AMD Intensify DPU Competition with Pensando Acquisition

If you've been following some of the DPU companies in the market, you've noticed that they're all coming out simultaneously as the cloud services infrastructure is starting to scale rapidly.

With the convergence of multiple requirements for virtualization, telemetry, networking, security, and storage, cloud service vendors are acquiring more equipment to ensure their hardware keeps pace with business growth. According to AMD's analysis, 30 to 80 percent of CPUs today run on cloud servers, not counting private clouds built by enterprises.

Even so, enterprises still want to have the capacity of a public cloud in general, not for deploying thousands of servers but for flexibility considerations.

Whereas service deployment on local servers can take days or even weeks, in the cloud, it only takes minutes, up to a few hours, to build the service containers you need, and the process has largely become automated.

Enterprises no longer have to worry about local servers are not enough, do not have to place orders waiting for the arrival of machines, or encounter a surplus of servers, only to sell used servers. The cloud server can be dynamically scaled on demand, the perfect solution to the pain point on the dosage.

What exactly does DPU solve?

However, CPUs, especially x86 CPUs, do not run efficiently in the cloud, but the total share is so high that this contributes to a large degree of waste. According to AMD's IDC server shipment data analysis, the number of wasted servers is close to 10 million per year. Such an inefficient configuration equates to nearly $75 billion in wasted equipment annually, with more than 65 TWh wasted power.

These big cloud service vendors earn a lot of money simultaneously and want to maximize the benefits. If you can release all the CPU resources in the hands of users, cloud service vendors can provide the same hardware resources to exceed competitors' performance. So Amazon is the first to take this step to free up valuable CPU resources with dedicated hardware, not just offload the network like a smart card, but also offload the storage, management, security, and monitoring loads together, and the work of these resources that occupy nearly 30% of the CPU will all be taken over by dedicated hardware.

NitroSystem Architecture / Amazon

This is the earliest form of DPU or the first case to hit the market and be widely used. Still, as a first mover, Amazon chose the name NitroSystem to include dedicated hardware in addition to the Nitro cards for various shunts and accelerated IO, as well as an additional Nitro security chip.

Because of this, Amazon's AWS has an additional ace up its sleeve with the number of customers already a step ahead and has created the 34% market share that AWS has today.

Acquired DPU company

In the past two years, AMD has snatched a lot of x86 server CPUs from Intel's mouth, and even Intel has admitted this. But both Intel's IPU and Nvidia's DPU have long been laid out. Xilinx's Alveo smart NIC alone is not enough to complete all CPU offloading tasks, so AMD decided to acquire a DPU unicorn, Pensando, outright.

Today vendors such as IBM Cloud, Microsoft Azure, Oracle Cloud, and VMware have deployed Pensando's first and second-generation DPUs, although they have not chosen to deploy them similarly. For example, Microsoft Azure mainly uses DPU-based smart switches to complete SDN (software-defined networking) decoupling, while Oracle Cloud and IBM Cloud use DPU cards in hardware and custom logical applications in software, in addition to the SDK and libraries provided by Pensando to facilitate customers to add software-defined services.

PensandoDPU Roadmap / AMD

From the roadmap, AMD PensandoDPU's architecture evolution is based on the network performance gradually until the third generation architecture, codenamed "Salina" DPU launched, will support up to 800G network, and the given time node of 2024 is also estimated due to AMD's forecast of 800G network began to lay the prediction. It is also worth mentioning that AMD PensandoDPU supports forward software compatibility, which means that older versions of the software can use the data generated by the new hardware.

Another great feature of AMD Pensando is its Distributed Services Platform (DSP), which AMD calls a next-generation infrastructure that can be used in any environment. On the underlying DPU programmable processor of the DPS, Pensando requires low power consumption, low latency/jitter, high bandwidth, and high scalability.

At the same time, the DSP is to provide an on-chip software service that provides flexible offload solutions from software, security, storage, monitoring, and telemetry, replacing the devices and black boxes already in the data center. Finally, the DSP has to provide a policy and service manager to cloud service vendors to create a containerized management platform and integrate it into existing infrastructure and controllers.

Final words

AMD, as the only remaining contestant in the x86 ring, has been taking the lead in advanced processes and 3D packaging to bring the CPU to its ultimate performance.

But nowadays, it has become a fact that performance improvement has become slower, and it isn't easy to get customers to buy the chip stack without substantial architectural improvements, so solutions like DPU to improve CPU utilization have become the first choice for data centers. In such a trend, even if Nvidia did not acquire Mellanox to use DPU, AMD will eventually go down this road.

前: Microwave oven transformer working principle and common troubleshooting

次の: FP8 Format | Standardized Specification for AI