Two years ago, planning AI infrastructure largely meant sizing for training. Then inference took the lead in shaping requirements. Now agents turn a single request into a shifting chain of retrieval, tool calls, code execution and results, arguably creating a shape-shifting workflow that changes how much compute the AI task requires, and where that compute is needed with every execution.

Infrastructure built around a single point in time risks being the wrong shape for what’s next. Flexibility is no longer a feature. It is a requirement at every layer of the stack, including the CPUs supporting the agentic pipeline.

Enterprises already understand this. Databases, virtualization, analytics, web services, technical computing and AI have long required different system profiles. Agentic AI brings more of those profiles into one workflow.

In response to this, a new AMD white paper evaluates 6th Gen AMD EPYC™ 9006 “Venice” server CPUs across general-purpose, enterprise, cloud-native, AI and high-performance computing workloads rather than relying on a selected few.

In SPECrate® 2026 Integer testing, the AMD EPYC™ 9996 server CPU delivers 1.2 times the per-core performance of an Nvidia Vera-based platform1 and 2.24 times its platform-level performance.2 Across enterprise and cloud-native testing – including server-side Java, OpenSSL, MongoDB, Redis, NGINX and transaction processing – AMD testing reports gains of 2.4x to 3.7x.3

Against the Intel® Xeon® 6980P processor, the AMD EPYC 9996 server CPU delivers 1.8x to 3.13x performance advantages across molecular dynamics, materials modeling and weather forecasting.4 In a modeled 100-kilowatt rack, it delivers an estimated 3.3 times the throughput of a Vera-based platform.5

Strong loaded per-core performance can accelerate latency-sensitive work, while core density can support concurrency and improve rack-level throughput. A processor portfolio lets customers optimize for both without imposing the same compromise across the entire fleet.

“Venice” is one part of a portfolio designed around that principle: four families running on a common software foundation, spanning 8-core edge deployments, 256-core flagship processors and rack-scale AI host nodes. Each role in an agentic pipeline can use the processor profile that fits it without creating a separate operating environment.

“Venice” is in production today. Major OEM platforms are on track to launch, and leading cloud providers begin deploying later this year.

Agentic AI will keep adding diversity. The answer to a more varied workload has never been less choice.

More:

  1. 9xx6-049: SPECrate®2026_int_base per-core comparison based on AMD internal estimates for the 6th Gen AMD EPYC 96c HF processor and NVIDIA Vera as of 7/22/2026. Preliminary performance estimates based on AMD engineering projections and subject to change. 2P 6th Gen AMD EPYC 96c HF processor, GCC 15.2, estimated SPECrate®2026_int_base score 1210 (6.3 per core). 2P NVIDIA Vera: 88 cores, GCC 15.2, estimated SPECrate®2026_int_base score 925 (5.3 per core). For: ~1.2x the performance / core Nvidia Vera performance from: https://nvdam.widen.net/s/nmw5vblpqd/nvidia_vera_cpu_architecture_whitepaper?nvid=nv-tblg-543584. SPEC®, SPEC CPU®, and SPECrate® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. Results may vary.

    9xx6-047: SPECrate®2017_int_base per-core comparison based on AMD internal projections and measurements as of 07/20/2026. 6th Gen AMD EPYC delivers leadership 1P SPECrate®2017_int_base per-core performance versus three competing server processors, each at the AMD core configuration noted. Versus NVIDIA Vera: 1P 96C 6th Gen AMD EPYC 9996 estimated score 1350, per-core 1350 / 96 = 14.06, versus 1P 88C NVIDIA Vera score 729.5, per-core 729.5 / 88 = 8.29.. Versus Intel Xeon 6980P: 1P 128C 6th Gen AMD EPYC estimated score 1565, per-core 1565 / 128 = 12.23, versus 1P 128C Intel Xeon 6980P score 1255, per-core 1255 / 128 = 9.80; . Versus Arm AGI: 1P 96C 6th Gen AMD EPYC 9996 estimated score 1350, per-core 14.06, versus 1P 136C Arm AGI score 972, per-core 972 / 136 = 7.15. Configurations: 6th Gen AMD EPYC (96C and 128C) using AOCC 5.1 with 16x64GB 4R MRDIMM 12800 MT/s; Intel Xeon 6980P using OneAPI with 12x64GB MRDIMM 8800 MT/s; Arm AGI using GCC 13 with 12x64GB MRDIMM 8800 MT/s; NVIDIA Vera using GCC 13 with 8x96GB SOCAMM2. SPEC® and SPECint® are registered trademarks of Standard Performance Evaluation Corporation. Learn more at spec.org. Results may vary.

