AMD EPYC™ 9005 Series Processors

The AMD EPYC 9005 Series is the 5th generation of AMD’s server processor family, codenamed “Turin”, offering 8 to 192 cores built on the “Zen 5” and “Zen 5c” architectures. Launched in October 2024 on the widely deployed SP5 socket, it covers cloud, enterprise, HPC and AI workloads, and includes one model built specifically to keep expensive GPUs busy.

Key benefits at a glance

Benefit	Why it matters
Up to 192 cores and 384 threads	More virtual machines and containers per socket, which is the whole game in cloud and hosting
Up to 2.3x the integer performance of competing CPUs	AMD reports leadership on SPECrate®2017_int_base, the benchmark most enterprise buyers reference
Up to 86% fewer racks for equivalent performance	Consolidating a legacy estate cuts power, cooling, floor space and per-core software licensing all at once
Same SP5 socket as EPYC 9004	Most models drop into validated platforms you already own
12-channel DDR5 and 6TB per socket	Memory capacity and bandwidth for in-memory databases, analytics and virtualisation
128 PCIe® Gen 5 lanes with CXL 2.0	Room for GPUs, NVMe and high-speed networking without contention
A dedicated GPU host CPU option	The 64-core EPYC 9575F boosts to 5GHz to keep accelerators fed and utilisation high

Full specifications

Specification	AMD EPYC 9005 Series
Core architecture	“Zen 5” (4nm) and “Zen 5c” (3nm)
Core count	8 to 192 cores, up to 384 threads
Socket	SP5
Memory	12-channel DDR5 up to 6400 MT/s
Maximum memory	Up to 6TB per socket
I/O	128 PCIe® Gen 5 lanes, CXL 2.0
Max L3 cache	Up to 512MB
Max TDP	500W
Sockets supported	1P and 2P

“Zen 5” or “Zen 5c”: which core do you need?

Unusually, both core types sit in the same product stack, so the choice comes down to what your workload rewards.

• “Zen 5” is the performance-oriented design on a 4nm process. It scales to 128 cores and suits general-purpose enterprise workloads, HPC and anything licensed per core, where you want the most work out of each one.
• “Zen 5c” is the density-optimised design on 3nm. It scales to 192 cores and 384 threads, aimed at cloud and scale-out workloads where throughput per rack beats single-thread speed.

Both benefit from roughly a 17% IPC uplift over “Zen 4”, plus a full 512-bit datapath for AVX-512 with an optional double-pumped mode that trades a little peak throughput for better efficiency.

What the performance figures mean in practice

AMD reports the following for EPYC 9005:

• Up to 2.3x the integer performance of leading competitive offerings on SPECrate®2017_int_base
• 33% more vCPUs from the 192-core EPYC 9965 than a 144-core Intel® Xeon® 6E “Sierra Forest” processor
• Up to 2x inference throughput versus the previous EPYC generation on CPU-hosted AI workloads
• Up to 86% fewer racks to match the integer performance of a legacy estate

That last figure is usually the one that carries a business case. Fewer racks means less power, less cooling, fewer software licences and less floor space, which is why consolidation remains the most reliable route to a defensible TCO argument.

Which EPYC 9005 model should you choose?

Model	Cores	Core type	Best suited to
EPYC 9965	192	“Zen 5c”	Maximum density; 384MB L3, 500W
EPYC 9755	128	“Zen 5”	Top “Zen 5” performance part; 512MB L3, 500W
EPYC 9575F	64	“Zen 5”	GPU host duties; boosts to 5GHz

The EPYC 9575F deserves a closer look if you are buying GPU servers. In AMD testing across eight inference models, an eight-GPU server hosted by dual EPYC 9575F CPUs delivered up to 13% faster time to first token and 6.6% higher inference throughput than the equivalent Intel Xeon 6960P system. The host CPU is a small fraction of the bill of materials in a GPU node, but it has an outsized influence on how much of your accelerator investment you actually use.

You do not always need a GPU

Plenty of AI work runs perfectly well on CPUs alone. Language models around 13 billion parameters and below, fraud analysis, image classification and recommendation systems all sit comfortably on CPU-only servers. Add EPYC 9005’s core density, memory bandwidth and native x86 compatibility, and it becomes a sensible platform for agentic AI workloads that orchestrate tools, retrieve information and monitor progress rather than train frontier models.

Will it fit your existing platform?

Almost certainly. EPYC 9005 uses the same SP5 socket as the 4th Gen EPYC 9004 Series, so most models drop into existing validated platforms with a firmware update. The one exception is the 192-core EPYC 9965, which needs a board designed for its power delivery requirements.