New AI  and Security features 

AI techniques, especially those related to Machine Learning (ML) and Deep Learning are increasingly useful for scientific computing as well.  The H100 new Transformer Engine can accelerate Transformer model training and inference by dynamically choosing between FP8 and 16-bit calculations, which may deliver up to 9x faster AI training and up to 30x faster AI inference speedups on large language models compared  to A100.
H100 also provides some features that enable safer multi-user environments, especially important in virtualized environments. The  new  Confidential Computing capability with MIG-level Trusted Execution Environments (TEE) supports up to seven individual GPU Instances, each with dedicated NVDEC and NVJPG units. Each Instance now includes its own set of performance monitors that work with Nvidia developer tools.  The H100 extends the Trusted Execution Environment with CPUs at full PCIe line rate.

ARM and Grace

Announced for 2023, Nvidia's ARM-based Grace CPU seems also noteworthy. USA's Los Alamos National Laboratory and the Swiss National Computing Centre have already announced plans for Grace-based supercomputers \cite{pressure}. ARM, originally a UK IP-only company, is now owned by the Japanese SoftBank Group Corp (SBC). Nvidia also tried buying ARM from them in a Fall 2020 agreement, but the deal was terminated in Feb. 2022 due to "regulatory challenges" in both US and China.
Intel still dominates the top500.org list of the world's largest supercomputers, but in 2021 Japan’s ARM-based Fugaku supercomputer (> seven million cores, running at 442 petaFLOPS) took the top spot.
The Grace CPU Superchip https://www.nvidia.com/en-us/data-center/grace-cpu/\cite{intro} features two Grace cores connected via  the NVLink-C2C technology thus providing up to 144 Arm v9 CPU cores. It claims to be the world’s first CPU using LPDDR5x memory with ECC and 1TB/s total bandwidth. Its 900 GB/s coherent interface is 7x faster than PCIe Gen 5.
Nvidia's Grace Hopper Superchip combines the Grace CPU and Hopper GPU architectures using Nvidia's NVLink-C2C technology to deliver a coherent CPU+GPU memory model. 
The system targets both HPC and  AI applications and can provide a 30x higher aggregate system memory bandwidth to GPU compared to the DGX A100. Both the Grace and Grace Hopper superchips will run Nvidia's software stacks,  including NVIDIA HPC, NVIDIA AI, and NVIDIA Omniverse™ .