  2. 9xx6-050: SPECrate®2026_int_base throughput comparison based on AMD internal estimates for the 2P 6th Gen AMD EPYC 9996 and 2P NVIDIA Vera as of 7/22/2026. Preliminary performance estimates based on AMD engineering projections and subject to change. 2P 6th Gen AMD EPYC 9996: GCC 15.2, estimated SPECrate®2026_int_base score 2070. NVIDIA Vera: 88 cores, GCC 15.2, estimated SPECrate®2026_int_base score 925. For ~2.2x the performance https://nvdam.widen.net/s/nmw5vblpqd/nvidia_vera_cpu_architecture_whitepaper?nvid=nv-tblg-543584. SPEC®, SPEC CPU®, and SPECrate® are registered trademarks of the Standard Performance Evaluation Corporation. See www.spec.org for more information. Results may vary.

  3. 9xx6-037: Redis 8.8 Open Source Stable request-throughput comparison based on AMD internal testing as of 7/20/2026. 1P AMD EPYC 9996 (256C) delivers up to 2.9x the Redis 8.8 request throughput of 1P Intel Xeon 6980P (128C) Software: Redis 8.8 Open Source Stable, redis-benchmark SET and GET, one Redis server per 8 vCPUs aggregated, KVM, Ubuntu 24.04. Configurations: 1P Intel Xeon 6980P: 1P Intel Xeon 6980P, 128 cores (256 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, Supermicro SYS-222HA-TN, BIOS v1.5, Max performance BIOS profile, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P Graviton5: 1P Graviton5 (AWS m9gd.metal-48xl), 192 cores (192 vCPU), SMT OFF, 768GB memory, Ubuntu 24.04, kernel Linux 6.17.0-1009-aws. 1P AMD EPYC 9965: 1P AMD EPYC 9965, 192 cores (384 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, AMD reference platform, BIOS RVOT1006C, TDP=500W, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P AMD EPYC 9996: 1P AMD EPYC 9996, 256 cores (512 vCPU), SMT ON, 64GB DDR5-8000 RDIMM fully populated, AMD reference platform, BIOS BKC10.3 90RC4, 600W Default CPU Power, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. Data presented: 1P Intel Xeon 6980P 42636320, 1P Graviton5 50057444, 1P AMD EPYC 9965 76703115, 1P AMD EPYC 9996 124130835.50. Ratio 9996 vs 6980P: 124130835.50 / 42636320 = 2.91 (rounds to 2.9x). Results may vary. Results for AWS Graviton5 were obtained from a publicly available cloud instance. Results compare AMD internal testing of the listed AMD and Intel systems against testing performed on the listed AWS EC2 Graviton5 bare-metal instance. Cloud instance results may be affected by cloud service configuration, instance availability, regional deployment, hypervisor/Nitro behavior, storage/network configuration, and other cloud provider variables. Starting with the 6th Gen AMD EPYC server processor family, AMD uses Default CPU Power to describe processor power consumption, succeeding AMD's historical TDP reference. Default CPU Power reflects total power consumed across the processor's compute and I/O dies for the stated performance target. Default CPU Power and TDP may both serve as processor power references for product comparison, platform planning, and performance-per-watt analysis. Results may not be directly comparable to physical server configurations due to differences in platform implementation, firmware, operating environment, memory configuration, and system tuning.

