Dai Weili re-emerged and established a chip company to focus on DPU

The History of Semiconductor Entrepreneurship of “Happy” Couple

Since Dai Weili and Zhou Xiuwen left Marvell in 2016, there has been little news about them. But now, we find that they have returned to the semiconductor circuit, or that they have never left the semiconductor circuit. Yes, they founded a DPU company. You must know that the two have deep knowledge in semiconductors. After 5 years, the two came out again and settled in DPU, which shows that they are optimistic about the prospects of DPU.

Regardless of the cloud speed race, DPU has now become the standard for hyperscalers. Alibaba and Baidu are already DPU users, while Google is suspected of putting them behind the scenes. Microsoft is doing similar things with FPGAs. VMware has also announced the integration of SmartNIC into the VMware Cloud Foundation Project Monterey project. In the face of the slowdown of CPU Moore’s Law, there is a growing consensus in the industry that CPUs cannot be wasted on some housekeeping functions such as network and storage, and must be focused on core computing. And SmartNIC can just uninstall everything from network functions to many data plane and control plane functions. DPU is not popular!

Founded DPU company, “dream is big”

In the past ten years, computing costs have been stable, but network and storage loads have been increasing, and most importantly, the gap between network performance and computing performance has been widening (as shown in the figure below). As early as 2018, more than 70% of the Ethernet ports shipped at a speed of about 10G/sec. At this performance level, the processing overhead required to provide services for the network interface began to become an issue.

Dai Weili re-emerged and established a chip company to focus on DPU

图源:THE NEXT PLATFORM

According to THE NEXT PLATFORM report, Microsoft has 1 million or more FPGA-based “Catapult” smart network cards, and Amazon also has the same number of “Nitro” smart network cards as its self-developed Arm processors. But looking at the entire hyperscalers, including enterprises, cloud builders, and telecommunications companies, there are about 12 to 14 million servers in total, but only 2 million to 3 million machines may have smart network cards. The possibility of these players self-developed smart network cards is almost zero, so there is still a huge market space.

Coupled with the fact that more and more customers are now hosting virtual machines on cloud servers, which makes cloud networking and security quite complicated. Therefore, some industry visionaries see the potential to offload some work to SmartNIC, which can free up CPU cores and further isolate customer workloads from the dangers of other tenants and the wider Internet. Dai Weili and Zhou Xiuwen also saw this opportunity.

The name of the company they established is [DreamBigSemi], which was established in 2019. From this name, we can see that their “dream is big”. DreamBigSemi mainly provides comprehensive low-cost, low-latency and high-throughput SmartNIC solutions for data center and storage acceleration. And the Dream Big Semiconductor team has an unparalleled track record in disruptive technology. Among them, Zhou Xiuwen is the executive chairman and Dai Weili is the executive vice chairman. In addition to the two we know, SOHAIL SYED is the CEO. SOHAIL SYED is also an amazing figure . Sohail is a strong entrepreneur who has successfully founded many companies including Questarium (acquired by Marvell) and FIRQuest (acquired by Corigine).

Dai Weili re-emerged and established a chip company to focus on DPU

Picture from DreamBig Semiconductor official website

DreamBig Semiconductor is headquartered in San Jose, California, but has offices in different regions of the world. It is worth mentioning that most of their team is based in Pakistan, which may be a performance of decoupling from China and the United States.

You know, Marvell, which was co-founded by the two, is now also very important in the semiconductor industry. This time DPU set off again, and Marvell may end up in a battle.

Marvell has been doing research on SmartNIC since a long time ago. As of September 2020, Marvell’s OCTEON-based LiquidIO® SmartNIC has shipped 1 million units. The OCTEON® series has also become the most widely deployed data processing unit (DPU) in SmartNIC cloud applications. Facing the already powerful Marvell, I don’t know how DreamBigSemi can get a share of it.

At present, Marvell’s latest solution is LiquidIO III, which has developed into a SmartNIC platform that combines the previously widely deployed OCTEON TX2 DPU and up to 36 Arm V8-based cores, 5 x100G network connections, and up to 2 PCI Express Gen 4×16 host interface and 6 DDR4 3200 controller channels. Oracle is one of the first customers of LiquidIO III.

In the early stage of DPU development, it is important to grab the market of cloud vendors as early as possible. It is understood that Marvell is also cooperating with a number of cloud hyperscale data center operators to develop customized solutions that enable them to combine their intellectual property rights with Marvell’s enhanced and widely deployed OCTEON LiquidIO SmartNIC DPU.

The History of Semiconductor Entrepreneurship of “Happy” Couple

Dai Weili and Zhou Xiuwen founded Marvell in 1995, and they met when they were studying at the University of California, Berkeley. Both are immigrants. Dai Weili was born in Shanghai and Zhou Xiuwen was born in Jakarta, Indonesia. The two have run Marvell for more than 20 years, and after achieving record profitability and scale in a short period of time, they listed the company in 2000. Marvell has also become the world’s top semiconductor company.

