Netcup Generation 12.5 vs G11 & G12 Architecture Guide: EPYC 9645 Zen 5, Storage & Pricing

On Netcup Root Servers (RS 500 through RS 8000), each allocated core maps to a dedicated physical processor thread pinned on an enterprise AMD EPYC host node, rather than the oversubscribed time-sliced vCores used on shared VPS lines. Across the lifespan of Generation 11 (G11), early G11 deployments utilized the 64-core AMD EPYC 7702P ("Rome" 7nm Zen 2 with 8-channel DDR4-3200 ECC) before late-G11 clusters introduced the 84-core AMD EPYC 9634 ("Genoa" 5nm Zen 4 with 12-channel DDR5-4800 ECC). In Generation 12 (G12) and Generation 12.5 (G12.5), Netcup standardized its entire Root Server fleet on the 96-core / 192-thread AMD EPYC 9645 ("Turin" 4nm TSMC N4P Zen 5).

Architecturally, the transition from EPYC 9634 (G11 Genoa) to EPYC 9645 (G12/G12.5 Turin) resolves an important Core Complex Die (CCD) symmetry bottleneck. The 84-core EPYC 9634 populated 12 CCDs with 7 active cores per CCD (32 MB L3 cache shared across 7 cores), whereas the 96-core EPYC 9645 populates all 12 CCDs with a uniform 8 active cores per CCD (32 MB L3 cache per 8-core CCD, 384 MB L3 total) running at a 2.40 GHz base clock and 3.70 GHz boost clock. When Netcup’s KVM hypervisor pins a 4-core RS 1000 G12.5 or an 8-core RS 2000 G12.5 onto a single CCD, all guest cores communicate across a unified 32 MB L3 cache domain without crossing the central I/O Die (IOD) Infinity Fabric.

For backend engineers, database administrators, and AI/ML practitioners, three microarchitectural changes inside the Zen 5 core on EPYC 9645 directly improve guest VM latency and throughput:

• ▪ Full 512-Bit AVX-512 Execution Engine: In Zen 4 (G11), 512-bit vector instructions were cracked into two 256-bit halves (double-pumped). In Zen 5 (G12/G12.5), AMD engineered a true 512-bit floating-point and vector datapath. Workloads utilizing AVX-512 VNNI/BF16—such as PostgreSQL pgvector HNSW indexing, Qdrant/Milvus vector similarity search, llama.cpp CPU inference, and OpenSSL/WireGuard cryptographic routines—see 18% to 35% higher single-core throughput.
• ▪ Wider Front-End Dispatch & 48 KB 12-Way L1 Data Cache: Zen 5 widens the instruction decode and dispatch pipeline from 6-wide to 8-wide and increases the L1 Data Cache by 50% (from 32 KB 8-way to 48 KB 12-way per core), significantly reducing L1 misses on high-concurrency Node.js, Go, and Rust microservices.
• ▪ 12-Channel DDR5-6000 ECC Memory Subsystem: While Genoa (G11) operated 12 channels of DDR5-4800 ECC (460.8 GB/s theoretical peak bandwidth per socket), Turin (G12/G12.5) supports 12 channels of DDR5-6000 ECC (up to 576 GB/s per socket), preventing memory-bus starvation when 24 customers on the same host node run concurrent Redis or PostgreSQL queries.

Why did Netcup introduce a .5 half-generation (G12.5) so shortly after launching G12? While G11 → G12 was a pure hardware upgrade (swapping EPYC 9634 for EPYC 9645 while keeping large NVMe allocations), G12 → G12.5 was a commercial and storage re-indexing driven by sharply rising enterprise PCIe 4.0/5.0 TLC/eTLC NAND flash prices and European datacenter power tariffs.

Specifically, in Generation 12.5, Netcup kept the exact same EPYC 9645 Zen 5 processor and DDR5 ECC RAM sizes per tier, but halved the base NVMe disk capacity across Root Server and VPS tiers and adjusted monthly contract pricing (most visibly on RS 2000, RS 4000, and RS 8000):

Examine the G12.5 voucher structure carefully and a striking architectural advantage emerges: RS 1000 G12.5 is the only Root Server tier granted a 2-Months-Free voucher (6874nc..., a 16.67% Year-1 reduction), whereas higher tiers receive 1 Month Free (8.33%) or €5.00 OFF.

