Higher transfer speed is also required as well as greater efficiency, which should result in a greater number of interactions in less time, optimizing the internal traffic of the servers.
The first thing we must be clear about is that this platform is based on an end-to-end network and is the most powerful ever created. For this, it stockpiles the Quantum-2 switch, a ConnectX-7 network adapter and a BlueField-3 DPU, all managed by the company’s software that is apparently included in the package.
What is sought here is so-called “predictive performance” as NVIDIA ensures proactive monitoring, congestion management to provide isolation from generated traffic and almost completely eliminating fluctuations in Quantum-2 performance.
For this, it collects a BlueField-3 DPU that achieves the best performance ever achieved in a component of this caliber and an HCA ConnectX-and compatible with PCIe 4.0 and 5.0 with single or dual ports at a speed of 400 Gbps, in addition to a better data protection with user isolation management.
The configuration has up to 64 400 Gbps ports or 128 200 GB / s ports, all on physical 32 octal OSFP connectors. The great advantage of this is that, being modular, you can choose between three different total configurations:
- 2,048 400 Gb / s ports or 4,096 200 Gb / s ports
- 1,024 400 Gb / s ports or 2,048 200 Gb / s ports
- 512 400 Gb / s ports or 1,024 200 Gb / s ports
The result of all this is a bi-directional performance that the company claims to reach 1.64 petabits per second, or what is equal, 5 times more than the Quantum InfiniBand of the generation that precedes it. All this in addition to reducing energy consumption by 7% and the physical space in data centers and the need for fabric switches by up to 6 times.
This is possible due to the TSMC node 7N, which has led to Quantum 2 having 57 billion transistors, being the chip with the highest number of them in NVIDIA’s history and logically surpassing the A100 by 3 billion.
BlueField 3 InfiniBand goes up to 22 billion transistors and offers 16 ARM cores 64-bit also with the 7N under its belt.
ConnectX-7 integrates 8 billion transistors in the same node TSMC 7N and integrates 16 cores and 256 Threads, which multiplies its performance by 4 and its Throughput by 2 if we compare it with a current GPU.
It also doubles the performance of RDMA, GPUDirect Storage, GPUDirect RDMA and In-Networking Computing, arriving in January next year to complete the NVIDIA platform.
NVIDIA DRIVE ORIN and Jetson AGX ORIN
Changing a bit of third we have the new Orin SoCs that will try to redefine robotics and the development ecosystems of the industry, where the automotive and energy sector in general are integrated of course.
What NVIDIA has shown is an approach and ease for a platform that diversifies and that has to reach most of the current and future industry, so the applications are incalculable and scalable. To position ourselves, Orin is Xavier’s successor and will arrive with two very clear and differentiated bets: DRIVE Orin and JETSON AGX ORIN for next-generation AI in general robotics.
NVIDIA DRIVE Orin
It is the second generation integrated supercomputing platform within DRIVE as such and its role has been very clearly defined:
- Record the data from the car cameras.
- Record radar and sensor data.
- Locate the car and its location.
- Plan and monitor driving and the road.
This specific SoC for the automotive industry is a chip that offers 254 TOPS and with that it is more than capable of autonomous driving, next-generation infotainment and interaction with AI for driver and passengers.
We are talking about a chip that is capable of driving 100% autonomous driving of any vehicle that integrates it up to level 5, the highest in the industry. As if that were not enough, DRIVE Orin accepts multiple complex applications through different operating systems, including Linux, QNX and Android, for example.
Focused on AI like its predecessor and beyond, this is how you can define that 100mm x 87mm SoC that is the Jetson AGX Orin, which according to NVIDIA offers up to 6 times the performance of its predecessor Xavier in the same compact format.
This SoC is specifically created to be housed in major delivery robots, logistics or autonomous manufacturing systems, as well as even large air vehicles. Get an AI performance of 200 TOPS in INT8 and has to its credit an Ampere GPU with 2048 Shaders and 64 Tensor Cores at a maximum frequency of 1 GHz.
Instead the CPU relies on nothing less than 12 ARM A78AE V8.2 64 Bit Cores with 3MB of L2 and 6MB of L3 with a maximum frequency of 2 GHz. The SoC, if already spectacular, adds to its credit 32GB of LPDDR5 with a bus of 256 bits to give a bandwidth of 204.8 GB / s, which allows you to control up to 6 cameras with 16 channels in total distributed with the following configurations:
- 2 x 4K at 60 FPS.
- 4 x 4K at 30 FPS.
- 8 x 1080p at 60 FPS.
- 16 x 1080p at 30 FPS.
All in H.265 and being cameras for encode, while for decode the configuration is different:
- 1 x 8K at 30 FPS.
- 3 x 4K at 60 FPS.
- 6 x 4K at 30 FPS.
- 12 x 1080p at 60 FPS.
- 24 x 1080p at 30 FPS.
If this was not already impressive in itself, the screen resolution that it will be able to handle in the new cars that integrate it is amazing: 1 x 8K at 60 FPS with multi-mode DP 1.4a or HDMI 2.1, without counting the multiple I / O that may have:
- 4x USB 2.0
- 4x UART
- 3x SPI
- 4x I2S
- 8x I2C
- 2x CAN
- DMIC & DSPK
Last and not least we have the configurations that can be selected according to their TDP, and that is that the manufacturers will have from the 15 watts, going through the 30 watts until the 50W. Jetson AGX Orin will arrive in the Q1 2022 at a price still unknown.