The Heart of Machine Report
Editor: Zenan, Du Wei
The highest-end RTX4090 has the highest cost-effectiveness, did you not expect it? Although
was criticized for being too expensive, when it comes to the strongest AI chips and high-end gaming GPUs, people still have to look at Nvidia first. On the evening of September 20, the GTC conference was held online, and the highly anticipated RTX 40 series graphics cards were finally officially released.
is slightly different from the past. This important Keynote seems simple and "short". Huang Renxun stands directly in the empty metaverse and starts speaking:
One and a half hours to finish work.
In this event, Nvidia showed the latest progress in RTX, AI chips and metaverse product line Omniverse, and also included their help to new breakthroughs in artificial intelligence field, as well as the implementation of a large number of applications.
Before everything started, Lao Huang first showed a fully interactive simulation environment, RacerX, which is built with Omniverse, with many physical material characteristics, ray tracing, smoke and flames. The most important thing is that "all have not been pre-rendered, they are running on a single-block GPU":
Huang Renxun said that all processing in real time is what the future game should look like.
can afford RacerX. The latest RTX 40 series graphics card, and AMD is also about to release a new generation of GPUs. Will Nvidia stay ahead this time? If the N card has stronger performance, what is the cost?
RTX 40 series GPU: doubled performance, doubled ray tracing capability
This is Nvidia's third generation RTX graphics card, adopts the new Ada Lovelace architecture .
After Nvidia switched to TSMC, the RTX40 series GPU that was added to the customized 4N process was stuffed into 76 billion transistors , 70% more than the previous generation Ampere. 5 years ago, Nvidia launched a programmable coloring GPU, which completely changed the world of computer 3D graphics. By 2018, Nvidia launched the RTX architecture, with the newly added RT Core used to accelerate real-time ray tracing, and Tensor Core used to process matrix operations, achieving unprecedented results. In the Ada Lovelace architecture just launched, all three processors have been improved and improved, including:
, and SM stream processors have added Shader Reordering and Execution Recording, which can reschedule tasks in real time, increase ray tracing speed by 2-3 times , and output 90TFLOPS on 4090, which has twice the performance compared to the previous generation.
, the third-generation RT Core achieves twice the intercourse performance between light and triangles. The new Opacity Micromap engine can increase the Alpha-test geometry performance by 2 times. The Micro-Mesh engine can improve the richness of geometry without causing more consumption of BVH construction and storage resources.
. The new fourth-generation Tensor Core can realize .4 petaFLOPs computing power, and AI performance doubles .
"Shader Execution Recording is like the out-of-order execution of the CPU back then," said Huang Renxun. "Ray tracing is difficult to process in parallel, and GPU has a high parallelism. SER can improve efficiency by rescheduling the shader load in real time, which can improve ray tracing performance by 2 to 3 times, and game performance is improved by 25%. "
But we know that ray tracing was once said by David Krik, former chief scientist of Nvidia, as a technology that "never come". Double performance cannot ensure that GPUs can maintain high frame rate in various large stand-alone games today. At this time, AI algorithms are needed.
DLSS Using the convolutional autoencoder AI model, high resolution can be automatically calculated based on the low-resolution picture output by the GPU, greatly reducing performance requirements. Nvidia introduced DLSS 3 in the Ada architecture, which can automatically fill frames while improving resolution.DLSS 3 contains four components: a new optical flow accelerator, game engine motion vector, convolutional autoencoding AI frame generator, and Reflex low-latency pipeline.
DLSS 3 processes the current frame and the previous frame at the same time. The optical flow accelerator provides the neural network with motion direction and speed information. Combining the motion vector of graphics and pixels, input neural network to generate intermediate frames.
"DLSS 3 generates a brand new frame without involving graphics pipeline processing. Compared with simple rendering, it can improve performance by up to 4 times," said Huang Renxun. "And games with bottlenecks on both CPU and GPU can benefit from it."
Nvidia showed the running effects of games such as Cyberpunk 2077 with DLSS 3, Microsoft Simulation Flying and other games:
years ago, the amount of RTX's data has increased by 16 times. Now some pixels are calculated, and most of them are speculated by AI.
In the demonstration, several games are directly multiplied by two. But it is worth noting that DLSS 3 and new hardware features are highly bound, and users of the 30 and 20 series cannot enjoy such performance.
