Home > News content

Replace Intel! The secret of NVIDIA data center special processor: one DPU replaces 125 x86 CPUs

via:驱动之家     time:2020/10/7 8:01:29     readed:1488

NVIDIA (NVIDIA) held the GTC 2020 conference online on Monday local time. NVIDIA CEO Huang Renxun once again released a number of new hardware products in his kitchen, including A new software-defined Data Center-specific processor DPU and its eco-product roadmap and announced that it would push ARM architecture in the server market with a view to gaining more market share in the most lucrative server processor market.

Over the past years, GPU chip has been the main product of NVIDIA, and has been widely used in graphics processing and other fields. But in recent years, they have also been widely used in the field of artificial intelligence training to help improve the speed of artificial intelligence for image recognition and other tasks. These chips are usually installed next to the Intel CPU to help speed up the calculation of AI related work. Now NVIDIA plans to replace Intel's CPU with a new software definable data center processor DPU.


Mr Wong said:

Under the plan, NVIDIA will launch two DPU chips in 2021: BlueField-2 and BlueField-2X.

Among them, BlueField-2 integrated eight 64-bit ARM-Cortex A72 kernels, Has 2 super long command word (VLIW) acceleration engines, Integrated NVIDIA Mellanox ConnectX-6Dx NIC( smart card), Can provide two 100 Gb/s network channels, and accelerate key data center security, networking, and storage tasks, Among them are isolation, trust root, key management, RDMA/RoCE、GPU Direct、 elastic block storage, data compression and so on.


the optimized BlueField-2DPU is described to uninstall critical network, storage and security tasks from the CPU, enabling enterprises to transform their IT infrastructure into state-of-the-art data centers. Such data centres can be accelerated, fully programmable and have


In terms of performance, NVIDIA claims that a single bluefield-2 DPU can provide data center services equivalent to that of 125 x86 CPU processors.

It is worth mentioning that, BlueField DPU was born out of the smart network chip SmartNIC technology that NVIDIA acquired for $6.9 billion, Mellanox its integration with ARM's processor IP, which has announced a $40 billion acquisition, BlueField IPU, formed Taking into account the speed and flexibility of software definition solutions, It also improves security, accelerates performance and improves efficiency. First generation BlueField IPU chips have been officially released in 2019. BlueField-2DPU seems to have evolved on the basis of the BlueField-2IPU announced in March this year.

The bluefield-2x DPU to be launched in the future will have all the key features of bluefield-2 DPU. At the same time, NVIDIA ampere GPU, which can be applied to data center security, network and storage tasks, is integrated into a system. AI can be used for real-time security analysis, including identifying abnormal traffic prompting to steal secrets, line speed encrypted traffic analysis, malicious host self-test, and dynamic security choreography automated response (soar).

NVIDIA company's head of computing Manuvir Das said,Adding GPU to bluefield-2x is to accelerate the application of artificial intelligence.


According to the introduction, BlueField-2DPU can currently provide samples, is expected to be used in 2021 in the new system of leading server manufacturers. BlueField-2X DPU is under development and is expected to be listed in 2021.


Then, in 2022, NVIDIA will launch bluefield-3 and bluefield-3x. These DPUs will be integrated into a system with NVIDIA's GPU, but they will still be separated on devices. In 2023, bluefield-4 will be truly integrated in the form of SOC and become a real on-chip data center processor.

In terms of computing power, NVIDIA said,BlueField-4 will be 600 times the BlueField-2. And the data NVIDIA called out really sounds a little too much

In addition to announcing several software defined DPU products for data centers, NVIDIA also launched DOCA (data center infrastructure on a chip architecture) based on DPU. Manuvir DAS, head of enterprise computing at NVIDIA, compares DOCA to CUDA (unified computing device architecture) in the server field.

CUDA is a general parallel computing architecture introduced by NVIDIA, which enables GPU to solve complex computing problems. which contains CUDA instruction set architecture (ISA) as well as parallel computing engines within the GPU. Developers can use C languages to write programs for CUDA architectures. CUDA architecture is also the key to NVIDIA success in GPU and artificial intelligence in recent years.


According to reports, NVIDIA's new DOCA architecture is based on open APIs, such as P4 for packet processing, dpdk for network, spdk for storage, CUDA and NVIDIA AI. DOCA can be seamlessly compatible and integrated with major OS and hypervisors, while programs written for DOCA can run on bluefield-2 DPU and all future versions.

NVIDIA said server vendors, including asus, source, dell technology, fujitsu, giga technology, xinhua san, wave, lenovo, yunda technology and ultramicro, all plan to integrate NVIDIA DPU products into their own server products.

In addition, NVIDIA also launched NVIDIA Jetson nano, a $59 edge computing suite, as well as a number of software tools including Omniverse and Maxine. It also announced the construction of cambridge-1, the most powerful supercomputer in British history, with GlaxoSmithKline.

From the layout of NVIDIA in recent years, from the traditional GPU graphics processor products to CUDA based artificial intelligence computing products, NVIDIA has brought great success. NVIDIA's US $6.9 billion acquisition of mellanox in 2019 is to prepare for further expansion into the data center market. We should know that mellanox's Ethernet products and Infiniband intelligent interconnection solutions have a large market share in the data center market. According to NVIDIA, these products of mellanox are used in more than half of the world's fastest supercomputers and many of the leading ultra large-scale data centers.

The NVIDIA's new $40 billion acquisition of ARM this year may also be aimed at competing head-on with Intel in the lucrative data center market.

This time, the launch of the new Bluefield DPU and DOCA architecture is the beginning of NVIDIA's challenge to Intel in the data center market.

Notably, after the announcement of the acquisition ARM, Huang Renxun publicly said, NVIDIA

The GTC 2020 theme speech concluded with NVIDIACEO Wong stressed:

At the same time, Huang Renxun announced a major move to promote arm platform, and NVIDIA is investing in the following three aspects: firstly, NVIDIA will complement arm in GPU, network, storage and security technology to build a complete acceleration platform; secondly, NVIDIA is cooperating with arm to develop platform for high-performance computing, cloud, edge and PC applications; thirdly, NVIDIA will invest in the following three aspects , NVIDIA is migrating NVIDIA AI and NVIDIA RTX engines to arm.

Wong said:



#NVIDIA#CPU processor#Huang Renxun#Data center

Link to the original text:Smart news

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments