Nvidia's future is not just about GPUs
Latest update time：2020-10-06
Reads：
Founded in 1993, NVIDIA is best known for its GPU. Especially in recent years, due to the popularity of AI, the attention paid to NVIDIA GPU has skyrocketed, and the industry's recognition of them in this field has reached an unprecedented level. But in fact, GPU is only the foundation of NVIDIA. After acquisitions and product line expansion in the past few years, NVIDIA has developed multiple product lines, and DPU is one of them.
DPU: Born for Data Movement
In the data center field, CPU and GPU are well-deserved "processor giants". They have established an unshakable position in the data center with their respective advantages. But in the view of Nvidia CEO Huang Renxun, DPU has become the third member of the data-centric accelerated computing model. CPU is used for general computing, GPU is used for accelerated computing, and DPU moves data around the data center for data processing.
The DPU mentioned here is the abbreviation of Data Processing Unit, which is one of the flagship products brought by Mellanox, an Israeli chip manufacturer acquired by Nvidia for US$6.9 billion.
According to NVIDIA, DPU is a new type of programmable processor with high performance, software programmability and multi-core CPU features. In this SoC, there is a high-performance network interface that can parse and process data at the rate of wired transmission and quickly transmit it to the GPU and CPU. More importantly,
Nvidia said that the DPU's rich, flexible and programmable acceleration engines can reduce and improve the performance of AI and machine learning applications. All of these DPU functions are essential to achieving isolated bare metal cloud native computing, and it will also define the next generation of cloud-scale computing. They further pointed out that the DPU can be used as a standalone embedded processor, but it is usually integrated into SmartNIC to support future servers.
From NVIDIA's introduction, we know that DPU can perform tasks such as network, storage and security that originally required CPU processing. This means that if DPU is used in the data center, a lot of CPU computing power can be released to execute a wide range of enterprise applications.
At the GTC conference held recently, NVIDIA brought its new DPU products BlueField-2, BlueField-2X, and the new DOCA SDK.
First, let's look at the BlueField-2 DPU. As shown in the figure below, the DPU has eight 64-bit Arm Cortex A72 cores, 2VLIM acceleration engines and Mellanox's ConnertX-6 Dx NIC, and integrates the industry-leading 50Gb/s PAM4 SerDes and PCIe Gen 4.0 interface, which allows it to provide two ports with speeds of 25Gb/s, 50Gb/s or 100Gb/s, or an Ethernet and InfiniBand connection with speeds of up to 200Gb/s.
Thanks to these configurations, the BlueField-2 DPU can accelerate security, networking and storage tasks in the data center, including isolation, root of trust, key management, RDMA/RoCE, GPUDirect, elastic block storage and data compression.
In addition to the BlueField-2 DPU, NVIDIA also brought the BlueField-2X DPU. In addition to all the features of the BlueField-2 DPU, this product also integrates an NVIDIA Ampere GPU, which allows it to use AI to perform tasks such as security, networking, and storage in the data center.
NVIDIA said that because the Ampere GPU uses NVIDIA's third-generation Tensor Core, it can use AI to perform real-time security analysis, including identifying abnormal traffic to prevent confidential data theft, line-speed encrypted traffic analysis, host self-checking to identify malicious activities, dynamic security processes and automatic response.
After introducing these two DPUs, NVIDIA also brought a new roadmap for DPUs. As shown in the figure below, NVIDIA's next two generations of DPUs will be available in 2022 and 2023, and the performance improvements they bring are very obvious.
In order to facilitate the development of DPU, NVIDIA also brought an SDK called DOCA (Data-Center-Infrastructure-on-a-ChipArchitecture). It is understood that DOCA provides developers with a comprehensive open platform to assist them in building software-defined, hardware-accelerated network, storage, security and management applications on the BlueField series DPU. DOCA has also been fully integrated into NVIDIA NGC, a software catalog that provides a convenient and containerized software development environment for third-party application developers, which means they can take advantage of DPU acceleration services in data centers and develop, certify and distribute applications to their customers.
JETSON: Embracing the AIoT Revolution
In Nvidia's future product layout, AIoT is also a key direction they will not miss. As Deepu Talla, the company's edge computing VP and GM, said, this is a market involving trillions of connections, and there is no reason for them to miss it.
As shown in the figure above, this is a product line that started in 2014. In March of that year, they released the first product in the Jeston series, Jeston TK1. This is the world's first mobile supercomputer for embedded systems, and its applications include computer vision, image processing, and real-time data processing.
According to NVIDIA, Jetson is their embedded system for the new generation of autonomous machines. It is a series of AI platforms suitable for all autonomous machines. The performance and energy efficiency provided by its system can increase the running speed of autonomous machine software and consume less power. According to NVIDIA's information, each system in the Jetson series is a complete modular system (SOM) with CPU, GPU, PMIC, DRAM and flash memory, and is scalable. For developers and users, they only need to choose the SOM that suits the functional requirements of the application scenario, and they can build a system based on it.
Since 2014, NVIDIA has launched six products, including TK1, TX1, TX2, AGX Xavier, Nano and XavierNX, for different application scenarios. Now, they have brought the Jetson Nano 2GB, which is priced at only $59. In NVIDIA's view, this is an AI and robotics starter kit that is very suitable for students, education and robotics enthusiasts.