    9xx6-038: NGINX request-throughput comparison based on AMD internal testing as of 7/20/2026. 1P AMD EPYC 9996 (256C) delivers up to 3.7x the NGINX request throughput of 1P Intel Xeon 6980P (128C) Software: NGINX served with wrk load generator (1KB file), one NGINX server per 8 vCPUs aggregated, KVM, Ubuntu 24.04. Configurations: 1P Intel Xeon 6980P: 1P Intel Xeon 6980P, 128 cores (256 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, Supermicro SYS-222HA-TN, BIOS v1.5, Max performance BIOS profile, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P Graviton5: 1P Graviton5 (AWS m9gd.metal-48xl), 192 cores (192 vCPU), SMT OFF, 768GB memory, Ubuntu 24.04, kernel Linux 6.17.0-1009-aws. 1P AMD EPYC 9965: 1P AMD EPYC 9965, 192 cores (384 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, AMD reference platform, BIOS RVOT1006C, TDP=500W, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P AMD EPYC 9996: 1P AMD EPYC 9996, 256 cores (512 vCPU), SMT ON, 64GB DDR5-8000 RDIMM fully populated, AMD reference platform, BIOS BKC10.3 90RC4, 600W Default CPU Power, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. Data presented: 1P Intel Xeon 6980P 10526326, 1P Graviton5 15929450, 1P AMD EPYC 9965 24907610, 1P AMD EPYC 9996 38451232. Results for AWS Graviton5 were obtained from a publicly available cloud instance. Results compare AMD internal testing of the listed AMD and Intel systems against testing performed on the listed AWS EC2 Graviton5 bare-metal instance. Cloud instance results may be affected by cloud service configuration, instance availability, regional deployment, hypervisor/Nitro behavior, storage/network configuration, and other cloud provider variables. Starting with the 6th Gen AMD EPYC™ server processor family, AMD uses Default CPU Power to describe processor power consumption, succeeding AMD's historical TDP reference. Default CPU Power reflects total power consumed across the processor's compute and I/O dies for the stated performance target. Default CPU Power and TDP may both serve as processor power references for product comparison, platform planning, and performance-per-watt analysis. Results may not be directly comparable to physical server configurations due to differences in platform implementation, firmware, operating environment, memory configuration, and system tuning.

    9xx6-039: MySQL 8.4.10 online transaction processing throughput comparison (HammerDB TPROC-C, a workload derived from the TPC-C Benchmark Standard. Results are based on TPC-C derived workloads and are not comparable to published TPC benchmark results.) based on AMD internal testing as of 7/20/2026. 1P AMD EPYC 9996 (256C) delivers up to 2.6x the MySQL 8.4.10 online transaction processing throughput (HammerDB TPROC-C) of 1P Intel Xeon 6980P (128C) Software: HammerDB TPROC-C on MySQL 8.4.10, 500 warehouse dataset, users scaled to vCPU count, KVM, Ubuntu 24.04. Configurations: 1P Intel Xeon 6980P: 1P Intel Xeon 6980P, 128 cores (256 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, Supermicro SYS-222HA-TN, BIOS v1.5, Max performance BIOS profile, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P Graviton5: 1P Graviton5 (AWS m9gd.metal-48xl), 192 cores (192 vCPU), SMT OFF, 768GB memory, Ubuntu 24.04, kernel Linux 6.17.0-1009-aws. 1P AMD EPYC 9965: 1P AMD EPYC 9965, 192 cores (384 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, AMD reference platform, BIOS RVOT1006C, TDP=500W, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P AMD EPYC 9996: 1P AMD EPYC 9996, 256 cores (512 vCPU), SMT ON, 64GB DDR5-8000 RDIMM fully populated, AMD reference platform, BIOS BKC10.3 90RC4, 600W Default CPU Power, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. Data presented: 1P Intel Xeon 6980P 15499304, 1P Graviton5 23075914, 1P AMD EPYC 9965 25168617, 1P AMD EPYC 9996 40804653. + Results for AWS Graviton5 were obtained from a publicly available cloud instance. Results compare AMD internal testing of the listed AMD and Intel systems against testing performed on the listed AWS EC2 Graviton5 bare-metal instance. Cloud instance results may be affected by cloud service configuration, instance availability, regional deployment, hypervisor/Nitro behavior, storage/network configuration, and other cloud provider variables. Starting with the 6th Gen AMD EPYC™ server processor family, AMD uses Default CPU Power to describe processor power consumption, succeeding AMD's historical TDP reference. Default CPU Power reflects total power consumed across the processor's compute and I/O dies for the stated performance target. Default CPU Power and TDP may both serve as processor power references for product comparison, platform planning, and performance-per-watt analysis. Results may not be directly comparable to physical server configurations due to differences in platform implementation, firmware, operating environment, memory configuration, and system tuning.

