Author | ZeR0
Edit | Moying
Xinjie reported on September 21 that last night, NVIDIA (NVIDIA) launched the new generation of GeForce RTX 40 series graphics cards .
As the world's first GPU based on the new NVIDIA Ada Lovelace architecture, the RTX 40 series has achieved a huge intergenerational leap in performance and efficiency.
Among them, the performance of modern gaming of the new flagship RTX 4090 GPU can be improved by up to 2 times compared to the previous generation 3090 Ti, and the performance of ray tracing games can be improved by up to 4 times. It is no problem to play 4K cyberpunk after is activated by DLSS 3.
Nvidia founder and CEO O Huang Renxun introduced on the special live broadcast of GeForce Beyond in the GTC conference keynote speech, which means that a new era of real-time ray tracing and neural network rendering using AI to generate pixels has arrived. There are three 40 series released in
. The flagship product RTX 4090 24GB will be launched on October 12, with a suggested retail price starting from 12,999 yuan . RTX 4080 16GB and RTX 4080 12GB will be launched in November, with the recommended retail prices starting from 9499 yuan and 7199 yuan respectively.
In comparison, the starting price of RTX 3090 starts at 11,999 yuan, the starting price of RTX 3090 Ti starts at 14,999 yuan, and the starting price of a top-end iPhone 14 Pro Max is 13,499 yuan.
From this perspective, the cost-effectiveness of RTX 4090 is "very good". Top graphics card suppliers such as
ASUS , Colorful , Generation , Gigabyte , Yingzhong , MSI , MSI and Souste will launch GeForce RTX 4090 and 4080 GPU standard and overclocking versions in China. The RTX 40 series GPU will also be sold through products such as the world's leading OEMs such as Acer , Alien, Asus, Dell , HP , Lenovo , and MSI.
NVIDIA will also launch the RTX 4090 and RTX 4080 (16GB) FE versions in limited edition to meet fans' needs.
These are just "appetizers" for GTC's keynote speech. Using the Ada Lovelace architecture, Nvidia launched the super chip DRIVE Thor for autonomous driving computing. The computing power is doubled compared with the previous generation DRIVE Orin, and the floating point performance reaches 2000 TFLOPS.
OVX computer , specially designed for meta-universe applications, has also been upgraded to the second generation, equipped with the new Ada Lovelace L40 data center GPU.
and the new microrobot Jetson Orin Nano, which is 80 times faster than the previous generation Jetson Nano.
In addition, Nvidia's flagship computing product H100 GPU released to data centers in April this year also ushered in key progress - is fully launched in .
is aimed at meta-universe applications. Nvidia has also further expanded its platform coverage through cloud services for the first time - releasing Nvidia 's first software and infrastructure-as-a-service (Iaash) product, Omniverse Cloud, providing comprehensive cloud services for the design, release, operation and experience of meta-universe applications.
11, 40 series graphics cards are all available in seconds! TSMC customized 4N process
Nearly 25 years ago, Nvidia launched a programmable coloring GPU, and GPU completely changed 3D graphics.
In 2018, at the global computer graphics and image top conference SIGGRAPH, NVIDIA launched the new GPU architecture NVIDIA RTX, which extends programmable shaders through two new processors - RT Core is used to accelerate real-time ray tracing, and Tensor Core is used to process matrix operations and accelerate AI.
Today, Nvidia's ultimate move that has been held back for 4 years - , the third-generation RTX architecture Ada Lovelace, has finally officially debuted!
This generation of RTX is named after the mathematician Ada Lovelace, who is recognized as the world's first computer programmer .
According to reports, Ada GPU can achieve 2 times the performance improvement of traditional rasterized gaming, and the performance improvement of ray tracing games can be as high as 4 times the performance improvement of . Compared with the previous generation Ampere architecture, Ada can bring more than 2 times the performance improvement of with the same power consumption.
"Ada is paving the way for future games based entirely on simulation," said Huang Renxun.
Today, Nvidia launched three GPUs based on the Ada Lovelace architecture: GeForce RTX 4090 is available in 24GB version, and GeForce RTX 4080 is available in 16GB and 12GB versions.
GeForce RTX 4090 GPU is the flagship product of the new GeForce RTX 40 series and is the world's first gaming GPU based on the new NVIDIA Ada Lovelace architecture.