Dai Weili re-emerged and established a chip company to focus on DPU

Until 2016, Marvell’s board of directors expelled both parties due to financial problems. In fact, this is just an example of a capital game. Jobs was also kicked out of the Apple company he founded by the board of directors.

Although both resigned their positions, they never left the semiconductor industry. On September 16, 2019, the electronic equipment multi-standard connection IP solution company announced that Dr. Zhou Xiuwen has joined its board of directors. In September 2020, NUVIA’s Series B financing raised US$240 million, with the participation of Dai Weili and Zhou Xiuwen.

Zhou Xiuwen is known as an Indonesian child prodigy. His enthusiasm for electronic products began very early. At the age of 13, he became a qualified radio maintenance technician in Indonesia. Since then, he has been designing components and systems. He is also an IEEE researcher. Its FLC and MoChi technologies are also quite leading.

Weili Dai is also one of the representative women sought after by the industry. Due to her contribution to technology and society, Newsweek named Ms. Dai as one of the “150 Women Who Shocked the World” and was named the “Leading Women” series by CNN International : Leading women’s technology charging, leading women’s principles of fairness and care, educating for future success, and leading women to inspire others. In addition, “Forbes” magazine listed Ms. Dai as one of the “100 Most Influential Women in the World”.

The climate of DPU has matured

Ten years ago, when the first major wave of hardware acceleration technology hit. GPU products began to appear. Now, as the logic unit of FPGA expands to more than 3 million, we also tightly couple it with other composable processing blocks for network, memory, storage, and computing. In this case, computing means implementing core clusters on the chip through SoC blocks or even ACAP. It has to be said that SoC and FPGA are now mature enough to become the basic technology of SmartNIC. It is with such progress that the second wave of hardware acceleration began to appear.

In the exploration of the DPU road, the benevolent sees the benevolence and the wise sees the wisdom, and the adopted architecture presents a multi-faceted blooming situation:

(1) Select the FPGA-based representation SmartNIC camp, they can be like processing network processing network computing and storage, as in the development, may have a high degree of programmability, such as CPU, the solution may be as fast as the development of the SoC solutions new features, such as Xilinx claims, which compared with Alveo U25 SmartNIC Arm-based multi-core, with the same power lower rate, the performance can be increased 10-fold.

However, there are also two types based on FPGA, one is FPGA-based smart network card, and the other is due to the need for backward compatibility, which has spawned another type of SmartNIC-FPGA-enhanced SmartNIC. It adds FPGA functionality to the NIC. There are three ways to develop FPGA-enhanced SmartNIC designs. The first is to simply connect the FPGA to the existing NIC. Another option is to design the next-generation SmartNIC ASIC and integrate the FPGA array on the chip. The third option is to add a high-speed chip-to-chip interconnection to the SmartNIC ASIC design and develop an FPGA chipset to connect to the SmartNIC ASIC. It is reported that the foreign company Fungible is studying the combination of hardware programmable FPGA and ASIC network controller.

(2) There are also options based on the Arm multi-core architecture. The advantage of this approach is that it can offload clearly defined tasks, such as standardized security and storage protocols, such as Broadcom and Nvidia . Nvidia also believes that GPU can benefit from the integration with DPU. Their EGX platform is making this effort. However, multi-core SmartNIC cards also have some limitations. First, they are based on software programmable processors. Due to the lack of processor parallelism, these processors are slower when used for network processing; second, the fixed-function hardware in these multi-core ASICs The engine lacks SmartNIC offloading the data plane programmability and flexibility that is increasingly needed. At the same time, the fixed-function engines in the multi-core SmartNIC ASIC cannot be extended to handle new encryption or security algorithms because they lack sufficient programmability and can only accommodate minor algorithm changes.

(3) Those who adopt a heterogeneous architecture believe that heterogeneity has higher flexibility and can bring more efficient data processing efficiency, as well as more direct use of interfaces and better virtualization support, such as the domestic Zhongkeyu According to the KPU architecture, they organize four types of heterogeneous cores to process network protocols, OLAPOLTP processing, machine learning, and secure encryption computing cores.

With the final rise SmartNIC market, we will work with the next wave based FPGA, or ARM core, or heterogeneous hardware accelerator architecture converge together . This will form a variety of superpositions in the accelerated market and change our view of the future of computing. Although it is currently unknown which way DreamBigSemi uses DPU. But one thing is that with the addition of Dai Weili and Zhou Xiuwen, the DPU competition is even more lively.

Posted by:CoinYuppie,Reprinted with attribution to:https://coinyuppie.com/dai-weili-re-emerged-and-established-a-chip-company-to-focus-on-dpu/
Coinyuppie is an open information publishing platform, all information provided is not related to the views and positions of coinyuppie, and does not constitute any investment and financial advice. Users are expected to carefully screen and prevent risks.

Like (0)
Donate Buy me a coffee Buy me a coffee
Previous 2021-07-21 13:59
Next 2021-07-21 14:01

Related articles