AMD Unveils New Epyc Server CPU in Latest AI Push

AMD Instinct MI300X (Image: AMD)

AMD has released its next-generation Epyc server processor: a high-performance, energy-efficient CPU that’s designed for cloud, enterprise, and artificial intelligence workloads, the company announced today (October 10).

Built with the new Zen 5 core architecture, the new fifth-generation AMD Epyc CPU features up to 192 cores and can be used as a standalone chip for general-purpose workloads or AI inferencing. The hardware can also be paired with AI accelerators like AMD’s Instinct Series GPUs for larger AI models and applications, executives said.

The new AMD Epyc 9005 series processor, formerly code-named Turin, provides up to 17% better instructions per clock (IPC) for enterprise and cloud workloads and up to 37% higher IPC for AI and high-performance computing workloads when compared to AMD’s Zen 4 chips that were first introduced two years ago, the company said.

With the release of the new processor, AMD will “once again take a huge generational leap in performance,” said Forrest Norrod, executive vice president, and general manager of AMD’s data center solutions business Group, during a pre-briefing with media and analysts.

At its Advancing AI event in San Francisco today, the company also announced new GPUs and data center networking solutions, including a new DPU and a NIC to speed AI applications. The chipmaker reiterated its plan to release a new GPU every year, starting with the AMD Instinct MI325X accelerator, which will be available during the fourth quarter of this year.

Related:AI Accelerated Servers Fuel Growth in Data Center Spending

At the event's keynote speech, AMD CEO Lisa Su said AI will enable many new experiences that will make computing an even more essential part of our lives, from accelerating medical discoveries and revolutionizing research to creating smarter, more efficient cities and enabling more resilient supply chains.

“It can really enhance productivity across virtually every industry, and our goal at AMD is to make AMD the end-to-end AI leader,” she said.

To enable that, AMD is “driving the bleeding edge of performance in CPUs, GPUs and high-performance networking,” Su said.

As for the new Epyc CPU, an enterprise looking to modernize its data center can replace seven legacy servers running Intel chips from four or five years ago with one new server running AMD’s new chip, she said.

“That significantly reduces the power you need in your data center. It lowers TCO by more than 60%, and when you add… the (reduced) enterprise software licensing costs, that means an enterprise can break even on their investments in as little as six to 12 months,” Su said.

Related:Data Center Chips in 2024: Top Trends and Releases

Analysts’ Take on AMD’s Announcements

Overall, analysts say AMD is doing what it needs to do to compete against rivals Intel and Nvidia – and it’s doing it very well. In fact, while Intel still dominates, AMD executives said they have captured 34% market share in the server CPU market.

“AMD just continues to execute year after year. They’ve gotten to the point where it’s just improvement, improvement, improvement,” said Jim McGregor, founder and principal analyst at Tirias Research.

Ian Cutress, chief analyst of More than Moore, agreed. “They’re hitting all the right corporate notes. They’re on track with everything they’ve talked about,” he said. “This event is not only about their newest generation CPU, it’s their yearly cadence with the GPU, and they’re talking about networking and the synergy going in between. They’re basically saying, ‘We’re still putting one foot in front of the other, and it turns out, we’re pretty good at it.’”

Analysts say AMD continues to deliver improvements year after year with its new Epyc server CPUs (Image: AMD)

Intel has done a good job with its roadmap and recent release of its Intel Xeon 6 CPUs and Gaudi 3 AI accelerator, but by capturing one-third of the data center CPU market, AMD has momentum on its side, McGregor said.

AMD is also doing well with its entry into the GPU market as an alternative to Nvidia’s GPUs, he said. Many enterprises are just starting to explore how to integrate GPUs and AI workloads into its data centers. There’s strong interest in AMD as another source for GPUs, he said. 

Related:AI Server Market Projected to Reach $180B by 2032

“AMD has momentum. They’re still growing, and as long as they still continue to execute on their roadmap, they’re in a very good position,” McGregor said.

Zen 5 Architecture

The company is using two different Zen 5 core architectures for its fifth-generation CPUs. Zen 5, built using the 4-nanometer manufacturing process, features up to 128 cores and is built for performance. Zen 5c, built using 3nm and features up to 192 cores, is designed for efficiency and optimized for parallelization and throughput, McGregor noted.

It’s very much like the strategy Intel took with its Intel Xeon 6 efficient cores (E-cores) and performance cores (P-cores), the hardware analyst said. 

The reason is that data center operators’ needs are changing because they have different types of workloads that have unique requirements and require different processors.

“Both Intel and AMD have developed that performance and efficiency core strategy,” McGregor said. “They realize they have to be more flexible because we’ve seen some hyperscalers develop their own processors for different applications. So this is kind of their response to the needs of not just the system vendors, but the end customers – the data centers.”

Staying On Message

AMD’s messaging at today’s event is that it can deliver a full infrastructure solution that includes CPUs, GPUs, DPUs, and networking, but the company needs to beef up its software, said Peter Rutten, research vice president in IDC’s worldwide infrastructure research organization.

AMD today said it continues to invest in and improve its AMD ROCm software stack for building AI and HPC applications running on its GPUs. However, Nvidia is far ahead with CUDA, Nvidia AI Enterprise, Nvidia NIM microservices, and Omniverse, McGregor said.

“AMD is basically saying we, too, can deliver you the entire infrastructure and software. That’s good. That’s what customers want,” Rutten said. “So you want those CPUs, GPUs, and fast networking. But I am worried about the actual developer story, the end user story. The software story is still getting short-changed and that should be a main focus.”

AMD’s GPU Roadmap and AI Networking Solutions

On the GPU front, the forthcoming AMD Instinct MI325X will offer 256GB of HBM3E memory and 6TB/s of memory bandwidth, which the company says is 1.8 times more capacity and 1.3 times more bandwidth than Nvidia’s H200 Tensor Core GPU.

AMD said server vendors are expected to begin shipping servers with the MI325X in the 2025 first quarter, including Dell Technologies, Hewlett Packard Enterprise, Lenovo, and Supermicro, and others are expected to begin shipping servers running the AMD.

After the MI325X, the company plans to release the Instinct MI350 series accelerator during the second half of 2025 and the MI400 series in 2026.

The MI350 series GPU will offer 288GB of HBM3E memory capacity and will provide a 35x increase in AI inferencing performance over AMD’s initial GPU – the MI300 series accelerator, the company said.

On the networking front, AMD announced the new AMD Pensando Salina DPU, an accelerator that takes over data processing tasks, such as networking and security, to free up CPU resources.

AMD’s new third-generation Pensando Salina DPU will provide twice the performance, bandwidth, and scale as its previous generation and is designed for the front-end of a data center network, which will improve the performance, efficiency, security, and scalability for data-driven AI applications, the company said.

For the back end of the network, which manages data transfer between accelerators and clusters, AMD announced the Pensando Pollara 400 NIC, which the company claims will be the first Ultra Ethernet Consortium (UEC)-ready AI NIC, and will reduce latency, improve throughput and prevent congestion.

The DPU and NIC are expected to be available during the first half of 2025.