AMD Takes On Nvidia's Vera Head-On With Its Venice CPU
AMD said on the 21st that its server CPU, the EPYC 9996 Venice, delivers up to 2.24 times higher platform performance than Nvidia's Vera. In recently released technical materials, AMD explained that the EPYC 9006 Venice, the flagship model of its sixth-generation server CPUs (the core semiconductors responsible for a computer's computation), posted platform performance (the overall system combining CPU, memory, and other components) up to 2.24 times higher than Vera in SPEC CPU 2026 integer throughput testing, an industry-standard benchmark. The result underscores AMD's determination to compete with Nvidia in the AI and data center server market by leveraging the Venice's 256 processing cores.
The gap also showed up in per-core performance. AMD said that on an equal 96-core basis, Venice delivered roughly 20% higher performance than Vera. Compared with the previous-generation 192-core EPYC 9965, the top-end Venice's processing performance improved by about 78%. Against Intel's Xeon 6980P, AMD claimed a 2.37-fold advantage.
However, the results AMD published are based on the company's own tests, and the 2.24-fold figure compares the platform throughput of the 256-core EPYC 9996 against the 88-core Vera, making it difficult to interpret as a difference in individual CPU performance. There were also differences in some testing conditions, such as the compiler used.
AI server competition spreads from GPUs to memory Behind this announcement lies a shift in the landscape of AI server competition. Nvidia is vertically integrating its AI data center computing platform by pushing not only its GPUs but also its own CPU, Vera. AMD is responding with a platform that combines its EPYC CPUs with Instinct GPUs.
AI data center memory demand, which had been concentrated on high-bandwidth memory (HBM), is now spreading to general-purpose server DRAM. In large-scale AI computation, GPUs handle the core calculations while CPUs perform data preprocessing, system management, and storage and networking tasks, so the capacity and bandwidth of the memory attached to the CPU also determine overall system performance. The server CPU competition is broadening beyond core counts to securing memory bandwidth.
Samsung Electronics and SK Hynix, which lead the global DRAM market, are competing not only in HBM but also in DDR5-based server DRAM and high-capacity, high-speed module markets. That said, no new supply contracts have been disclosed showing that AMD specifically selected either company's products for the Venice platform. An industry official said, "As computation in AI servers increases, demand for capacity and bandwidth will inevitably rise not only for GPU-attached HBM but also for the server DRAM connected to CPUs," adding, "The pace at which next-generation server platforms spread will be a key variable in the growth of the high-performance DDR5 and MRDIMM markets."