    9xx6-040: OpenSSL cryptographic throughput comparison based on AMD internal testing as of 7/20/2026. 1P AMD EPYC 9996 (256C) delivers up to 2.5x the OpenSSL cryptographic throughput of 1P Intel Xeon 6980P (128C) Software: OpenSSL cryptographic throughput workload, KVM, Ubuntu 24.04. Configurations: 1P Intel Xeon 6980P: 1P Intel Xeon 6980P, 128 cores (256 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, Supermicro SYS-222HA-TN, BIOS v1.5, Max performance BIOS profile, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P Graviton5: 1P Graviton5 (AWS m9gd.metal-48xl), 192 cores (192 vCPU), SMT OFF, 768GB memory, Ubuntu 24.04, kernel Linux 6.17.0-1009-aws. 1P AMD EPYC 9965: 1P AMD EPYC 9965, 192 cores (384 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, AMD reference platform, BIOS RVOT1006C, TDP=500W, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P AMD EPYC 9996: 1P AMD EPYC 9996, 256 cores (512 vCPU), SMT ON, 64GB DDR5-8000 RDIMM fully populated, AMD reference platform, BIOS BKC10.3 90RC4, 600W Default CPU Power, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. Data presented: 1P Intel Xeon 6980P 552433496.01, 1P Graviton5 545579628, 1P AMD EPYC 9965 1073669574.93, 1P AMD EPYC 9996 1389064585.24. Results for AWS Graviton5 were obtained from a publicly available cloud instance. Results compare AMD internal testing of the listed AMD and Intel systems against testing performed on the listed AWS EC2 Graviton5 bare-metal instance. Cloud instance results may be affected by cloud service configuration, instance availability, regional deployment, hypervisor/Nitro behavior, storage/network configuration, and other cloud provider variables. Starting with the 6th Gen AMD EPYC™ server processor family, AMD uses Default CPU Power to describe processor power consumption, succeeding AMD's historical TDP reference. Default CPU Power reflects total power consumed across the processor's compute and I/O dies for the stated performance target. Default CPU Power and TDP may both serve as processor power references for product comparison, platform planning, and performance-per-watt analysis. Results may not be directly comparable to physical server configurations due to differences in platform implementation, firmware, operating environment, memory configuration, and system tuning.

    9xx6-041: MongoDB throughput comparison based on AMD internal testing as of 7/20/2026. 1P AMD EPYC 9996 (256C) delivers up to 3.5x the MongoDB throughput of 1P Intel Xeon 6980P (128C) Software: MongoDB throughput workload, KVM, Ubuntu 24.04. Configurations: 1P Intel Xeon 6980P: 1P Intel Xeon 6980P, 128 cores (256 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, Supermicro SYS-222HA-TN, BIOS v1.5, Max performance BIOS profile, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P Graviton5: 1P Graviton5 (AWS m9gd.metal-48xl), 192 cores (192 vCPU), SMT OFF, 768GB memory, Ubuntu 24.04, kernel Linux 6.17.0-1009-aws. 1P AMD EPYC 9965: 1P AMD EPYC 9965, 192 cores (384 vCPU), SMT ON, 64GB DDR5-6400 RDIMM fully populated, AMD reference platform, BIOS RVOT1006C, TDP=500W, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. 1P AMD EPYC 9996: 1P AMD EPYC 9996, 256 cores (512 vCPU), SMT ON, 64GB DDR5-8000 RDIMM fully populated, AMD reference platform, BIOS BKC10.3 90RC4, DEFAULT CPU POWER 600W, CPPC on, power determinism, Ubuntu 24.04, kernel Linux 6.17.0-35-generic. Data presented: 1P Intel Xeon 6980P 1794003, 1P Graviton5 2088232, 1P AMD EPYC 9965 3590497, 1P AMD EPYC 9996 6195941. Results compare AMD internal testing of the listed AMD and Intel systems against testing performed on the listed AWS EC2 Graviton5 bare-metal instance. Cloud instance results may be affected by cloud service configuration, instance availability, regional deployment, hypervisor/Nitro behavior, storage/network configuration, and other cloud provider variables. Results for AWS Graviton5 were obtained from a publicly available cloud instance. Starting with the 6th Gen AMD EPYC™ server processor family, AMD uses Default CPU Power to describe processor power consumption, succeeding AMD's historical TDP reference. Default CPU Power reflects total power consumed across the processor's compute and I/O dies for the stated performance target. Default CPU Power and TDP may both serve as processor power references for product comparison, platform planning, and performance-per-watt analysis. Results may not be directly comparable to physical server configurations due to differences in platform implementation, firmware, operating environment, memory configuration, and system tuning.

