跳到正文
The Elec· JEONG IL JOO·· 10 天前AI 評分57

AMD 發佈第六代 EPYC Venice,稱 EPYC 9996 基準效能最高達 Nvidia Vera 系統的 3.75 倍

AMD Launches 6th-Gen EPYC Venice, Claims 3.75 Times Nvidia Performance

AI 導讀

AMD 正式推出採用 Zen 6 架構的第六代 EPYC「Venice」伺服器 CPU,產品現已商用,主要伺服器廠商已開始出貨相關系統。AMD 公佈的雙插槽 EPYC 9996 在 SPECrate 2017 integer 測試中得分 5,470,搭載 Nvidia Vera CPU 的系統得分為 1,459,效能最高高出 3.75 倍。

正文

AMD officially launched its sixth-generation EPYC “Venice” processors. (Photo: AMD)

AMD officially launched its sixth-generation EPYC “Venice” processors. (Photo: AMD)

AMD officially launched its sixth-generation EPYC “Venice” processors. (Photo: AMD)
AMD officially launched its sixth-generation EPYC “Venice” processors. (Photo: AMD)

AMD has officially launched its sixth-generation EPYC “Venice” server central processing unit (CPU), aiming to expand market share by offering better price-performance efficiency than competing products.

AMD unveiled key details of its sixth-generation EPYC processors based on the Zen 6 microarchitecture on Nov. 10. The announcement came 10 months after Chief Executive Officer Lisa Su surprised audiences by showcasing the physical silicon at her CES 2026 keynote in January.

Codenamed Venice, the sixth-generation EPYC processors are now commercially available, with major server manufacturers beginning shipments of Venice-based systems. Hyperscale cloud providers including Microsoft Azure and Oracle Cloud Infrastructure (OCI) also plan to make virtual machine instances powered by the processors available to general customers soon.

AMD has divided the Venice lineup into three main categories: graphics processing unit (GPU) host nodes, multi-agent artificial intelligence (AI) workloads and general-purpose enterprise computing. Each category is optimized for its respective workloads.

(Photo: AMD)
(Photo: AMD)

The GPU host-node models feature higher single-core clock speeds and increased PCI Express (PCIe) 6.0 bandwidth. A high-frequency 96-core model delivers 1.8 to 1.9 times the AI inference output throughput on Nvidia B200 GPUs compared with fifth-generation EPYC processors.

The multi-agent AI models, by contrast, use lower clock speeds while maximizing core density, thread counts and power efficiency per agent. The general-purpose high-performance models focus on per-socket performance and cost efficiency.

AMD also disclosed benchmark results for the EPYC 9996, its general-purpose high-performance model. In the SPECrate 2017 integer benchmark, the AMD system scored 5,470 points, compared with 1,459 points for a system powered by Nvidia’s Vera CPU, representing a performance advantage of up to 3.75 times. Both systems used dual-socket configurations.

Compared with Intel’s sixth-generation Xeon Granite Rapids 6980P, the AMD processor delivered 2.74 times higher performance based on the geometric mean of multiple benchmarks. The results included 3.45 times higher planning and orchestration performance in the TPCx-AI benchmark, 2.83 times higher web gateway request throughput measured using the WRK benchmark and NGINX web server, 2.65 times higher performance across various enterprise tools and network workloads, and 2.5 times higher ephemeral tool execution performance.

AMD EPYC Embedded 9006 Series SP7. (Photo: AMD)
AMD EPYC Embedded 9006 Series SP7. (Photo: AMD)

The EPYC 9996 features 256 cores and a default CPU power (DCP) rating of 600 watts. Nvidia’s Vera has 176 cores and a thermal design power (TDP) rating of 250 to 450 watts, while Intel’s Xeon 6980P has 128 cores and a 500-watt TDP rating.

AMD priced the EPYC 9996 at $14,904, compared with $13,955 for Intel’s Xeon 6980P, positioning its processor as offering better performance for the price. The AMD processor also supported more active agents per watt than competing products.

AMD also announced the sixth-generation EPYC Embedded 9006 Series SP7 CPUs targeting the embedded computing market.

來源:The Elec · thelec.net