![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
- Joined
- Dec 19, 2005
- Messages
- 19,054
“Generational
Everything in this section is AMD's own number, measured or estimated by AMD, and should be treated as such until independent testing lands.
A note on benchmark versions first. AMD quotes SPEC CPU 2026 for the headline integer and floating-point figures, which is the new suite, and its own footnotes record that SPEC retires SPECrate 2017 on November 17, 2026. This is the last launch where both will be quoted side by side, and 2017 and 2026 scores are not comparable to each other.
Against 5th Gen "Turin," AMD claims for the 6th Gen EPYC 9006 series:
- 2.0x integer compute, SPECint 2026
- 2.4x floating point, SPECfp 2026
- 2.7x memory bandwidth, STREAM
- 1.4x HPC compute, HPL
- 1.3x matrix compute, DGEMM
Against the Competition
AMD's competitive comparisons mostly target the Intel Xeon 6980P, with NVIDIA Vera and AWS Graviton5 appearing in several charts.
Workload category claims, with Turin shown for reference:
- Cloud native—3.3x Intel, geomean of 8 workloads (Turin 1.6x)
- Enterprise—2.6x Intel, geomean of 5 workloads (Turin 1.5x)
- Storage and database—3.3x Intel, geomean of 6 workloads (Turin 1.7x)
For agentic workloads, AMD broke the pipeline into stages and measured each against Xeon 6980P: gateway and stream response 2.83x, plan/route/verify 3.45x, retrieve context and similarity search 2.38x, enterprise tools 2.65x, ephemeral tools 2.50x. The consolidated geomean is 2.74x.
Against NVIDIA Vera, AMD claims roughly 2.8x agent density, 1.3x per-core performance and 1.3x memory bandwidth. Against published SPECrate 2026_int_base results, a 2P EPYC 9996 scores 2370, and a 2P EPYC 9686F scores 1270 at 96 cores per socket.
Efficiency
AMD's efficiency claims are the ones that most directly target the power-constrained data center. It quotes a 40% generational performance per watt lift, and up to 3.3x the performance per watt of 6th Gen Intel on general purpose servers. For agent sandboxes, up to 2.8x agents per watt, rising to 3.3x when UPP is enabled.
At rack level, comparing a Helion rack of 256-core Venice at 400 W against Vera at 88 cores and 450 W, AMD claims roughly 1.45x cores per rack, 2.0x performance per megawatt and 2.9x threads per megawatt.
What This Means for Gamers
None of these numbers predict gaming performance. They are throughput benchmarks on 256-core parts, and the competitive ones lean on core count rather than per-core speed.
The IPC claim is the only figure that travels, and even it needs care. AMD measured roughly 20% on enterprise and cloud workloads and roughly 31% on AI workloads against Zen 5, at fixed frequency. Games are neither. Zen 5 claimed 16% IPC and delivered a few percent in games at launch. Educated guess: a desktop Zen 6 lands in the low-to-mid teens in games, with the cache and complex changes contributing more than the core itself.
Once again, the 20% and 27% IPC gain figures presented here are not representative of performance in a consumer/client environment.
Reading These Numbers
Two things are worth keeping in mind. First, a large share of AMD's lead comes from core count rather than per-core performance. A 2P Venice brings 512 cores against 256 on a 2P Xeon 6980P, and the throughput results scale accordingly. That is a real advantage for anyone whose workload scales with threads, and much less relevant for anyone licensing software per core.Second, the memory configurations differ between compared systems in several footnotes, usually DDR5-8000 or MRDIMM-12800 for Venice against DDR5-6400 for the comparison part. That is a fair representation of what each platform supports, but it means the memory subsystem is doing part of the work in any bandwidth-sensitive result.“
Source: https://www.techpowerup.com/review/amd-zen-6-technical-deep-dive/