  4. 9xx6-031: GROMACS (benchPEP) performance comparison based on AMD internal testing as of 07/15/2026. Software: AOCC 5.1 with AOCL 5.1 (lp64) and OpenMPI 4.1.5 on the AMD EPYC parts; Intel compiler 2023.0.0 with MKL 2023.0.0 and OpenMPI 4.1.5 on Intel Xeon 6980P. All results are the mean of 3 runs. Intel Xeon 6980P: 2P (128C per socket, 256C total, SMT off, 1 thread per core), 24x 64 GiB DDR5 at 8800 MT/s (1536 GiB total), BIOS 1.4, Supermicro SYS-222HA-TN, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.55.1). AMD EPYC 9755 (5th Gen): 2P (128C per socket, 256C total, SMT off, 1 thread per core), 24x 64 GiB RDIMM DDR5 at 6400 MT/s (1536 GiB total), BIOS TVOT1009C, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.52.1). AMD EPYC 9965 (5th Gen): 2P (192C per socket, 384C total, SMT off, 1 thread per core), 24x 64 GiB RDIMM DDR5 at 6400 MT/s (1536 GiB total), NPS4, BIOS TVOT100BB, High Performance Mode, power determinism, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). 6th Gen AMD EPYC 9996 (RDIMM): 2P (256C per socket, 512C total, SMT off, 1 thread per core), 32x 64 GiB RDIMM DDR5 at 8000 MT/s (2048 GiB total), NPS2, BIOS TNIV0090D, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). 6th Gen AMD EPYC 9996 (MRDIMM): 2P (256C per socket, 512C total, SMT off, 1 thread per core), 32x 64 GiB MRDIMM DDR5 at 12800 MT/s (2048 GiB total), NPS2, BIOS TNIV0090D, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). Data presented (ns/day): 6980P 2.937; 9755 4.425; 9965 5.663; 9996 RDIMM 8.249; 9996 MRDIMM 9.195. Results may vary due to factors including system configurations, software versions and BIOS settings.

    9xx6-032: NAMD (20stmv) performance comparison based on AMD internal testing as of 07/15/2026, results in ns/day. Software: AOCC 5.1 with AOCL 5.1 (lp64) and OpenMPI 4.1.5 on the AMD EPYC parts; Intel compiler 2023.0.0 with MKL 2023.0.0 and OpenMPI 4.1.5 on Intel Xeon 6980P. All results are the mean of 3 runs. Intel Xeon 6980P: 2P (128C per socket, 256C total, SMT off, 1 thread per core), 24x 64 GiB DDR5 at 8800 MT/s (1536 GiB total), BIOS 1.4, Supermicro SYS-222HA-TN, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.55.1). AMD EPYC 9755 (5th Gen): 2P (128C per socket, 256C total, SMT off, 1 thread per core), 24x 64 GiB RDIMM DDR5 at 6400 MT/s (1536 GiB total), BIOS TVOT1009C, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.52.1). AMD EPYC 9965 (5th Gen): 2P (192C per socket, 384C total, SMT off, 1 thread per core), 24x 64 GiB RDIMM DDR5 at 6400 MT/s (1536 GiB total), NPS4, BIOS TVOT100BB, High Performance Mode, power determinism, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). 6th Gen AMD EPYC 9996 (RDIMM): 2P (256C per socket, 512C total, SMT off, 1 thread per core), 32x 64 GiB RDIMM DDR5 at 8000 MT/s (2048 GiB total), NPS2, BIOS TNIV0090D, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). 6th Gen AMD EPYC 9996 (MRDIMM): 2P (256C per socket, 512C total, SMT off, 1 thread per core), 32x 64 GiB MRDIMM DDR5 at 12800 MT/s (2048 GiB total), NPS2, BIOS TNIV0090D, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). Data presented (ns/day): 6980P 0.556; 9755 0.858; 9965 1.125; 9996 RDIMM 1.72; 9996 MRDIMM 1.71. Results may vary due to factors including system configurations, software versions and BIOS settings.

    9xx6-034: Quantum ESPRESSO (Ta205) performance comparison based on AMD internal testing as of 07/15/2026, results in wallclock time (s). Software: AOCC 5.1 with AOCL 5.1 (lp64) and OpenMPI 4.1.5 on the AMD EPYC parts; Intel compiler 2023.0.0 with MKL 2023.0.0 and OpenMPI 4.1.5 on Intel Xeon 6980P. All results are the mean of 3 runs. Intel Xeon 6980P: 2P (128C per socket, 256C total, SMT off, 1 thread per core), 24x 64 GiB DDR5 at 8800 MT/s (1536 GiB total), BIOS 1.4, Supermicro SYS-222HA-TN, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.55.1). AMD EPYC 9755 (5th Gen): 2P (128C per socket, 256C total, SMT off, 1 thread per core), 24x 64 GiB RDIMM DDR5 at 6400 MT/s (1536 GiB total), BIOS TVOT1009C, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.52.1). AMD EPYC 9965 (5th Gen): 2P (192C per socket, 384C total, SMT off, 1 thread per core), 24x 64 GiB RDIMM DDR5 at 6400 MT/s (1536 GiB total), NPS4, BIOS TVOT100BB, High Performance Mode, power determinism, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). 6th Gen AMD EPYC 9996 (RDIMM): 2P (256C per socket, 512C total, SMT off, 1 thread per core), 32x 64 GiB RDIMM DDR5 at 8000 MT/s (2048 GiB total), NPS2, BIOS TNIV0090D, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). 6th Gen AMD EPYC 9996 (MRDIMM): 2P (256C per socket, 512C total, SMT off, 1 thread per core), 32x 64 GiB MRDIMM DDR5 at 12800 MT/s (2048 GiB total), NPS2, BIOS TNIV0090D, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). Data presented (wallclock time (s)): 6980P 2864.897; 9755 2287.51; 9965 2831.003; 9996 RDIMM 1798.487; 9996 MRDIMM 1594.417. Results may vary due to factors including system configurations, software versions and BIOS settings.

