelligence is supported by AMD.
DGX-A100 is based on a total of 8 NVIDIA Tesla A100 GPUs that implement next-generation Tensor Cores. This new platform will have 40GB HBM2 per GPU, adding up to a total of 320GB HBM2 in total. The whole set will offer 5 PETAFLOPs of power for Artificial Intelligence and 10 PETAOP (INT8). 6 NVLinks have also been integrated into the system so that the GPUs work as 1 and can share memory.
The success of Zen architecture
NVIDIA has opted for AMD's EPYC 7742 processors that are based on TSMC's 7nm lithography. These processors have 64 cores and 128 threads and two have been mounted in this system, thus adding 128 cores and 256 threads. Each of the processors works at a base frequency of 2.25GHz and reaches 3.4GHz in Boost mode.
1TB DDR4 has been provided for the entire system and up to 15TB in U.2 drives and 2 M.2 NVMe 1.92TB SSDs. Something very interesting is that it has 8 Mellanox single port connections for 200GB / s HDR (Infiniband) and a dual port Mellanox switch.
You can already buy these systems starting at $ 199.999 onwards. But the most relevant thing is that Intel has lost a very important customer that AMD has taken thanks to its excellent processors.
Source: wccftech
