NVIDIA’s Next-Gen GPU Architecture Powers The DRIVE AGX Orin – 17 Billion Transistors, 12 ARM Hercules CPU Cores and 200 TOPs
NVIDIA has just announced their latest DRIVE AGX Orin platform which is powered by the Orin SOC, a chip featuring NVIDIA's next-generation GPU architecture, ARM Hercules CPU cores, and lots of AI performance. The platform is aimed at autonomous vehicles and robots and should deliver much better performance and higher efficiency than the DRIVE Pegasus boards.
NVIDIA's Next-Generation GPU Architecture & ARM Hercules Cores Power The Orin SOC Featured on DRIVE AGX Orin
The platform consists of the next-generation Orin SOC which replaces Drive PX Pegasus. The Drive PX Pegasus board was comprised of multiple Xavier SOCs and Turing GPUs and featured 320 TOPs at 500W. The singular Orin SOC aims to deliver 200 TOPs of performance on a single chip, featuring a massive 17 billion transistors and delivering roughly 7 times the performance of NVIDIA's Xavier SOCs. For comparison, the biggest consumer-grade GPUs that are currently in production, the Volta GV100, and the Turing TU102 have transistor counts of 21.1 billion and 18.6 billion, respectively.
NVIDIA hasn't specifically stated which next-generation GPU they are using on the Orin SOC but it should be noted that NVIDIA's Ampere GPUs are slated for launch in the coming year, featuring a sub-12nm process node and offering a much bigger architectural upgrade than the one we saw from Volta to Turing. The Xavier SOC was also the first product announced which publicly mentioned the Volta GPU so it's no surprise that its successor, Orin, would be the first product announced with the next-gen Ampere GPU.
UPDATE: NVIDIA has confirmed to us that the Orin SOC utilizes an 8nm process node that aligns with Samsungs 8LPP node. This also confirms what Jensen Huang, CEO of NVIDIA, stated, that the majority of NVIDIA orders would be handled by TSMC while a small portion, such as Orin, would be handled by Samsung.
Xavier featured a total of 7 billion transistors on a 16nm die along with 8 custom ARM64 cores. The Orin SOC is a much denser chip than Xavier which confirms that it isn't just a 12nm shrink, but a 7nm (EUV) part, which allows for such a drastic uplift in its transistor density. ARM has also shown in their roadmap that Hercules CPUs can be made on either 7nm or 5nm process, giving NVIDIA the flexibility needed to build the Orin SOC. The Hercules cores are also stated to be 10% faster than the Deimos 7nm cores.
“Creating a safe autonomous vehicle is perhaps society’s greatest computing challenge,” said Jensen Huang, founder and CEO of NVIDIA. “The amount of investment required to deliver autonomous vehicles has grown exponentially, and the complexity of the task requires a scalable, programmable, software-defined AI platform like Orin.”
“NVIDIA’s long-term commitment to the transportation industry, along with its innovative end-to-end platform and tools, has resulted in a vast ecosystem — virtually every company working on AVs is utilizing NVIDIA in its compute stack,” said Sam Abuelsamid, principal research analyst at Navigant Research. “Orin looks to be a significant step forward that should help enable the next great chapter in this ever improving technology story.”
In a slide shared by Dylan522P over at his Twitter feed, various DRIVE configurations for the Orin SOC are listed. It seems like Orin will have a few TDP / workload-optimized variants with the base 1-Camera variant offering 36 TOPs at 15W, the 4-Camera variant offering 100 TOPs at 40W, a 2 chip variant offering 400 TOPs at 130W and the top-end 2 Orin + 2 discrete GPU variant offering up to 2000 TOPs at 750W. It looks like the TDP of the 200 TOPs, single-SOC lies somewhere around 60-70 Watts.
A next-generation GPU, seemingly the same iteration featured on the Orin SOC, was teased at the GTC 2019 event, featuring around 800 TOPs at 310W. This is taken from the fact that two Orin chips offer 400 TOPs at 130W which leaves the top DRIVE SKU (also featuring 2 Orin chips) with a spare 620 Watts. If each discrete GPU has a TDP of 310W, this would mean that they are delivering a total theoretical INT8 output of 800 TOPs each. Currently, the Quadro RTX 8000 which features the top-of-the-line TU102 GPU, has a total computational power of 206.1 TOPs at 295W. This is just about a 3.5x increase in terms of performance over the Turing GPU architecture which showcases just what to expect from NVIDIA's next-gen GPU.
NVIDIA states that the DRIVE AGX Orin is compatible and scalable to Level 2 and up to Level 5 self-driving vehicles. It is also designed to handle numerous applications and DNN / AI workloads while achieving systematic safety standards such as ISO 26262 ASIL-D. Just like Xavier before it, the Orin SOC is scheduled to be featured in vehicles going in production in 2022 which is two years after its announcement. This means that the production of Orin should commence earlier than that and very likely in 2021.
NVIDIA Drive PX Generation Comparison:
|Product Name||NVIDIA Drive PX||NVIDIA Drive PX 2||NVIDIA Drive Xavier||NVIDIA Drive Pegasus||NVIDIA Drive AGX Orin|
|SOC Name||Tegra X1||Parker||Xavier||Xavier||Orin|
|Process Technology||20nm SOC||16nm FinFET||12nm FinFET||12nm FinFET||TBA|
|SOC Transistors||2 Billion (Tegra X1)||N/A||7 Billion (Xavier)||7 Billion (Xavier)||17 Billion (Orin)|
|GPU Architecture||Maxwell (256 Core)||Pascal (256 Core)||Volta (512 Core)||Volta (512 Core)||Ampere?|
|CPU||16 Core ARM CPU||12 Core ARM CPU||8 Core ARM CPU||16 Core ARM CPU||12 Core ARM CPU|
|CPU Architecture||8x Cortex A57|
8x Cortex A53
8x Cortex A57
|Carmel ARM64 8 Core CPU (8 MB L2 + 4 MB L3)||Carmel ARM64 8 Core CPU (8 MB L2 + 4 MB L3)||ARM Herclues Cores|
|Compute DLTOPs||N/A||20 DLTOPs||30 TOPs||320 TOPs||200 TOPs|
|Total Chips||2 x Tegra X1||2 x Tegra X2|
2 x Pascal MXM GPUs
|1 x Xavier||2 x Volta|
2 x Turing
|1 x Ampere|
|System Memory||LPDDR4||8 GB LPDDR4 (50+ GB/s)||16 GB 256-bit LPDDR4||LPDDR4 + GDDR6||N/A|
|Graphics Memory||N/A||4 GB GDDR5 (80+ GB/s)||137 GB/s||1 TB/s||200 GB/s|
Xavier was also announced back in 2016 but featured in products available by end of 2017 and the beginning of 2018. With that said, the GPUs powered by Volta were also available a year later (Titan V) so that might give us a hint at the launch of GPUs powered by NVIDIA's next-generation GPU architecture which is Ampere. With CES 2020 closing in, we will get to hear more details from NVIDIA about their future GPU architecture and next-generation product roadmap.