Lisa Su On How AMD Is Building ‘Workload-Optimized’ CPUs

AMD CEO Lisa Su explains how the chipmaker is building ‘workload-optimized’ products like the Ryzen 7 5800X3D CPU for desktop gaming and the EPYC Milan-X CPUs for technical computing, which she believes will help the company set itself apart from rival Intel.

AMD CEO Lisa Su said the chipmaker is putting a greater emphasis on creating processors that are optimized to provide better performance for specific workloads.

Partners can expect to see new workload-optimized CPUs from AMD later this year, with the Ryzen 7 5800X3D CPU for desktop gaming set to arrive this spring and the EPYC Milan-X CPUs for technical computing workloads launching by next month.

[Related: For AMD’s Commercial PC Head, Patience Begets Ryzen Pro Growth]

In a roundtable with journalists in early January, Su said new technologies developed by AMD like 3-D stacking technology are helping the company develop processors that provide a boost for specific workloads over those that serve a broader array of applications.

In the case of the Ryzen 7 5800X3D and EPYC Milan-X processors, AMD is using its 3-D stacking technology to create 3-D chiplets that have an additional cache fused on top of the processor.

This will allow the Ryzen 7 5800X3D to provide, on average, 15 percent faster gaming performance compared with the Ryzen 9 5900X, which came out in 2020, the company has previously said. EPYC Milan-X CPUs, on the other hand, will provide, on average, a 50 percent performance boost for technical computing workloads compared with the third-generation EPYC CPUs that launched last year.

“I do expect over time that you’ll see [3-D chiplets] in more of our portfolio, but part of what we’re trying to build is a portfolio that really is workload-optimized, so it depends on what you’re trying to do,” Su said. “If you’re in a case where you want more cores, you may not need the additional cache. And then on the data center side, when you have these scientific compute workloads that can take advantage of the additional cache or on the consumer side with gaming, we would use it. So think of it as an option, but an option that will go more broadly over our portfolio over time.”

The number of products with 3-D chiplets will be small in comparison to AMD’s overall portfolio in the beginning, but Su said the chipmaker will expand its use of the technology over time, which will help the company amortize the costs of research and development.

“As we take, for example, the 3-D chiplet technology into higher and higher volume, we will expect to see economies of scale, and that’s how you’ll see some of these technologies go across more of the product line,” she said.

However, 3-D chiplet technology is only one of many tools in AMD’s “toolbox” for building workload-optimized products, according to Su. Having a wide range of tools is important, she said, because it gives AMD’s architects more flexibility in how they design products.

“I think all those tools in the toolbox are actually extremely good for the consumer because what comes out is something that’s actually a lot more optimized because we have all of these choices,” she said.

Another tool AMD plans to use for workload-optimized products is the chipmaker’s Zen 4c architecture. The architecture is an offshoot of Zen 4, which will be used for AMD’s fourth-generation EPYC processors, code-named Genoa, that are set to launch this year.

While EPYC CPUs using Zen 4 will be used for a broader range of workloads, processors using Zen 4c will be designed for cloud-native applications. Code-named Bergamo, these processors will feature up to 128 cores, 32 more cores than Genoa’s maximum core count, while also providing “significantly improved power efficiencies and breakthrough performance per socket.”

“We have all of these things in our tool chest, and we’re using the right technology for the right application,” Su said.

While AMD is tapping a wider of set a technologies to remain competitive against Intel, Intel itself has been ramping up new ways of making chips. This includes Intel’s own 3-D packaging technology, which debuted in a small-volume mobile CPU lineup in 2020 before being discontinued a year later. Intel’s first high-volume use of 3-D packaging is expected in 2023 with the Meteor Lake client CPUs.

As for workload-optimized chips, Intel’s latest Xeon Scalable processors that were released last year include models that are optimized for networking applications. The chipmaker also plans to launch a variant of its next-generation Xeon Scalable processors, code-named Sapphire Rapids, with High Bandwidth Memory that will provide a major performance boost for high-performance computing.

Alexey Stolyar, CTO of International Computer Concepts, a Northbrook, Ill.-based HPC system integrator, told CRN that while he expects a small portion of his customers will want servers with EPYC Milan-X CPUs, they will see a “huge benefit” for targeted workloads like computational fluid dynamics and electronic design automation, the latter of which is used to design semiconductor chips.

“We’re really focusing the Milan-X parts around those and with software partners like Ansys, Altair and Siemens to really show the benefit where those CPUs shine,” he said.

Stolyar said International Computer Concepts is hoping to expand its HPC customer base by combining Milan-X servers with a turnkey HPC software solution called LMX Cloud, which is developed by the company’s European sales and integration representative, Define Tech.

“That’s really the approach that we’ve been doing with AMD: to target a bigger audience that would typically stay away from HPC because it’s too difficult while making it easy and accessible,” he said.

Stolyar’s larger concern is one that he and other AMD partners have long held: supply.

“I think this year’s not going to be easy in terms of supply for people,” he said.

On AMD’s earnings call earlier this week, Su said the company has made significant investments in capacities for wafers, substrates and back-end-resources. She said company is also working “much more closely” with customers, which has resulted in the chipmaker receiving forecasts “multiple quarters and, in some cases, multiple years out” for the kind of capacity they need.

“We feel very good about our progress in the supply chain to meet the 2022 guidance, and our goal is, frankly, to have enough supply to satisfy the demand out there,” she said.