NVIDIA's newest AI chip is the NVIDIA Grace Hopper Superchip, which is based on the NVIDIA Hopper architecture. It is a 500-trillion-transistor chip that can deliver up to 1.5 exaflops of performance. It is the world's first AI chip to use a unified memory architecture, allowing it to access CPU and GPU memory simultaneously. This makes it ideal for AI applications that require large amounts of memory, such as natural language processing and machine learning.
The Grace Hopper Superchip is also the first AI chip to use a new type of interconnect called NVLink. NVLink can transfer data at up to 100GB/s. This makes it possible to connect multiple Grace Hopper Superchips together to create even more powerful AI systems.
The release is expected to be available in 2024.
The NVIDIA Grace Hopper Superchip is significantly faster than the NVIDIA H100 GPU. For example, in the GPT-J inference benchmark, the Grace Hopper Superchip delivers up to 17% more performance.
In comparison, the IBM Northstar is a 170-billion-transistor chip that can deliver up to 150 gigaflops of performance. It is based on the IBM POWER9 architecture, which is a traditional CPU architecture that has been optimized for AI workloads. The POWER9 architecture includes a number of features that make it well-suited for AI workloads, such as a large number of cores and a high memory bandwidth.
Here is a table that summarizes the key features of the NVIDIA Grace Hopper Superchip and the IBM Northstar:
As you can see, the Grace Hopper Superchip and the Northstar are both powerful AI chips that are designed for different workloads. The Grace Hopper Superchip is better suited for AI applications that require large amounts of memory, while the Northstar is designed for AI applications that require real-time performance, such as autonomous driving and medical imaging.