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

Among the many spate of bulletins from NVIDIA at the moment as a part of their fall GTC 2022 occasion, the corporate is delivering a shocking shake-up to their DRIVE automotive SoC plans. Efficient instantly, NVIDIA has cancelled Atlan, their deliberate post-Orin SoC for 2025 vehicles. As a replacement, NVIDIA is saying Thor, an much more highly effective SoC set to launch in the identical 2025 timeframe.

NVIDIA’s Atlan SoC was first revealed again at spring GTC 2021, the place NVIDIA introduced it as their next-generation automotive SoC to succeed the (now present) Orin SoC. On the time of its announcement, Atlan was deliberate to be a high-performance SoC providing 1000 TOPS of INT8 inference efficiency, utilizing a next-generation (Lovelace) GPU design and next-generation Grace CPU design. The chip even built-in a BlueField DPU as a networking and safety processor, with the intent of delivering a single SoC that would deal with all the compute capabilities required by a self-driving automotive.

Atlan-that-was

However no matter Atlan was meant to be, now it’s no extra. As of NVIDIA new DRIVE SoC roadmap, Atlan has been scrapped. As a replacement is a brand new SoC, Thor, which is slated to be much more highly effective than what NVIDIA had deliberate for Atlan.

Just like the Atlan announcement in 2021, NVIDIA is just releasing a handful of particulars about Thor this far forward of its launch. The high-level particulars embrace that, whereas not naming particular NVIDIA CPU and GPU architectures, the SoC is tapping options that had been first launched with the Grace CPU, Ampere GPU structure, and Lovelace GPU structure. In the meantime, NVIDIA’s weblog publish on the matter does go one additional, stating that the SoC makes use of an Automotive Enhanced (AE) model of Arm’s so far secretive Poseidon CPU core. Of what little we learn about Poseidon, it’s a next-generation high-performance CPU core being developed by Arm that might be used of their next-generation Neoverse V-series platform, changing the just-released Neoverse V2.

From a efficiency standpoint, Thor is slated to ship 2 PFLOPS (2000 TFLOPS) of floating level inference efficiency utilizing the newly-standardized FP8 knowledge format. Which, though not apples-to-apples with Atlan’s 1000 TFLOPS INT8 determine, nonetheless represents twice the throughput for 8-bit precision computing. The SoC’s tensor cores can even characteristic NVIDIA’s transformer engines, permitting the SoC to additional speed up the processing of transformer networks.

Notably, packing in all of this efficiency goes to make Thor a really cumbersome chip. Whereas NVIDIA shouldn’t be saying the method node, they’re already saying that it’ll use 77 billion transistors, which is all of three billion transistors lower than their new flagship GH100 GPU. NVIDIA’s efficiency claims don’t point out whether or not matrix sparsity is in use, however even whether it is that can put Thor at half the FP8 efficiency as NVIDIA’s flagship GPU. All of which underscores NVIDIA’s excessive efficiency objectives for the deliberate SoC.

And whereas NVIDIA’s mockup for the chip reveals it on an AGX board in a single chip configuration, at the moment’s announcement additionally makes express point out of NVLink chip-to-chip (NVLink-C2C) chip interconnect know-how. This can be a curious point out, since NVIDIA’s key artwork doesn’t present Thor as being chiplet-based. This will likely imply that NVIDIA is as a substitute going to make use of NVLink-C2C for much more highly effective multi-chip DRIVE AGX boards (ala Pegasus), or it may very properly be that Thor is a chiplet-based design and NVIDIA is being purposely generic with their artwork.

Past that, NVIDIA shouldn’t be provided any additional technical particulars on the SoC. So particulars on the reminiscence kind used, the GPU structure, and different useful blocks stays to be seen.

At this level, NVIDIA has additionally not detailed why they cancelled Atlan instead of Thor. Thor is actually a extra highly effective design and would appear to include some newer options not discovered on (or no less than by no means disclosed for) Atlan. Whether or not this implies NVIDIA is in some way pulling in what would have been a post-Atlan chip or if they’ve tossed Atlan as a result of their prospects want higher AI inference efficiency for his or her self-driving automobiles stays to be seen.

{Hardware} upgrades apart, what is obvious is that NVIDIA is designing Thor for a similar market phase as Atlan. Which is to say it’s meant to be a high-performance single-chip design for dealing with all the computational wants of a self-driving automotive, from infotainment methods and sensor fusion as much as the precise self-driving algorithms themselves. As with Atlan, the purpose is to switch what are presently separate computer systems inside a automotive with a single pc that does all of it, leveraging the useful security design strategies with intensive isolation (together with MIG) to maintain the separate duties from interfering with one another.

Maybe essentially the most shocking factor of all, nonetheless, is that this variation in SoCs shouldn’t be anticipated to impression NVIDIA’s SoC supply date. NVIDIA says that they’ll have Thor in automakers’ palms for his or her 2025 automobiles, the identical timeframe that Atlan was supposed to reach in. So whereas the satan is within the particulars, at a excessive stage NVIDIA is aiming to ship Thor close to the identical time as they’d have delivered Atlan. It’s value noting, nonetheless, that whereas NVIDIA had beforehand introduced that Atlan can be sampling in 2023, no such announcement has been made about Thor. As such, Thor’s sampling date could find yourself being later than what would have been Atlan’s.