NVIDIA officials also stated that the original intention of designing the Jetson Nan 2GB open kit was to integrate the teaching and learning of AI, and those interested in it can use it to develop projects in the fields of robotics and smart Internet of Things. To support this work, NVIDIA will also provide free online training and AI certification programs, which will support thousands of developers in providing more open source projects, development methods and videos in the vibrant Jetson community.
NVIDIA further pointed out that NVIDIA JetPack™ SDK provides support for the new Jetson Nan 2GB, which allows developers to carry out diversified development based on this demand.
Developer Platform: Nvidia’s Weapon
In order to allow developers to bring its hardware to various fields, NVIDIA has made a lot of investment in software, and CUDA is its most successful representative. As many industry insiders have pointed out, NVIDIA's continued investment in CUDA is the reason why it can make great progress in the AI era, which is also the reason why they invested in DOCA development for DPU.
In order to make it more convenient for developers, NVIDIA has brought more development platforms. For example, in order to better bring AI to edge applications, NVIDIA launched the EGX AI platform in 2019. The platform can sense, understand and process data in real time without sending the data to the cloud or data center first.
As a high-performance and scalable platform, EGX can expand from a small NVIDIA Jetson Nano to all server clusters equipped with NVIDIA GPUs, providing support from 0.5TOPS to 10,000 TOPS, and can provide real-time speech recognition and other complex AI experiences for hundreds of users. According to them, server vendors including Dell, Inspur, Lenovo and Supermicro also provide support for the NVIDIA EGX AI platform. This allows large industries such as manufacturing, health, retail, logistics, agriculture, telecommunications, public safety and broadcast media to benefit from the EGX AI platform and accelerate their AI deployment.
Nvidia said its EGX platform is being expanded to combine NVIDIA Ampere GPUs and BlueField-2 DPU capabilities on a single PCIe card, providing enterprises with a versatile platform to build secure accelerated data centers.
At GTC2020, NVIDIA also brought the new NVIDIA Maxine, a cloud-native streaming video AI platform.
According to reports, based on this platform, service providers are expected to bring new AI features to more than 30 million web conferences held every day, including gaze correction, super resolution and noise reduction. Since the data is processed in the cloud rather than on local devices, end users can enjoy new features without any dedicated hardware.
"The Maxine platform will greatly reduce the bandwidth required for video calls. Unlike traditional streaming video that transmits pixels across the entire screen, AI software will analyze the key facial points of each person on the call and then intelligently adjust the face in the other side of the video. This allows streaming video to be transmitted over the network with less traffic," Nvidia emphasized.
The Maxine platform integrates multiple NVIDIA AI SDK and API technologies. In addition to NVIDIA Jarine, the Maxine platform also leverages the NVIDIA DeepStream high-throughput audio and video streaming SDK and the NVIDIA TensorRTTM SDK to achieve high-performance deep learning inference. AI audio, video, and natural language features provided in the NVIDIA SDK used in Maxine
It was developed through hundreds of thousands of training runs on NVIDIA DGX systems, further demonstrating its leadership.
In addition, NVIDIA also brought the Omniverse platform.
According to reports, Omniverse is the world's first 3D simulation and collaboration platform based on NVIDIA RTX. It brings together NVIDIA's breakthroughs in graphics, simulation, and AI, integrates the physical and virtual worlds, and can simulate reality in real time with realistic details.
Using the platform, remote teams can collaborate on projects simultaneously, such as architects iterating on 3D architectural designs, animators modifying 3D scenes, and engineers collaborating on self-driving car development, just as easily as they could edit a document together online.
Omniverse is supported by many major software leaders including Adobe, Autodesk, Bentley Systems, Robert McNeel & Associates and SideFX, and NVIDIA plans to work with other leading software providers so that all artists and designers can choose the applications they need on Omniverse.
Thanks to its leading strength in GPU, NVIDIA has firmly established itself as the leading chip supplier in the AI market. Coupled with the company's investment in software and hardware as mentioned above, I believe that NVIDIA will definitely have a place in the data center field and the AIoT market in the future.
*Disclaimer: This article is originally written by the author. The content of the article is the author's personal opinion. Semiconductor Industry Observer reprints it only to convey a different point of view. It does not mean that Semiconductor Industry Observer agrees or supports this point of view. If you have any objections, please contact Semiconductor Industry Observer.
Today is the 2449th issue of content shared by "Semiconductor Industry Observer" for you, welcome to follow.
Recommended Reading
★ Thoughts on insufficient wafer fab capacity
★ Will the third generation semiconductors really become popular?
Semiconductor M &A pool
Semiconductor Industry Observation
" The first vertical media in semiconductor industry "
Real-time professional original depth
Scan the QR code , reply to the keywords below, and read more
Wafer｜IP｜ SiC｜M&A｜RF｜TSMC｜Nvidia｜Apple
Reply
Submit your article
and read "How to become a member of "Semiconductor Industry Observer""
Reply Search and you can easily find other articles that interest you!
京公网安备 11010802033920号