AMD Reveals CDNA Server GPU Architecture, Updated EPYC Roadmap

'What you can expect is the beginning of a new roadmap that will take our compute architecture forward for high-performance computing and machine learning,' AMD CEO Lisa Su says of the new CDNA data center GPU architecture.

AMD has revealed its new CDNA data center GPU architecture for high-performance computing and machine learning as well as its roadmap plans for future EPYC and Ryzen processors over the next few years.

The Santa Clara, Calif.-based company made the disclosures Thursday during its Financial Analyst Day, where AMD CEO Lisa Su said the next few years is about "building the best on the technology side as well as the business side."

[Related: AMD Wraps Up 2019 With Strong 7nm Ryzen, EPYC And Radeon Sales]

The company also disclosed plans to use "X3D" chip packaging for future processors to drive 10 times greater compute density and shared details of its third-generation Infinity Architecture that will improve CPU-GPU connectivity and a second-generation RDNA architecture that will support ray tracing.

During her talk, Su said these investments will propel AMD to a larger market opportunity and faster growth than ever before.

"The opportunities are larger, the impact we can make on the industry is larger, and our resources are much stronger, so if you think about those things and what we've been able to accomplish , it's really exciting to think about what we'll be able to accomplish," she said.

By 2023, AMD expects the total addressable market for its client, server and semi-custom products to reach $79 billion, according to Su. The biggest opportunity will be the data center market, representing $35 billion. PCs, on the other hand, will represent a $32 billion market while gaming, which includes game consoles, will be $12 billion.

Su said the company expects a roughly 20 percent compounded annual growth rate for the next five years after experiencing a 14 percent growth rate over the last few years. The company also expects higher gross margins at over 50 percent and higher operating margins in the mid-20 percent range.

"We believe we are underrepresented in the [total addressable market]. There's a lot more we can do," the CEO said.

CDNA, RDNA 2 Architectures Detailed

With AMD's newly revealed CDNA, which stands for Compute DNA, Su said the company decided to create a GPU architecture for the data center separate from its RDNA graphics architecture, because "there's really an opportunity to optimize" for HPC and machine learning workloads in servers.

Like the company's Zen architecture that powers Ryzen client and EPYC server processors, CDNA will have a multi-generational GPU roadmap.

"What you can expect is the beginning of a new roadmap that will take our compute architecture forward for high-performance computing and machine learning," Su said.

At the top of CDNA's feature list is support for second-generation Infinity Architecture, which will enable eight-way GPU connectivity.

CDNA will also include a reduction in total cost of ownership and improved performance-per-watt over previous GPU architectures. It will also include enhanced enterprise reliability, availability and serviceability as well as improved security and virtualization features.

Previously, AMD used a common architecture, GCN, for GPUs on both the client side and the data center side. But the company determined that moving beyond the general-purpose GPU approach would allow it to focus on domain-specific optimizations, according to David Wang, senior vice president of engineering in AMD's Radeon Technologies Group.

"This allows us through domain-specific optimization to achieve optimal efficiency for gaming and for compute," he said, which means that users on both sides "don’t have to pay for features they don’t need so it's win-win."

This means that while AMD's current 7-nanometer Radeon Instinct data center GPU is based on the older GCN architecture, future products in the category will be based on CDNA, starting later this year with 7nm Radeon GPUs that will be optimized for HPC and machine learning and use second-generation Infinity Architecture, according to Wang.

AMD teased a second-generation CDNA architecture, or CDNA 2, which will use third-generation Infinity Architecture and support exascale applications that demand exaflops of performance.

"We have very, very strong roadmap for data center GPUs," Wang said.

An important aspect of enabling high performance in the data center is ROCm, also known as AMD's Open Compute Platform, which will improve multi-GPU scaling in its fourth iteration due out this year.

As for the gaming and client graphics side, AMD currently has its first 7nm Radeon GPUs in the market that are based on the first-generation RDNA architecture. But starting later this year, the company plans to start releasing Radeon GPUs that will use second-generation RNDA, or RNDA 2.

One central feature of RDNA 2 will be hardware-accelerated ray tracing, a method that was first brought to the gaming and content creation market by Nvidia for simulating realistic lighting and shadows in real time. It will also boost performance-per-watt and add support for variable rate shading.

EPYC, Ryzen Roadmaps Updated With New Products

On top of the GPU news, AMD updated roadmaps for its EPYC server and Ryzen desktop processors while stating that the company expects to reach double-digit share in the server CPU market by the second quarter.

Forrest Norrod, senior vice president and general manager of AMD's Data Center and Embedded Solutions Business Group, said the company is on track to achieve this momentum in a market still dominated by Intel because of the superior performance and efficiency of AMD's EPYC processors.

"We knew that to be considered for the data center we had to be not just a provider of high-performance components. We had to be a reliable partner," he said.

Even when compared to Intel's new Xeon Cascade Lake Refresh processors, AMD's second-generation EPYC processors still have a "performance lead and unmatchable performance-per-dollar advantage," Norrod said.

AMD said it is on track to ship its third-generation EPYC processors, code-named Milan, by the end of the year, Norrod said the CPUs will be based on AMD's 7nm Zen 3 architecture and include improvements that will allow the chipmaker to cover virtually every workload in the data center market.

The company is also actively developing its fourth-generation EPYC processors, code-named Genoa, which Norrod said will use AMD's 5nm Zen 4 architecture and is expected to start shipping by 2022.

But it's no longer just about driving high performance with traditional CPUs, Norrod said. Accelerated computing workloads increasingly rely on a deeper integration between CPUs and GPUs, which is why the chipmaker is working on the CDNA and Infinity architectures, according to the executive.

While CDNA will allow up to eight Radeon GPUs scale more efficiently, CDNA 2 will be the true game-changer in how it smashes the barriers to reach new peaks in performance, according to Norrod. These new CDNA 2 GPUs, combined with EPYC Genoa CPUs, will power the U.S. Department of Energy's El Capitan supercomputer that is scheduled to be delivered in 2023.

"We get something truly special extending to Infinity Architecture 3 to couple the CPUs and GPUs together into one unified data view, where this not only provides additional performance […] it allows programmers to stop worrying about the explicit movement of data," Norrod said.

In addition to the new EPYC disclosures, the company said it will launch fourth-generation Ryzen processors for desktop PCs that will be based on the company's 7nm Zen 3 architecture by the end of this year.