Marvell OCTEON 10 (CN10K)

Marvell OCTEON 10 (CN10K)

The Marvell OCTEON 10 (silicon code CN10K) is a programmable network processor / DPU (Data Processing Unit), not a merchant switch ASIC. It combines ARM64 Neoverse N2 CPU cores with network I/O, crypto, and scheduling engines on TSMC 5 nm silicon. Several ODM platforms use it for carrier edge routing, 5G gateways, and uCPE appliances.

Field Value
Silicon family Marvell OCTEON 10 (CN10K)
Process node TSMC 5 nm
CPU cores 8–36 Arm Neoverse N2 (ARMv9)
Packet buffer External DDR5 DRAM (software-managed)
On-chip SRAM packet buffer Not published — no public figure from Marvell
DDR5 channels 2 (CN102/CN103) to 12 (DPU400)
Max forwarding bandwidth ~1 Tbps (CN106S integrated switch) / ~200 Gbps (CN103)
Target use cases Carrier edge routing, 5G gateways, uCPE, SmartNIC/DPU

Buffer Architecture

CN10K's buffer model is fundamentally different from switch ASICs like Broadcom Trident or Tomahawk, which have a fixed-size on-chip SRAM packet buffer pool.

How packets are buffered

The CN10K uses three internal blocks to manage packet flow:

Packets in flight reside in external DDR5 DRAM, not in a fixed on-chip SRAM pool. The NPA maintains rings of free buffer pointers; NIX DMA engines move packets directly into DRAM buffers. The on-chip L2/L3 cache (8–72 MB depending on variant) acts as a hot-path cache for recently accessed buffer regions, not as a dedicated packet FIFO.

Effective packet buffer capacity = installed DDR5 DRAM, which is configurable at platform design time:

Platform CN10K Variant DDR5 Capacity Notes
Asterfusion ET3608 CN103XX 16 GB (expandable to 48 GB) 2x 100G QSFP28
Asterfusion ET3600 2× CN103XX ~32 GB (16 GB per NP) 4x 100G QSFP28

At 100 Gbps line rate (~12.5 GB/s), even the base 16 GB DDR5 configuration provides over a 1-second buffer — orders of magnitude more than any on-chip SRAM switch ASIC.

DDR5 bandwidth headroom

A CN103XX with dual DDR5-5600 channels provides ~89 GB/s of peak memory bandwidth. At 200 Gbps aggregate throughput (~25 GB/s), there is substantial headroom for simultaneous packet reads, writes, and descriptor updates without memory contention.

Latency tradeoff

The DDR5 buffer model trades latency for capacity. DDR5 access latency is ~80–100 ns, compared to ~5–10 ns for on-chip SRAM. For latency-sensitive ultra-low-latency trading or HPC workloads this matters, but for carrier edge/routing applications where deep buffering and programmability are the priority, DDR5 is the right tradeoff.

Comparison: CN10K vs. Switch ASICs

Dimension Broadcom Tomahawk 5 Broadcom Jericho 3 Marvell OCTEON 10 CN103
Architecture Fixed-function switch ASIC Fixed-function NP (VOQ) Programmable NP/DPU
CPU cores None None Up to 8 Neoverse N2
On-chip SRAM packet buffer 165.2 MB ~48 MB OCB Not published
Off-chip buffer None Up to 32 GB HBM3 16–48 GB DDR5 (external)
Max bandwidth 51.2 Tbps 3.2 Tbps ~200 Gbps
Target Hyperscale DC switching SP core / deep buffer Carrier edge / uCPE / DPU
Protocol programmability Limited (P4 pipeline) Limited Full (DPDK / VPP / eBPF)

Data Quality Note

Marvell does not publish an on-chip packet buffer SRAM size (in MB) for any CN10K variant in open product documentation. The L2/L3 cache figures in product briefs (8–72 MB) are CPU caches, not packet buffer pools. All buffer capacity figures on this site for CN10K platforms refer to the external DDR5 DRAM size, which is confirmed from platform vendor specifications.

See Also

← Back to the main buffer table