AMD takes on Intel and Nvidia with new 256-core EPYC server CPUs
Network World ·
As enterprises race to deploy AI agents and expand AI infrastructure, AMD has launched a new generation of EPYC server processors built to challenge both Intel’s dominance in enterprise CPUs and Nvidia’s ambitions in data-center computing. The 6th Gen EPYC lineup offers up to 256 cores per socket and supports PCIe Gen 6 and DDR5 memory, with AMD touting major gains in performance, efficiency and AI throughput. “With 6th Gen AMD EPYC, customers can standardize on the proven x86 foundation while choosing a workload-optimized solution to run more agents, keep accelerated infrastructure fed and power the business-critical applications they depend on,” Dan McNamara, senior vice president and general manager, Compute and Enterprise AI, AMD, said in a statement . Systems powered by the new CPUs are now shipping from leading server manufacturers and original design manufacturers (ODMs), while cloud instances from Microsoft Azure and Oracle Cloud Infrastructure (OCI) will be available soon. The hardware mechanics The EPYC 9006 SP7 range comprises 10 processor models , with configurations ranging from 64 to 256 cores and up to 512 threads per socket. The portfolio includes high-density processors for workloads requiring large numbers of simultaneous computing tasks, as well as high-frequency variants with boost speeds of up to 5.3GHz. Across the SP7 range, the processors support 16 channels of DDR5 memory and PCIe Gen 6 connectivity, offering twice the bandwidth per lane compared with PCIe Gen 5. “Rather than simply increasing core counts, AMD aims to deliver more computing power within existing power and infrastructure limits. This makes the platform particularly relevant for cloud computing, virtualization, high-performance computing (HPC), enterprise databases, data analytics and AI infrastructure,” said Manish Rawat, semiconductor analyst at TechInsights. Benchmarking the competition AMD is pitching the new processors against Intel’s Xeon and Nvidia’s Vera CPUs, claiming significant performance advantages in its internal and published benchmark comparisons. The company claims a server powered by its flagship 256-core EPYC 9996 delivers up to 3.7 times the performance of a system powered by Nvidia’s Vera CPU. Against Intel’s Xeon 6980P, and up to 2.2 times the performance and twice the performance per dollar. For AI agent workloads, AMD also claims its high-density EPYC processors can deliver up to 3.3 times more agents per CPU watt than Nvidia Vera and 2.5 times more than Intel Xeon 6980P when hosting agent sandboxes. Meanwhile, its high-frequency processors can deliver up to 1.8 times the LLM output throughput of the previous-generation EPYC processors in CPU-offloaded inference workloads. AMD EPYC, Intel Xeon, and Nvidia Vera represent three different approaches to the evolving data-centre CPU market, added Rawat. “AMD competes through high core density, memory capabilities and power efficiency, while Intel leverages its established enterprise ecosystem, platform compatibility and customer relationships. Nvidia is taking a different approach with Vera, an Arm-based CPU designed for AI and accelerated computing infrastructure. Unlike AMD and Intel, which target a broad range of general-purpose server workloads, Nvidia is positioning Vera as part of an integrated AI computing platform.” For instance, Danish Faruqui, CEO at Fab Economics, recommends AMD’s high-core-count EPYC processors for cloud consolidation, VDI and AI agent workloads, while he says the high-frequency EPYC models are better suited to databases and enterprise applications where software licensing costs are a concern. For memory-intensive HPC and scientific workloads, he recommends EPYC 9006, while recommending Nvidia Vera for tightly integrated AI training and inference clusters. Intel Xeon 6 remains an option for enterprises looking to maintain existing x86 infrastructure, particularly in air-cooled racks, he said. What CIOs need to evaluate For CIOs evaluating servers powered by the new EPYC processors, the choice will depend on the workloads they intend to run, rather than simply opting for the highest core count. “AI agents and cloud applications may benefit from more cores, while some databases need faster individual cores. They should compare performance, electricity consumption, memory capacity, software compatibility, and cost. The best processor is the one that runs their applications most efficiently,” said Pareekh Jain, CEO at EIIRTrend & Pareekh Consulting. AMD’s flagship EPYC 9996 is priced at around $14,900, broadly comparable to high-end Intel Xeon processors, Jain noted. Nvidia Vera pricing is less transparent, he noted. Jain also advised enterprises to calculate the total cost of ownership over three to five years. This includes hardware, electricity, cooling, software licensing, maintenance, and migration costs. AMD’s higher computing capacity could allow companies to replace several older servers with fewer new ones. However, expensive software licensing and migration costs could reduce these savings. Enterprises running servers that are three to six years old, facing high electricity bills, or expanding their AI workloads have stronger reasons to consider upgrading. Companies with relatively new servers, however, may not see enough savings to justify replacing them.