Gaming PC

NVIDIA Drops DRIVE Atlan SoC, Introduces 2 PFLOPS DRIVE Thor for 2025 Autos

NVIDIA is making some amazing changes to its DRIVE automotive SoC plans in a series of announcements announced today as part of its fall GTC 2022 event. NVIDIA immediately canceled Atlan, his post-Orin SoC for 2025 cars. Instead, NVIDIA is announcing his even more powerful SoC, Thor, which will launch in 2025.

NVIDIA’s Atlan SoC was first shown at GTC 2021 in the spring. So he announced NVIDIA as the next generation automotive SoC to replace the (current) Orin SoC. At the time of its announcement, Atlan was to be a high-performance SoC offering 1000 TOPS of his INT8 inference performance using next-generation (Lovelace) GPU design and next-generation Grace CPU design. The chip integrated a BlueField DPU as a networking and security processor with the aim of providing a single he SoC capable of handling all the computational functions required for self-driving cars.

Atlan the Was

But whatever Atlan was supposed to be, now it’s gone. As of NVIDIA’s new DRIVE SoC roadmap, Atlan has been discontinued. In its place is a new SoC, Thor, which is slated to be even more powerful than what his NVIDIA had planned for Atlan.

Similar to Atlan’s announcement in 2021, NVIDIA has only released a handful of details ahead of Thor’s release. The high-level details do not name specific NVIDIA CPU and GPU architectures, but the SoC takes advantage of features first introduced in the Grace CPU, Ampere GPU architecture, and Lovelace GPU architecture. Including. Meanwhile, his NVIDIA blog post on the matter goes even further, stating that the SoC uses his Automotive Enhanced (AE) version of Arm’s top-secret Poseidon CPU core. Very little we know about Poseidon but it’s the next generation of high performance CPU cores that Arm is developing and will replace his just released Neoverse V2 for the next generation of his Neoverse V series used on his platform.

NVIDIA ARM SoC spec comparison
Thor Atlan orin
CPU core Arm “Poseidon AE” Grace Next 12x Arm CA-78AE “Hercules”
GPU core you bet Ampere-Next iGPU
Ampere iGPU
(2048 CUDA cores
64 Tensor Cores)
FP8/INT8 DL OPS 2000 TOPS (FP8) 1000 TOPS (INT8) 275 Tops (INT8)
manufacturing process ? ? samsung 8nm
transistor 77B ? 17B
TDP ? ? 60W

From a performance perspective, Thor New standardized FP8 data formatThat’s not the same as Atlan’s 1000 TFLOPS INT8 figure, but it still represents twice the throughput of 8-bit precision computing. The SoC’s tensor cores will also feature his NVIDIA’s Transformer Engine, allowing the SoC to further speed up the Transformer network.

Especially when you pack all this performance into it, the Thor becomes a very bulky chip. NVIDIA hasn’t announced a process node, but already he says it will use 77 billion transistors. That’s 3 billion fewer transistors than the new flagship GH100 GPU. NVIDIA’s performance claims don’t indicate if matrix sparsity is used, but even if it is, Thor gets half of his FP8 performance of his NVIDIA’s flagship GPU. All of this highlights his NVIDIA’s ultimate performance target for his planned SoC.

NVIDIA’s mockup of the chip shows it on an AGX board in a single-chip configuration, but today’s announcement also explicitly mentions NVLink chip-to-chip (NVLink-C2C) chip interconnect technology. . This is an interesting mention, as NVIDIA’s key art doesn’t indicate that Thor is chiplet-based. This could mean that NVIDIA will instead use his NVLink-C2C for its even more powerful multi-chip DRIVE AGX boards (ala Pegasus). Or that Thor is a chiplet-based design and NVIDIA intentionally ditched their art.

Beyond that, NVIDIA has not provided any further technical details about the SoC. Therefore, details of the type of memory used, GPU architecture, and other functional blocks are not yet known.

At this point, NVIDIA has also not elaborated on why it canceled Atlan instead of Thor. Thor is certainly a stronger design and seems to incorporate some new features not found in Atlan (or at least unpublished). It remains to be seen if this means that his NVIDIA is somehow embracing post-Atlan chips, or if customers ditched Atlan because they needed better AI inference performance for self-driving cars. .

Hardware upgrades aside, it’s clear that NVIDIA is designing Thor for the same market segment as Atlan. In other words, it is intended to be a high-performance single-chip design to handle all the computational needs of autonomous vehicles, from infotainment systems and sensor fusion to the actual autonomous driving algorithms themselves. As with Atlan, the goal is to replace the separate computers currently in the car with a single computer that does everything, with extensive isolation (including MIG) so that separate tasks do not interfere with each other. It is to utilize the functional safety design method.

But perhaps the most surprising thing is that this SoC change is not expected to impact NVIDIA’s SoC delivery times. NVIDIA says it will put the tally of 2025 cars in the hands of automakers. This is the same time frame that Atlan was supposed to appear. So the devil is in the details, but at a high level NVIDIA aims to deliver Thor in much the same way. Time they would have delivered to Atlan. However, while NVIDIA had previously announced that Atlan would be sampling his 2023, it should be noted that no such announcement has been made for Thor. As such, Thor’s sampling date may be later than Atlan’s.

Related Articles

Leave a Reply

Your email address will not be published.

Back to top button