As part of the fall GTC 2022 event, NVIDIA announced a change in plans for its DRIVE series of automotive systems-on-a-chip. The company has canceled the release of Atlan SoC, which was supposed to be released after Orin for 2025 cars. Instead, NVIDIA announced a more powerful version of the SoC Thor, which is scheduled to launch in the same period in 2025.
The Atlan chip was unveiled at the GTC 2021 spring event when NVIDIA announced its next-generation automotive SoC. At the time of the announcement, Atlan was planned to be a high performance SoC offering 1,000 teraflops of performance in INT8 computing. The chip combined the next generation GPU (Lovelace) and the next generation Grace CPU. The chip even integrated a BlueField DPU as a network and security processor. Atlan was planned to be a single system-on-a-chip that could handle all the computational functions needed for an unmanned vehicle. Now this mission is entrusted to Thor. One chip will be able to perform all the functions that previously required the use of several different processors.
However, in the new NVIDIA DRIVE roadmap, the Atlan chip has been dropped, and a new version of Thor SoC has come in its place, which should be even more powerful. So far, the company is not naming the CPU and GPU architectures used, although it does say the SoC uses features that were first introduced in the Grace CPU, Ampere and Lovelace GPU architectures. However, SoC Thor uses the Automotive Enhanced (AE) version of Arm’s yet-to-be-announced Poseidon CPU core. So far, the only thing known about Poseidon is that it is a high-performance computing core being developed by Arm, which is planned to be used in the next generation Neoverse V platform.
The Thor chip is expected to deliver 2 petaflops (2000 teraflops) floating point performance using the new FP8 data format. SoC tensor cores will be equipped with NVIDIA converters, which will allow the chip to further accelerate the processing of related tasks.
All these high performance components will make Thor a very bulky chip. So far, NVIDIA has not specified the manufacturing process, but has already said that the chip will have 77 billion transistors. This is only 3 billion transistors less than the new flagship GH100 GPU.
The use of NVLink chip-to-chip (NVLink-C2C) inter-chip connection is also mentioned. Although Thor’s image does not indicate that the chip is made of chiplets. This may indicate that NVLink-C2C can be used to create even more powerful multi-chip boards (or NVIDIA simply did not specify the chiplet configuration for Thor).
NVIDIA hasn’t given a timeline for Thor samples, but says the chips will be in the hands of automakers in time to integrate them into 2025 lineup vehicles.
Source: anandtech