    9xx6-035: WRF (conus2.5km) performance comparison based on AMD internal testing as of 07/15/2026, results in mean time (s). Software: AOCC 5.1 with AOCL 5.1 (lp64) and OpenMPI 4.1.5 on the AMD EPYC parts; Intel compiler 2023.0.0 with MKL 2023.0.0 and OpenMPI 4.1.5 on Intel Xeon 6980P. All results are the mean of 3 runs. Intel Xeon 6980P: 2P (128C per socket, 256C total, SMT off, 1 thread per core), 24x 64 GiB DDR5 at 8800 MT/s (1536 GiB total), BIOS 1.4, Supermicro SYS-222HA-TN, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.55.1). AMD EPYC 9755 (5th Gen): 2P (128C per socket, 256C total, SMT off, 1 thread per core), 24x 64 GiB RDIMM DDR5 at 6400 MT/s (1536 GiB total), BIOS TVOT1009C, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.52.1). AMD EPYC 9965 (5th Gen): 2P (192C per socket, 384C total, SMT off, 1 thread per core), 24x 64 GiB RDIMM DDR5 at 6400 MT/s (1536 GiB total), NPS4, BIOS TVOT100BB, High Performance Mode, power determinism, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). 6th Gen AMD EPYC 9996 (RDIMM): 2P (256C per socket, 512C total, SMT off, 1 thread per core), 32x 64 GiB RDIMM DDR5 at 8000 MT/s (2048 GiB total), NPS2, BIOS TNIV0090D, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). 6th Gen AMD EPYC 9996 (MRDIMM): 2P (256C per socket, 512C total, SMT off, 1 thread per core), 32x 64 GiB MRDIMM DDR5 at 12800 MT/s (2048 GiB total), NPS2, BIOS TNIV0090D, Red Hat Enterprise Linux 10.1 (Linux 6.12.0-124.56.1). Data presented (mean time (s)): 6980P 1.51; 9755 1.066; 9965 1.084; 9996 RDIMM 0.671; 9996 MRDIMM 0.52. . Results may vary due to factors including system configurations, software versions and BIOS settings.

  5. 9xx6-021: Rack-level general purpose performance comparison within a 100 kW rack power constraint based on AMD analysis as of July 2026 comparing 2P AMD EPYC™ 9996 (512 total cores) vs 2P Nvidia Vera (176 total cores, Olympus).. Performance reflects geometric mean across 6 workloads: estimated SPECrate® 2017_int_base, Server-Side Java® Multi-JVM Max jOPS, Web serving (NGINX/WRK), Key-Value store (Redis), In-memory caching (Memcached), and Relational DB (TPROC-C). All platforms use 2-socket nodes. System manufacturers may vary configurations, yielding different results. See https://www.amd.com/content/dam/amd/en/documents/solutions/ai/methodology-description.pdf for full derivations and assumptions. TPROC-C workload is an open-source workload derived from TPC-Benchmark™ Standard, and as such is not comparable to published TPC-C™ results, as the results do not comply with the TPC-C Benchmark Standard. TPC, TPC Benchmark and TPC-C are trademarks of the Transaction Processing Performance Council. Preliminary performance estimates based on AMD engineering projections or measurements as of July 22, 2026 and subject to change. SPEC® and SPECint® are registered trademarks of Standard Performance Evaluation Corporation. Learn more at spec.org.

TopicsArtificial IntelligenceData Center #epyc
Share / cite

Press inquiries: corporate.pressinquiry@amd.com