RTX 4090 has 76 billion transistors, 16384 CUDA cores and 24 GB high-speed Micron GDDR6X video memory. It continues to run at more than 100 FPS in 4K resolution games, and has achieved significant improvements in power consumption, silence, heat dissipation, etc.
In the complete ray tracing game, the performance improvement of the RTX 4090 with DLSS 3 can reach 4 times compared with the previous generation of flagship GPU RTX 3090 Ti with DLSS 2.
In modern games, the performance of RTX 4090 is as high as 2 times , while maintaining the same 450W power consumption as the RTX 3090 Ti. One of the keys to the performance soaring in
is that Ada has introduced the new NVIDIA DLSS 3 super resolution technology . This function can use low-resolution content as input without affecting image quality and response speed, and use AI technology to create more high-quality frames.
Huang Renxun said that playing modern ray tracing games like "Cyberpunk 2077" requires more than 600 ray tracing calculations per pixel to determine the lighting, which is as high as 16 times higher than the first ray tracing games launched four years ago. However, the number of transistors responsible for such computing in GPUs has not increased year-on-year. With AI, Nvidia has improved its performance by 16 times in four years.
Whether it is a game that requires high GPU performance or a game that is restricted by CPU, it will benefit from this technology. 3D artists can render a complete ray tracing environment with precise physics and realistic materials without agents and view the effects in real time. The configurations of the two sub-flagship RTX 4080s of
are obviously widespread from the RTX 4090.
RTX 4080 16GB has 9728 CUDA cores and 16 GB high-speed Micron GDDR6X video memory, and its performance in modern games can reach 32 times that of the GeForce RTX 3080 Ti; at lower power, has stronger performance than the GeForce RTX 3090 Ti.
RTX 4080 12GB has 7680 CUDA cores and 12GB Micron GDDR6X video memory, and its performance is the same as the 3090 Ti.
2. 7 major technological innovations, bringing the performance of RTX 40 series, this time the performance of RTX 40 series GPUs has been greatly improved, and there is a series of technical innovations behind it.
1. Architecture improvements: Nvidia and TSMC have jointly created the GPU-optimized 4N custom process , enabling the RTX 40 series to integrate 76 billion transistors and more than 18,000 CUDA core , which is 70% more than the previous generation Ampere, and the performance power consumption ratio is as high as 2 times.
2, SM Streaming Multiprocessor: has shader capabilities up to 90 TFLOPS, and its throughput exceeds that of the previous generation of products.
3, shader Execution Reorder (SER): improves execution efficiency by rescheduling the shader load on the fly, thereby better utilizing GPU resources. This technology can reschedule tasks in real time, and is called "the same major innovation as the out-of-order execution of CPU" by Huang Renxun. It can improve ray tracing performance by 2-3 times and improve overall gaming performance by 25%.
4, third-generation RT Cores: has an effective ray tracing computing power of 191 TFLOPS, 2.8 times that of the previous generation of products.
The third generation RT Cores can provide 2x the intersecting performance between light and triangles, as well as two new important hardware units. Opacity Micromap engine improves the Alpha-Test geometry performance of ray tracing by 2 times; Micro-Mesh engine can dynamically generate microgrids to generate additional geometry, which can improve the richness of geometry without the cost of traditional complex geometry processing performance and storage costs.
5, fourth-generation Tensor Cores: has added Hopper FP8 Transformer Engine, and the FP8 tensor processing performance is as high as 1.4 Petaflops, which is 5 times the acceleration performance of the previous generation using FP8.
6, Ada optical flow accelerator: brings a 2x performance improvement, allowing DLSS 3 to predict motion in the scene, allowing neural network to improve the frame rate while maintaining image quality.
7, dual NVIDIA encoder (NVENC) shortens the output time by up to half and supports AV1. OBS, Blackmagic Design DaVinci Resolve, Discord and more companies are already using the NVENC AV1 encoder.
3. 2000 TFLOPS, the strongest autonomous driving super chip is here!
Before launching the new generation of autonomous driving chips, Huang Renxun reviewed his achievements as usual: Nvidia launched Xavier in 2018, was the world's first robot processor designed for deep learning. Every two years, Nvidia will release a new generation of processors with a leap in performance. Last year, Nvidia released Altan to 1,000 TOPS.
