Dell PowerEdge C6520 vs C6620 — Density Compute and HPC Sleds
The C-series is Dell’s density compute line — four 2-socket sleds in a 2U chassis (C6400 for 15G, C6600 for 16G). The target customer is HPC, CDN, large-scale web infrastructure, and hyperscale-style workloads where compute-per-rack-U is the metric that matters. This guide covers the C6520 (15G, Ice Lake) versus the C6620 (16G, Sapphire/Emerald Rapids).
Density by design
| C6520 sled + C6400 chassis | C6620 sled + C6600 chassis | |
|---|---|---|
| Generation | 15G | 16G |
| Released | March 2021 | January 2023 |
| Chassis size | 2U (C6400) | 2U (C6600) |
| Sleds per chassis | 4 | 4 |
| Sockets per sled | 2 | 2 |
| Sockets per 2U | 8 | 8 |
| CPU | Xeon Scalable Gen 3 (Ice Lake-SP) | Xeon Scalable Gen 4/5 (Sapphire / Emerald Rapids) |
| Max cores per sled | 80 (dual 40-core) | 128 (dual 64-core) |
| Max cores per 2U chassis | 320 | 512 |
| DIMM slots per sled | 16 | 16 |
| Memory per sled | 2TB DDR4 | 4TB DDR5 |
| Drive bays per sled | 6 SFF / 4 LFF | 8 SFF NVMe (E3.S option) |
| PCIe | Gen 4 | Gen 5 |
| Chassis PSU | 2 × 1600W or 2400W | 2 × 2800W Titanium |
| OEM support | EOS 2024-09-30, EOSL 2028-06-30 | Active |
Density economics
The C-series is purchased by customers optimising for cores-per-U and dollars-per-core at scale. In one 2U chassis:
- C6520 (full populate): 8 sockets × 40 cores = 320 cores, 8TB RAM, 24 drive bays.
- C6620 (full populate): 8 sockets × 64 cores = 512 cores, 16TB RAM, 32 drive bays all NVMe.
Contrast with stacking 8 × R650 (15G) in 8U: same core count but 4× the rack space. Or 8 × R660 (16G) — 8U to match the C6620’s 2U. Rack rent, power distribution, and cooling all favour C-series density at scale.
Where C-series is the wrong answer
- Mixed workload — the sleds are less flexible per-host than rack servers. Fewer PCIe slots, limited local storage, no GPU.
- Small deploys under 4 sleds — you are paying for chassis you don’t fill.
- Regulated enterprises that want per-host physical separation — sleds share chassis PSU and fabric.
- Any GPU attachment — C-series is CPU-density first, GPU support is minimal.
Memory gap
C6520: DDR4-3200 RDIMM, 16 DIMMs per sled, 2TB ceiling. PN samples: 6VDNY, 74FPM, 75X1V, HF6GX.
C6620: DDR5-4800/5600 RDIMM, 16 DIMMs per sled, 4TB ceiling. PN samples: AC239377, AC239378, AC239379, AC448843.
The DDR5 bandwidth jump matters for HPC codes that live in memory — finite-element simulations, genomics pipelines, large-scale CFD.
Storage — the real architectural shift
C6520 sleds ship with 6 × SFF or 4 × LFF local drives. SAS or SATA, with NVMe available on a subset of SKUs via direct-attach. Use case: OS boot + scratch space + local cache. Primary storage lives on external SAN / parallel filesystem.
C6620 sleds go all-NVMe, up to 8 × SFF or E3.S. Local-NVMe-first is the new model, and the sled-per-chassis NVMe count doubled. For Lustre or BeeGFS deployments where each node contributes local-NVMe to a distributed namespace, C6620 is a meaningful capability jump.
Representative NVMe PNs: 08M01, 0D4GH, 0HVC7, 0MNMV. E3.S: 1VK3C, 3NVF2, 831H5, G27W5.
Power distribution
C6400 chassis takes 2 × 1600W or 2 × 2400W shared PSU. C6600 takes 2 × 2800W Titanium. Redundancy is at chassis level, not per-sled — a failed PSU affects all 4 sleds’ power delivery margin. For most HPC customers this is acceptable; for critical commercial workloads, rack is the safer choice.
Stocked PSU PNs: 095HR5, 685W7, 95HR5, 0GDXX, 450-AGFW, 450-AKKZ, 7GN3C.
Networking
C-series typically ships with high-density 25/100 GbE OCP NIC 3.0 mezzanines per sled. InfiniBand / Omni-Path options exist for HPC customers needing low-latency fabric. C6620 supports PCIe 5.0 NICs — 400 GbE single-port or 2 × 200 GbE become realistic for the first time in this form factor.
Procurement realities
C-series is lower-volume than rack — both new and refurb channels are thinner. Chassis + sled pairing matters: a bare C6620 sled without the C6600 chassis is inert. ICD stocks both chassis types in Dubai for GCC and North Africa HPC customers, with lead times 7–14 days on full 4-sled configurations.
Decision framework
- HPC cluster, greenfield, 32+ nodes → C6620. DDR5 + PCIe 5.0 + all-NVMe pays back immediately.
- HPC cluster, scale-out of existing C6520 fleet → C6520 (keep fleet consistent) until a full refresh cycle.
- Hyperscale-style web/CDN at 100+ nodes → C6620 for density efficiency.
- Anything mixed-workload or under 4 nodes → skip C-series, use R660/R760.
For HPC / density-compute quotes, submit RFQ with node count, CPU SKU target, and interconnect. ICD typically responds same-day for C-series pricing.