Why we're publishing this: "enterprise CPU" is unverifiable marketing until a model number is attached. Every figure below is from the Intel or AMD spec sheet, and the node assignments are what our Proxmox inventory actually reports — you can cross-check all of it.
The fleet, spec for spec
| NODE | CPU | CORES/THR | CLOCK | MEMORY | L3 |
|---|---|---|---|---|---|
| Delhi N3 | AMD EPYC 7C13 (Milan, Zen 3) | 64C / 128T | 2.0 / 3.675 GHz | 8-ch DDR4-3200 | 256 MB |
| Mumbai N1 | Xeon Platinum 8269CY (Cascade Lake) | 26C / 52T | 2.5 / 3.8 GHz | 6-ch DDR4-2666 | 35.75 MB |
| Delhi N1 | Xeon Gold 6138 (Skylake-SP) | 20C / 40T | 2.0 / 3.7 GHz | 6-ch DDR4-2666 | 27.5 MB |
| Delhi N2 (Noida) | Xeon E5-2699 v4 (Broadwell) | 22C / 44T | 2.2 / 3.6 GHz | 4-ch DDR4-2400 | 55 MB |
What each chip is actually good at
EPYC 7C13 — the density and bandwidth node. A Zen 3 cloud-SKU part: 64 cores with the strongest per-core IPC in the fleet, 8 memory channels (~205 GB/s theoretical), 256 MB of L3 and PCIe Gen4. Database engines, compilation, and anything memory-bandwidth-bound do best here. It's also the youngest architecture in the rack (2021 vs 2016–2019 for the Xeons) — better per-watt, better per-core.
Xeon Platinum 8269CY — the single-thread sprinter. A Cascade Lake cloud-custom SKU (built for hyperscale buyers, never on Intel's public price list) with the fleet's highest boost at 3.8 GHz and Speed Select profiles. Game-server ticks — Minecraft, FiveM — are single-thread-bound, which is why it fronts Mumbai. Node carries 251 GB RAM with SSD-RAID pools.
Xeon Gold 6138 — the AVX-512 workhorse. Two AVX-512 FMA units per core (the E5 and EPYC have none); video transcode, compression and vector workloads behave differently on it. Delhi N1 backs it with a dedicated NVMe pool.
Xeon E5-2699 v4 — the capacity engine. The oldest silicon here (Broadwell, 2016) and honest about it: it's paired with 504 GB of node RAM — the largest in the fleet — making it the RAM-density node for big pooled workloads where thread count beats thread speed.
How to verify instead of trusting this page
Node CPUs are printed on our Mumbai and Delhi location pages, pulled live from Proxmox — if hardware changes, the page changes. Inside any VM, lscpu reports the real model and top's steal column reports real contention. For game workloads, /tps at your peak hour is the benchmark that matters more than any of our tables.