Write new article from this text only in smaller paragraphs
The chip manufacturer Nvidia continues to rely on the driving force AI hardware, which is currently bringing the company billions in profits. At the keynote speech at the Computex IT trade fair, Nvdia CEO Jensen Huang presented the DGX GH200 server system, in which 256 Grace Hopper combination accelerators calculate. These are connected via NVLink and should offer a computing power of one exaflops (1 trillion operations per second). The first DGX-GH200 systems should be available by the end of the year. (Picture: Nvidia) However, this specification refers to the 8-bit data type FP8 used for AI calculations and is therefore not comparable with the performance specifications of supercomputers, which are determined with double-precision floating-point numbers (FP64). There, a DGX-GH200 system achieves around 17 petaflops (0.017 exaflops). Launch customers Google, Meta and Microsoft The Grace GH200 hopper chips used consist of a Grace CPU with 72 Neoverse V2 ARM cores and the H100 GPU with Tensor cores, which are housed in one package. They communicate with each other via NVLINK at 900 GB/s, as well as with the server’s other GH200 chips. In total, the 256 combination accelerators can access 144 TB of shared memory. With Grace Hopper, the ARM processor and H100 GPU (on the right with HBM memory) share a common package. (Picture: Nvidia) This is why DGX GH200 are particularly suitable for training large models for generative AI applications. These include language models such as ChatGPT or image generators. Unsurprisingly, the initial customers are Google, Meta and Microsoft. But Nvidia itself also wants to build an AI supercomputer called Helios, which consists of four DGX-GH200 servers. MGX open server architecture For smaller servers, Nvidia has also introduced the MGX open server architecture. Server manufacturers can combine Nvidia’s complete GPU portfolio with Grace CPUs, GH200 or x86 processors in one, two or four height units. The chip manufacturer cooperates with Asrock Rack, Asus, Gigabyte, Pegatron, QCT and Supermicro. The MGX server architecture, optimized for GPU and computing accelerator, allows various designs. (Picture: Nvidia) (chh) Home