Today, Huang Renxun released a new big move - NVIDIA DRIVE Thor has a throughput of 2 times that of Atlan, the integer peak performance can reach 2000 TOPS, and the FP8 accuracy peak performance can reach 2000 TFLOPS, while reducing the overall system cost. The goal is the automaker's 2025 model.
achieves this goal due to three factors: Grace CPU, Hopper GPU and Ada Lovelace GPU. Hopper's integrated Transformer engine helps accelerate computing, and the invention of multi-instance GPUs in Ada will help centralize on-board computing resources, which can reduce costs by hundreds of dollars.
Thor can be configured in multiple modes, all of its computing power can be used for autonomous driving workflows, or part of it is used for cockpit AI and infotainment, and the other part is used for driver assistance.
Thor's multi-computing domain isolation allows concurrent, time-sensitive multi-processes to run without interruption. The vehicle can run Linux, QNX and Android on a single computer.
The current car's parking, active safety, driver monitoring, camera mirroring, clustering, infotainment and other functions are controlled by different computing devices, and these functions can be uniformly supported by Thor in the future.
Two DRIVE Thors can also use the latest NVLink-C2C chip interconnection technology to "stitch" into a more powerful chip, as the overall platform for running a single operating system.
Back to Nvidia's second-generation robot processor DRIVE Orin, which has been adopted by more than 40 manufacturing companies of cars, trucks, driverless taxis and shuttle buses. The basic processing assembly line of autonomous driving cars can be applied to various robot systems .
Jetson series is a robot computer created by Nvidia with 1 million developers. At this GTC conference, Huang Renxun announced the launch of a micro robot computer Jetson Orin Nano, which is 380 times faster than the previous generation Jetson Nano.
There are mobile robots and robot systems that observe moving objects. Nvidia's edge AI platform Metropolis has downloaded 1 million times, with more than 1,000 application partners around the world. Orin is also a robot processor for the industrial-grade IGX Edge AI platform where Metropolis runs.
Global large-scale industrial automation company uses Metropolis and Orin IGX for its industrial edge computing platform.
In addition to robot development, Orin IGX is also an ideal computing platform for medical imaging applications. NVIDIA Clara Holoscan, running on Orin IGX, is a low-latency imaging processing platform that contains libraries for data processing, AI model training, simulation and robot development applications. More than 70 leading medical device companies, startups and medical centers are all developed on Clara Holoscan.
Activ Surgical, Proximie and Moon Surgical will build their surgical robot system on NVIDIA Clara Holoscan running on the Orin IGX platform.
4. Sword to the meta-universe: The second-generation OVX computer upgrades the Ada architecture and launches the first Iaas cloud service
for meta-universe applications. Huang Renxun announced the launch of second-generation OVX computer , powered by the new Ada Lovelace L40 data center GPU and enhanced network technology to provide breakthrough real-time graphics, AI and digital twin simulation functions.
With the help of 48GB super large frame buffer, the second generation OVX with 8 L40 GPUs will be able to complete the super large Omniverse virtual world simulation. The L40 GPU has entered mass production in full. The second generation of OVX systems will be available to the market early next year.
In addition to the exclusive hardware of the metaverse, Nvidia has also created its first IaaS product, Omniverse Cloud service, , which can be connected to Omniverse applications running on the cloud, on-premises or devices. Individuals or teams can experience the ability to design and collaborate with the service in one click without any local computing power.
The new Omniverse container is now available for cloud deployment, including Replicator for generating synthetic data, Farm for extending rendering farms, Isaac Sim for building and training AI robots, and more.
After Nvidia's Isaac platform for autonomous mobile robots enters the cloud, users can obtain cloud-ready Omniverse VMI virtual machine images and Isaac containers on NGC and deploy them to any public cloud.
5. From the cloud to supercomputer, H100 is fully put into production
Finally, let’s take a look at the latest progress in accelerated computing towards data centers and high-performance computing.
Huang Renxun said that the NVIDIA platform now has 3.5 million developers, and 12,000 startups are opening new businesses based on NVIDIA's products. NVIDIA provides acceleration for about 3,000 applications through 550 SDKs and AI models. "Overall, the total value of the industries we serve is about $100 trillion."
is aimed at data centers. Nvidia's latest flagship product, H100 Tensor Core GPU, released in April this year, has entered mass production of .