Nvidia has announced several conventional hardware specifications for 40 series graphics cards: the RTX 4090 uses AD102 GPU, has 16384 CUDA cores and 24GB GDDR6X video memory, with a default TDP of 450W; the RTX 4080 16GB version has 9728 CUDA cores and a TDP of 320W; the RTX 4080 12G version has 7680 CUDA cores and a TDP of 285W. In terms of power consumption, the process efficiency of switching from Samsung 8nm to TSMC 4N has been significantly improved, the performance of the same level has been improved, and the power supply requirements have not been improved. How much has the performance improved compared to the previous generation? Through SER optimization, larger-scale chips, and the chip acceleration frequency increased from 1.7GHz to 2.52GHz, the performance of the RTX 4090 can be doubled compared to the 3090Ti; if you look at ray tracing, the performance will be improved by 4 times. Huang Renxun said that the performance of Ada is twice that of Ampere.
and further down, the RTX 4080 can achieve twice the performance of 3080Ti when turning on DLSS.
Finally, the price is: the RTX 4090 public version is priced at $1599 and will be sold on October 12; the RTX 4080 16GB is priced at $1199, and the 12GB version is priced at $899.
To sum up, the price of 90 has almost not increased, and the price of 80 has increased by $500. For domestic users, the price of the RTX 40 series is like this: starting from 12,999 yuan for 4090 yuan, 9,499 yuan for 4080 (16GB) and 7,199 yuan for 4080 (12GB).
It looks like the price of the non-public version of RTX 4090 will be around 15,000.
However, this generation of graphics cards has something to pay attention to: it seems that the 12G version of 4080 seems to have changed the original 70Ti.
Nvidia Omniverse is connected to the 3D world
In addition to GPU and AI, Nvidia is also the leader of the metaverse. Huang Renxun introduced a series of progress in Omniverse.
Omniverse is a platform for Nvidia to build and run metacosmic applications, and it works at the intersection of digital and real world. Omniverse is also a large 3D database in real time, building a shareable 3D world. Omniverse is also a computing platform where you can write applications running on it, and these applications become a portal to enter the virtual world.
Today, Huang Renxun released a series of major updates about his Omniverse platform, which supports Ada Lovelace GPUs, making a huge leap in ray tracing and large-scale scene performance.
is first a new neural rendering tool based on GAN and diffusion model. OmniGraph is a graphical rendering engine that controls behavior, actions, and actions in a programmatic way.
followed by a major update to Omniverse Physics, which can be used to handle the motion of complex multi-connected component objects.
and the brand new Cloud XR supports the powerful ray tracing function of Ada in VR. There is also the first SimReady material library for data generation and digital twin simulation.
Replicator is one of the most popular Omniverse applications to generate synthetic data to train autonomous cars, robots and various computer vision models. Finally, there is the new Omniverse JT connector, which is a large application that allows Omniverse to be used in industry and manufacturing.
It can be said that Omniverse is an enterprise platform suitable for the entire product life cycle from product design and styling to engineering planning, manufacturing, marketing and operations. Just as the internet connects various websites, Omniverse connects the 3D world.
Huang Renxun showed at the meeting that some companies used Omniverse to create digital twins for factories, logistics warehouses, automated production lines and industrial plants. We can explore some of the following scenarios.
Omniverse computing platform consists of three parts: an RTX computer for creators, designers and engineers; an OVX server, used to host connections to the Nucleus database and run virtual world simulations; NVIDIA GDN, enter the Omniverse portal.
Through GeForce Now, Nvidia has built a global graphics delivery network (i.e. GDN). The network covers 100 regions and provides a responsive, ultra-fast RTX graphic content delivery network (i.e. CDN) that efficiently connects Internet video. NVIDIA GDN can efficiently connect interactive graphics, and combine NVIDIA RTX PC and cloud-based NVIDIA GPU to create a global Omniverse computing platform.
NVIDIA Omniverse Cloud is a software and infrastructure-as-a-service suite for designing, publishing and experiencing meta-universe applications on any device anytime, anywhere. At the meeting, Huang Renxun presented Rimac, a pioneer in supercar and advanced electric vehicle solutions, and how it leverages Omniverse Cloud to enable collaborative workflows for 3D teams and provide users with advanced 3D experiences.
Huang Renxun said that NVIDIA Omniverse Cloud is an IaaS product that can be connected to the cloud, on-premises and run Omniverse applications on a single device. Replicator and Farm can also run on the cloud, where Farm is the expansion engine for rendering farms. Currently, users can use Replicator and Farm containers on AWS.
