AMD Intuition MI300X (Picture: AMD)
AMD has launched its next-generation Epyc server processor: a high-performance, energy-efficient CPU that’s designed for cloud, enterprise, and synthetic intelligence workloads, the corporate introduced at the moment (October 10).
Constructed with the brand new Zen 5 core structure, the brand new fifth-generation AMD Epyc CPU options as much as 192 cores and can be utilized as a standalone chip for general-purpose workloads or AI inferencing. The {hardware} may also be paired with AI accelerators like AMD’s Intuition Collection GPUs for bigger AI fashions and functions, executives mentioned.
The brand new AMD Epyc 9005 sequence processor, previously code-named Turin, offers as much as 17% higher directions per clock (IPC) for enterprise and cloud workloads and as much as 37% greater IPC for AI and high-performance computing workloads when in comparison with AMD’s Zen 4 chips that have been first launched two years in the past, the corporate mentioned.
With the discharge of the brand new processor, AMD will “as soon as once more take an enormous generational leap in efficiency,” mentioned Forrest Norrod, govt vice chairman, and normal supervisor of AMD’s information middle options enterprise Group, throughout a pre-briefing with media and analysts.
At its Advancing AI occasion in San Francisco at the moment, the corporate additionally introduced new GPUs and information middle networking options, together with a brand new DPU and a NIC to hurry AI functions. The chipmaker reiterated its plan to launch a brand new GPU yearly, beginning with the AMD Intuition MI325X accelerator, which might be out there through the fourth quarter of this yr.
On the occasion’s keynote speech, AMD CEO Lisa Su mentioned AI will allow many new experiences that can make computing an much more important a part of our lives, from accelerating medical discoveries and revolutionizing analysis to creating smarter, extra environment friendly cities and enabling extra resilient provide chains.
“It could actually actually improve productiveness throughout nearly each trade, and our objective at AMD is to make AMD the end-to-end AI chief,” she mentioned.
To allow that, AMD is “driving the bleeding fringe of efficiency in CPUs, GPUs and high-performance networking,” Su mentioned.
As for the brand new Epyc CPU, an enterprise trying to modernize its information middle can change seven legacy servers working Intel chips from 4 or 5 years in the past with one new server working AMD’s new chip, she mentioned.
“That considerably reduces the facility you want in your information middle. It lowers TCO by greater than 60%, and once you add… the (diminished) enterprise software program licensing prices, which means an enterprise can break even on their investments in as little as six to 12 months,” Su mentioned.
Analysts’ Tackle AMD’s Bulletins
Total, analysts say AMD is doing what it must do to compete towards rivals Intel and Nvidia – and it’s doing it very nicely. In actual fact, whereas Intel nonetheless dominates, AMD executives mentioned they’ve captured 34% market share within the server CPU market.
“AMD simply continues to execute yr after yr. They’ve gotten to the purpose the place it’s simply enchancment, enchancment, enchancment,” mentioned Jim McGregor, founder and principal analyst at Tirias Analysis.
Ian Cutress, chief analyst of Greater than Moore, agreed. “They’re hitting all the proper company notes. They’re on observe with all the things they’ve talked about,” he mentioned. “This occasion is just not solely about their latest technology CPU, it’s their yearly cadence with the GPU, and so they’re speaking about networking and the synergy getting into between. They’re principally saying, ‘We’re nonetheless placing one foot in entrance of the opposite, and it seems, we’re fairly good at it.’”
Analysts say AMD continues to ship enhancements yr after yr with its new Epyc server CPUs (Picture: AMD)
Intel has completed a great job with its roadmap and up to date launch of its Intel Xeon 6 CPUs and Gaudi 3 AI accelerator, however by capturing one-third of the info middle CPU market, AMD has momentum on its facet, McGregor mentioned.
AMD can be doing nicely with its entry into the GPU market as a substitute for Nvidia’s GPUs, he mentioned. Many enterprises are simply beginning to discover how one can combine GPUs and AI workloads into its information facilities. There’s sturdy curiosity in AMD as one other supply for GPUs, he mentioned.