H100 contains 80 billion transistors and adopts a variety of innovative technologies such as the new Hopper architecture, Transformer engine, second-generation multi-instance GPU, confidential computing, fourth-generation NVIDIA NVLink interconnection, DPX instructions, etc., which can be used to accelerate ultra-large-scale AI model training such as advanced recommendation systems and large-scale language models.
According to reports, H100 enables enterprises to reduce the deployment cost of AI. Compared with the previous generation of A100, when providing the same AI performance, the energy efficiency can be improved by 3.5 times, the overall cost of ownership will be reduced to 1/3, and the number of server nodes used will also be reduced to 1/5.
NVIDIA Global Technology Partners plan to launch the first batch of products and services based on the NVIDIA Hopper architecture in October. More than 50 server models are expected to be available by the end of this year, and dozens of models will be available in the first half of 2023.
AWS, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure will be the first to deploy H100-based instances on the cloud starting next year. The new generation of supercomputers from several world-leading higher education and research institutions will also adopt the H100.
DGX H100 system is now available for ordering. The peak performance of the FP8 accuracy of this system can reach 32 PFlops. Each DGX system includes NVIDIA Base Command and NVIDIA AI Enterprise software, enabling cluster deployment from a single node to NVIDIA DGX SuperPOD.
In terms of software support, H100 now includes a five-year license to NVIDIA AI Enterprise software suite, which will optimize the development and deployment of AI workflows and ensure that users can obtain the AI frameworks and tools needed to build AI chatbots, recommendation engines, visual AI, etc.
Some of the world's leading large language models and deep learning frameworks are being optimized on the H100. These frameworks combined with the Hopper architecture can significantly improve AI performance and reduce the training time of large language models to several days or even hours.
6. Launch two large language models Cloud services to assist in biomedical research
Large language model (LLM) is one of the most important AI models today. With LLM, users can perform specific tasks efficiently by simply fine-tuning the model with fewer samples. The Hopper architecture helps to lower the training and deployment thresholds of LLM.
Today, Nvidia launched Nemo LLM cloud service , for training large language models.
Nemo contains a series of pre-trained basic models built by the community. Its API can generate learned prompt embedding tables and optimized microservices. It can be deployed locally and in the cloud, suitable for one GPU, multiple GPUs and multiple nodes. Register now and you will be able to get the first experience of this service in October.
Nvidia has also launched BioNeMo LLM service , for training and deploying large biomolecular language models at supercomputing scale.
Leading pharmaceutical companies, biotech startups and cutting-edge biological researchers are using BioNeMo LLM services and frameworks to develop AI applications for generating, predicting, and understanding biomolecular data to better understand diseases and find treatments.
NVIDIA BioNeMo LLM service will provide 4 pre-trained language models:
1, ESM-1:, the protein LLM originally published by Meta AI Labs can process amino acid sequences and ultimately generate characterizations for predicting various protein properties and functions. It also improves scientists' ability to understand the structure of protein .
2, OpenFold: This is a sota protein modeling tool created by the Openfold alliance jointly established by academia and industry. It can provide its open source AI workflow through BioNeMo services.
3, MegaMolBART: , a generative chemical model trained on 1.4 billion molecules, can be used for reaction prediction, molecular optimization and the generation of new molecules.
4, ProtT5: This model was developed in collaboration under the leadership of RostLab, the University of Technology of Munich, and NVIDIA is also one of the participants in the project. PortT5 extends the function of protein LLM such as ESM-1b to sequence generation.
These models are optimized for inference and will provide a first-hand experience through the cloud API running on NVIDIA DGX Foundry.
Conclusion: Nvidia has become a full-stack computing company
GPU invented by Nvidia in 1999, which stimulated the growth of the PC game market, redefined computer graphics cards, and fueled the wave of popularization of modern AI. The newly launched Ada Lovelace GPU has improved all three RTX processors as neural rendering engines, bringing new productivity tools to gamers and virtual world creators.
can be seen that Nvidia has developed into a full-stack computing company. Whether it is accelerated computing or computer graphics, it has achieved performance breakthroughs through innovative superposition in architecture, design and algorithms. At the same time, AI technology has penetrated into every corner of Nvidia products, used to combine with more technological innovations, promote breakthroughs in more new AI applications in the science and industry, and provide impetus for the development of digital economy .