The new generation of autonomous driving chip Drive Thor
In the field of autonomous driving, car companies all need stronger computing power, and the performance of each generation of NVIDIA products will double.
At present, intelligent machine development has set off a wave of AI, and the participation of deep learning has opened a new door to the improvement of system capabilities. Everything has become completely different from how the software is developed to how it runs. Therefore, it is imperative to create a new generation of processors. Nvidia Xavier is the world's first autonomous super chip designed for deep learning, and has since made a huge leap in processor performance every two years.
At the same time, in order to expand the field of autonomous driving and improve driving safety, sensors are facing a synchronous growth in terms of quantity and resolution. At the same time, more complex AI models were introduced, all of which drove Nvidia to continuously improve performance.
In 2021, Nvidia launched the 1000 TOPS SoC - Atlan. Today, Huang Renxun said its position has been replaced by Thor. Thor has twice the throughput of Atlan and has more than twice the delivery performance of Atlan . Achieving these goals is inseparable from three factors: Grace, Hopper and Ada Lovelace, which provides the amazing Transformer engine, the rapid transformation of ViT, and the multi-instance GPU in Ada helps centralize on-board computing resources, reducing costs by hundreds of dollars.
Nvidia Drive Thor uses many new technologies, which can be configured in multiple modes, using its 2000 TOPS and 2000 TFLOPs all for autonomous driving workflows. It can also be configured as part of the cockpit AI and infotainment and part of the assisted driving. Multi-computing domain isolation in Thor allows concurrent, time-sensitive multi-processes to run without interruption.You can run Linux, QNX, and Android on one computer.
In addition, Thor also concentrates many computing resources to reduce costs and power consumption while achieving a functional leap. Currently, car parking, active safety, driver surveillance, camera mirroring, clustering and infotainment are controlled by different computing devices. In the future, these capabilities will no longer be controlled by separate computing devices, but will be controlled by software that runs on Thor and continues to improve over time.
Thor chip is expected to be launched in 2025.
NVIDIA Drive is an end-to-end platform for the development and deployment of autonomous vehicles. In terms of development, it includes Replicator synthetic data generation, Drive Sim and Drive Map, and in-car AI applications, AI computers and Hyperion autonomous vehicle reference architecture.
NVIDIA Drive has ushered in a series of feature updates, first of which is an AI workflow called Neural Reconstruction Engine, which has become a major feature of Drive Sim. It can build 3D scenes based on recorded sensor data, and enhance them by manually created or AI-generated content after importing the 3D scene into Drive Sim. Additionally, this workflow from video to 3D geometry can run on OVX systems.
Dynamic display from video to 3D workflow. Another important feature of
Drive Sim is hardware in the loop, which means we can run the entire on-board software stack in the AI factory. It can also simulate the interior environment, and future cars will not only have simple dashboards, but also a surround display that combines digital design with physical design, so that automotive engineers, software engineers and electronic engineers can collaborate in Drive Sim while running all the actual computers and software stacks.
Drive Sim has become a virtual design studio.
In addition, Nvidia has also made outstanding progress in developing Drive end-to-end autonomous driving system , such as Replicator synthetic data generation, AI model improvement, Drive Map autonomous driving fleet map construction, urban and highway driving and parking.
Drive Map for autonomous driving fleet map construction.
new microrobot system-level module
Drive Orin is the second generation of autonomous vehicle computing chip launched by Nvidia. It seems to be very successful at present and has been used by more than 40 cars, trucks and driverless taxis . Jetson is Nvidia's robotics with 1 million developers and uses about 6,000 companies.
At today's GTC conference, Huang Renxun announced the launch of a micro robot system-level module chip, Jetson Orin Nano, which is 80 times faster than the previous Jetson Nano. Jetson Orin Nano can run the NVIDIA Isaac robot stack and have a ROS 2 GPU acceleration framework.
Huang Renxun also introduced its edge AI platform Metropolis, which can interpret data from cameras, lidar and other IoT sensor , improving the security and efficiency of warehouses, factories, retail stores and cities.
From industry to scientific research, autonomous driving to the metaverse, Nvidia's business has long expanded from GPU to countless fields, and has a leading position in many places. For ordinary consumers, graphics cards are no longer used to play games.
Now, the new generation of GPUs has been launched. Will you choose the RTX30 after the price reduction, or buy the new ones instead of the old ones?