“AMD has momentum. They’re nonetheless rising, and so long as they nonetheless proceed to execute on their roadmap, they’re in an excellent place,” McGregor mentioned.
Zen 5 Structure
The corporate is utilizing two completely different Zen 5 core architectures for its fifth-generation CPUs. Zen 5, constructed utilizing the 4-nanometer manufacturing course of, options as much as 128 cores and is constructed for efficiency. Zen 5c, constructed utilizing 3nm and options as much as 192 cores, is designed for effectivity and optimized for parallelization and throughput, McGregor famous.
It’s very very like the technique Intel took with its Intel Xeon 6 environment friendly cores (E-cores) and efficiency cores (P-cores), the {hardware} analyst mentioned.
The reason being that information middle operators’ wants are altering as a result of they’ve various kinds of workloads which have distinctive necessities and require completely different processors.
“Each Intel and AMD have developed that efficiency and effectivity core technique,” McGregor mentioned. “They notice they need to be extra versatile as a result of we’ve seen some hyperscalers develop their very own processors for various functions. So that is type of their response to the wants of not simply the system distributors, however the finish clients – the info facilities.”
Staying On Message
AMD’s messaging at at the moment’s occasion is that it may ship a full infrastructure resolution that features CPUs, GPUs, DPUs, and networking, however the firm must beef up its software program, mentioned Peter Rutten, analysis vice chairman in IDC’s worldwide infrastructure analysis group.
AMD at the moment mentioned it continues to spend money on and enhance its AMD ROCm software program stack for constructing AI and HPC functions working on its GPUs. Nonetheless, Nvidia is way forward with CUDA, Nvidia AI Enterprise, Nvidia NIM microservices, and Omniverse, McGregor mentioned.
“AMD is principally saying we, too, can ship you the whole infrastructure and software program. That’s good. That’s what clients need,” Rutten mentioned. “So that you need these CPUs, GPUs, and quick networking. However I’m anxious concerning the precise developer story, the tip person story. The software program story continues to be getting short-changed and that must be a major focus.”
AMD’s GPU Roadmap and AI Networking Options
On the GPU entrance, the forthcoming AMD Intuition MI325X will provide 256GB of HBM3E reminiscence and 6TB/s of reminiscence bandwidth, which the corporate says is 1.8 occasions extra capability and 1.3 occasions extra bandwidth than Nvidia’s H200 Tensor Core GPU.
AMD mentioned server distributors are anticipated to start delivery servers with the MI325X within the 2025 first quarter, together with Dell Applied sciences, Hewlett Packard Enterprise, Lenovo, and Supermicro, and others are anticipated to start delivery servers working the AMD.
After the MI325X, the corporate plans to launch the Intuition MI350 sequence accelerator through the second half of 2025 and the MI400 sequence in 2026.
The MI350 sequence GPU will provide 288GB of HBM3E reminiscence capability and can present a 35x enhance in AI inferencing efficiency over AMD’s preliminary GPU – the MI300 sequence accelerator, the corporate mentioned.
On the networking entrance, AMD introduced the brand new AMD Pensando Salina DPU, an accelerator that takes over information processing duties, equivalent to networking and safety, to liberate CPU assets.
AMD’s new third-generation Pensando Salina DPU will present twice the efficiency, bandwidth, and scale as its earlier technology and is designed for the front-end of an information middle community, which is able to enhance the efficiency, effectivity, safety, and scalability for data-driven AI functions, the corporate mentioned.
For the again finish of the community, which manages information switch between accelerators and clusters, AMD introduced the Pensando Pollara 400 NIC, which the corporate claims would be the first Extremely Ethernet Consortium (UEC)-ready AI NIC, and can scale back latency, enhance throughput and forestall congestion.
The DPU and NIC are anticipated to be out there through the first half of 2025.