In G12.5, RS 1000 costs €18.26/mo net / €21.73/mo gross (€15.22/mo net / €18.11/mo gross effective in Year 1 with the 2-Months-Free voucher) for 4 dedicated Zen 5 cores, 8 GB DDR5 ECC RAM, and 128 GB NVMe. Stepping up to RS 2000 G12.5 doubles the hardware (8 cores, 16 GB RAM, 256 GB NVMe) at €34.20/mo net / €40.70/mo gross (€31.35/mo net / €37.31/mo gross effective with 1 Month Free).

Consequently, two horizontal RS 1000 G12.5 servers (totaling 8 dedicated Zen 5 cores, 16 GB DDR5 ECC RAM, 256 GB NVMe across 2 independent nodes, and 2× 2.5 Gbit/s network uplinks) cost €30.44/mo net (€36.22/mo gross) effective in Year 1 using two 2-Months-Free vouchers (6874nc...). That is cheaper than a single RS 2000 G12.5 (€31.35/mo net / €37.31/mo gross) while giving you multi-node high availability, two independent IPv4 addresses, and 5.0 Gbit/s aggregate throughput!

When choosing between VPS 2000 G12.5 (€22.62/mo net / €26.92/mo gross) and RS 1000 G12.5 (€18.26/mo net / €21.73/mo gross), engineers must weigh CPU scheduling guarantees against total RAM allocation. On Root Servers, KVM pins guest vCPUs to dedicated physical EPYC 9645 threads with hardware-isolated cache behavior, whereas VPS instances share host CPU time slices across neighboring tenants:

• ▪ Choose RS 1000 G12.5 (€15.22/mo net / €18.11/mo gross effective Year 1) when CPU latency predictability, sustained multi-core compilation, high PPS network routing, or AVX-512 throughput matters. Its 4 cores are dedicated EPYC 9645 threads with near-zero %steal time.
• ▪ Choose VPS 2000 G12.5 (€20.74/mo net / €24.68/mo gross effective Year 1) or VPS 4000 G12.5 (€34.93/mo net / €41.58/mo gross effective Year 1) when your workload is memory-capacity bound rather than CPU-bound (for example, running 15–25 low-CPU Docker containers, Java/Spring Boot heaps, or in-memory caches that need 16 GB to 32 GB of RAM at a fraction of the RS 2000/4000 price).

Once your server boots, you can verify whether your KVM guest is running on a G12/G12.5 AMD EPYC 9645 ("Turin") host or an older G11 AMD EPYC 9634 ("Genoa") / EPYC 7702P ("Rome") host, inspect AVX-512 instruction flags, and verify SCSI UNMAP discard support using the following diagnostic commands.

When inspecting lscpu output inside a Netcup KVM guest, pay close attention to the CPU family and cache topology fields: AMD Zen 5 ("Turin") reports CPU family: 26 (0x1A) with 48 KiB L1d and 1 MiB L2 cache per core, whereas Zen 4 ("Genoa") reports CPU family: 25 (0x19) with 32 KiB L1d cache, and Zen 2 ("Rome") reports CPU family: 23 (0x17) without AVX-512 support:

# 1. Inspect the exposed CPU model, family (Family 26 = Zen 5 Turin; Family 25 = Zen 4 Genoa), and caches:
lscpu | grep -E "Model name|Vendor ID|CPU family|Model:|Thread|Core|Socket|L1d|L2|L3|Flags"

# 2. Verify native AVX-512 instruction set extensions exposed to the KVM guest:
grep -o -E 'avx512[a-z0-9_]*' /proc/cpuinfo | sort -u | tr '\n' ' '; echo ""

# 3. Inspect QEMU/KVM SMBIOS chassis & block device geometry:
sudo dmidecode -t system -t processor | grep -E "Manufacturer|Product Name|Version"
lsblk -o NAME,SIZE,TYPE,FSTYPE,DISCLAYER,MODEL

Vendor ID:                AuthenticAMD
Model name:               AMD EPYC 9645 96-Core Processor
CPU family:               26
Thread(s) per core:       1
Core(s) per socket:       4
L1d cache:                192 KiB (4 instances)
L2 cache:                 4 MiB (4 instances)
L3 cache:                 32 MiB (1 instance)
avx512_bf16 avx512_bitalg avx512_vbmi avx512_vbmi2 avx512_vnni avx512_vpopcntdq avx512bw avx512cd avx512dq avx512f avx512ifma avx512vl