9621 Agnes Crossing, Lake Suzanneview, New Mexico Island 84604-9295.
With the introduction of the Zen architecture and chiplets, AMD made a strong comeback. Today, chiplets not only enable rapid innovation, but also help reduce the energy consumption
With the introduction of the Zen architecture and chiplets, AMD made a strong comeback. Today, chiplets not only enable rapid innovation, but also help reduce the energy consumption of server chips. What makes the small chip components so important and how do they help AMD gain a market position?
AMD launched its first chips based on the Zen microarchitecture in 2017. Since then, the manufacturer has been building on Zen. Both the Ryzen PC CPUs and the Epyc data center chips are based on various iterations of the architecture.
Epyc is gaining market share in data centers at a time when they are under fire from the general public. For example, some concerned Irish people recently took data centers to the Supreme Court over their impact on electricity consumption and therefore on the climate. In other words, AMD is competing for market share in a rather controversial sector.
“There are indeed components and solutions that consume more power than before,” notes Alexander Troshin, EMEA Product and Marketing Manager, Enterprise and HPC Server at AMD. “But overall, CPUs are becoming more efficient.”
AMD uses specific efficiency targets for this. By 2025, the CPU specialist’s chips must be thirty times more efficient than those from 2020. “Today we are already at a factor of 23,” says Troshin.
He says that data center energy consumption can be drastically reduced by replacing five-year-old servers with new products. “Depending on the situation, you can sometimes reduce a server park by two-thirds. Where you needed 1,000 servers, you can now do that with 300 to 400 servers that consume the same or even less power, which is mainly due to the architectural improvements in the CPUs.”
Troshin points out that such a renewal of the server park creates room for investment in AI without necessarily increasing the overall energy consumption of a data center. “If you first optimize what you already use, you can then add servers for AI and maybe even use less overall.”
AMD is working on two fronts to increase the efficiency of Epyc processors. “The first aspect has to do with the production process,” explains Troshin. “We always use the latest process through partners such as the chip manufacturer TSMC.”
More than fifteen years ago, AMD had its own factories, just like Intel does today. But in 2008, the CPU designer divested the factories through a new company that now goes by the name GlobalFoundries. GlobalFoundries has long been a key supplier to AMD, but the company has now relied on TSMC for its modern processors.
“Thanks to the sale of our factories, we can work with different market participants,” says Troshin. “It is up to them to make the necessary investments to keep up with the latest technology, because that is their core business.” AMD therefore has the freedom to turn to the most advanced chip manufacturer currently available.
The potentially limited capacity of a partner is a disadvantage, as you share it with the competition. On the other hand, there is an important advantage: flexibility. Intel, for example, relies primarily on its own factories, but experienced a stagnation in the development of its production facilities when it made the leap to ten nanometers. AMD’s biggest competitor, for example, fell behind technologically.
We’ve already described in a previous article why the manufacturing process of a chip is so important. The finer the resolution of the process, the smaller the components and thus the transistors on the chip. A smaller transistor switches faster, uses less power, and generates less heat. By baking chips on the most advanced production line available, AMD ensures that Epyc CPUs (just like Ryzen) are built as efficiently as possible.
The process is one pillar of the innovation that AMD is implementing. The second has to do with chiplets. “When we introduced the chiplet design with Zen, it was a pretty radical decision. In 2017, chiplets were not an obvious option. Historically, there have been technical limitations, including in terms of networking the chiplets,” says Troshin.
The chiplets proved to be an excellent choice. While other chip designers were still sticking to a monolithic design, AMD enjoyed the flexibility of different building blocks. “Thanks to the chiplets, we can quickly develop components that are optimized for specific workloads,” explains Troshin.
Troshin: “For example, the Epyc 9000 series has 128 cores but a lower clock speed. The chip was designed to support as many users as possible on a server. The 7000 series uses the same architecture with the same architecture Core chiplet dies (CCDs) but a slightly different layout that is useful for the edge. The Epyc 4000 series again has the same architecture with the same CCD chiplets but more limited memory and IO capabilities tailored to SMBs. Chiplets allow us to use the same core design and still optimize CPUs for specific purposes.”
The chiplets perform so well thanks to AMD Infinity Fabric, which “glues” all the components together. It creates a lightning-fast data highway between all the chiplets. The approach allows the manufacturer to introduce targeted innovations without the need for a new monolithic architecture each time.
AMD is not the only one saying that chiplets are a spectacular idea. It took several years, but now other manufacturers are also using the architecture. Intel is taking important steps towards a chiplet design with Meteor Lake, and Nvidia is also using chiplets for the first time with its Blackwell GPUs. Driven by AMD’s success, small subchips as the basis for larger CPUs and GPUs are gradually becoming the norm.
Intel, AMD and the other players must continue to be at the cutting edge of technology with their designs. More horsepower with less power, more efficiency and higher density are important demands from customers large and small.
In addition, the nanometer war is raging with full force: TSMC, Intel and Samsung want to build the most modern production lines so that the most advanced chips can roll off the production line for at least one or two generations. This is ideal for AMD: the larger the supply, the greater the chance that they can sell their chip volume to one supplier.
AMD has taken advantage of the momentum today, but will continue to struggle to maintain it. “We’ve seen good growth over the last six years,” Troshin says. “According to some sources, we already have a third of the server market. We’re seeing a lot of adoption, but we’re challenging the status quo. Our competition is very strong.”
“In the first generations, we also had a limited solution portfolio,” he notes. “If customers want a certain density in a server and we can’t provide it, the story ends quickly. It takes time to develop a broad solution portfolio and that certainly played a role in AMD’s adoption in the first two to three generations. Today, we have a very comprehensive range.”
After all, it takes more than just the right portfolio of solutions, no matter how flexible they are. Troshin: “The ecosystem is very important. We have to work with ISVs and OEMs. If a particular solution is not yet officially supported by a third party, that is a deterrent.”
Troshin sees AMD now in a favorable position thanks to the flexible chiplet design and the collaboration with TSMC. AMD seems to be increasingly able to break the habits of interested parties with its own solutions. “The market is improving. I see that we will do more and more in the future.”
Source: IT Daily
As an experienced journalist and author, Mary has been reporting on the latest news and trends for over 5 years. With a passion for uncovering the stories behind the headlines, Mary has earned a reputation as a trusted voice in the world of journalism. Her writing style is insightful, engaging and thought-provoking, as she takes a deep dive into the most pressing issues of our time.
