Nvidia is not a beginner in the field of processors, it announced the first processor product in 2008, which started the Tegra track. However, with the products of this SoC series intended primarily for mobile use, it did not manage to break through significantly – perhaps with the only exception, which is the Tegra X1 from 2015, on which stands the Nintendo Switch game console. Other models – whether designed for smartphones, tablets, laptops or the automotive industry – were not so lucky. This time, Nvidia comes up with one major change: For the first time in its processor history, it will not be a SoC, but (i) a stand-alone processor.
Nvidia calls her Grace and his (or her) first implementation targets HPC / AI systems, where it will act as a chassis for the same company’s computing GPUs. Nvidia hasn’t revealed much yet. We only know that the product is targeting 2023, will use LPDDR5x memory with a data throughput of over 500 GB / s, the connection between CPUs will reach> 600 GB / s and the connection between CPU and GPU will provide an interface with a transfer capacity of more than 900 GB / s promises scores over 300 in SPECrate_2017_int_base thanks to ARM generation cores Neoverse.
These numbers are of particular interest in the transfer capacity between CPU and memory, because processors, unlike graphics chips, do not benefit as much from the volume of data transfers as from latencies, so tens of GB / s and a server segment lower than hundreds of GB / s are usually enough. Grace thus practically doubling the transmission capacity with memory compared to current products.
According to VideoCardz, Nvidia also plans to report SoC Atlan combining processor Grace and the graphics cores of any of the successors of the generation Ampere.
Naming Grace could be related to the naming of the graphic generation Hopper.
The US Navy Admiral began programming as early as Mark I, and later with UNIVAC. She declined a professorship to spend the next 30 years at the forefront of the IT world. She was able to turn her idea that programming should be as easy as English into the creation of the FLOW-MATIC programming language, from which the well-known COBOL later emerged. After leaving the Navy, she worked as a consultant for DEC until her death at the age of 85.
Graphic generation Hopper was originally expected as a successor Ampere, but before Christmas there was information that it was to be preceded by a generation Lovelace. The time of their release is indirectly indicated by the published roadmap of Nvidia, albeit from the point of view of datacenters:
Ampere-next in this context there would be a generation Lovelace. The position in 2022 practically refutes its release at the end of this year, which some sources claimed. Nvidia is traditionally the first to start producing computing GPUs, and only with a certain delay (1-2 quarters) do GPUs start playing. However, the start of production of computational GPU generation is not excluded Lovelace at the end of this year with the release at the beginning of next year. This could mean, among other things, the release of a game series Lovelace around the middle of 2022.
Ampere-next-next painted for 2024 could then be a generation Hopper, which is expected from the end of last year in the period 2023-2024. Chiplets are expected to be deployed, but more is not known, and Nvidia, as can be seen, no longer reveals the names of generations – only the dates when it will release “something”.
Furthermore, the roadmap announces the timing of the next generations of DPU Bluefield. The Data Processing Unit, as Nvidia refers to this line of chips, is a legacy of the absorbed Mellanox. The goal of these chips integrating ARM cores is not to solve demanding tasks, but to take on less demanding tasks, which then do not have to deal with the main processor (s). These include software security tasks, software-defined networks and storage, or the presence of specialized accelerators. The third generation of Nvidia is expected in 2022, the fourth at about the turn of 2023/2024.