This morning, NVIDIA also held a press conference at the Taipei Computer Show. Naturally, there is no RTX 40 Series graphics card. It mainly talks about the progress in the data center. The self-developed ARM processor grace released at the GTC conference in March has new content, and more than a dozen partners such as ASUS, Gigabyte and ultramicro will launch grace server products next year.
Grace is an arm super processor developed by NVIDIA. It is based on arm neorse and has two versions. One is CPU + GPU, which integrates 72 core grace processor and hopper architecture GPU, and the other is CPU + CPU dual architecture, which has up to 144 CPU cores.
Based on these two different processors, servers also have different configurations, HGX grace with dual CPUs can integrate 1TB lpddr5x memory, 1TB / s bandwidth, TDP power consumption of 500W, air-cooled or water-cooled design, and each rack can accommodate 84 nodes**
The number of HGX grace hopper cores of CPU + GPU is halved, and the memory of lpddr5x is only 512gb. However, with the addition of 80GB hbm3 memory of graphics card, the total bandwidth can reach 3.5tb/s. The price is that the power consumption is 1000W, and each rack can accommodate 42 nodes.
NVIDIA does not exclude x86 from its friends in CPU selection. AMD's second-generation epyc 7742 processor is even selected in the previous generation products. After all, 64 core is very good and powerful, but now that it has its own grace CPU, it naturally has to show its advantages. It can be said that it punches Intel and kicks AMD.
According to the information of NVIDIA on different occasions, compared with Intel's 10nm ice lake server chip, grace super chip has twice the performance and 2.3 times the energy efficiency advantage**
NVIDIA is not soft on AMD. Compared with the processing of epyc 7742, grace super chip is in specrate_ 2017_ int_ Base performance test is 1.5 times higher, leading all kinds of.
