PicOS N8650-32OD — 32 x 800G AI/HPC Switch
An AI/HPC data centre switch with 32 x 800G OSFP ports on Broadcom Tomahawk 5. Lossless fabric with RoCEv2, PFC and ECN.
PicOS N8650-32OD — 32 x 800G AI/HPC Switch
An AI/HPC data centre switch with 32 x 800G OSFP ports on Broadcom Tomahawk 5. Lossless fabric with RoCEv2, PFC and ECN.
- 32 x 800G OSFP ports, 25.6 Tb/s switching capacity (1U)
- A 64-port, 51.2 Tb/s variant (2U) in the same family
- Broadcom Tomahawk 5 silicon
- Lossless AI fabric with RoCEv2, PFC, ECN and MLAG
The N8650-32OD is a 1U AI/HPC switch built on Broadcom Tomahawk 5 silicon: 32 x 800G OSFP ports and 25.6 Tb/s of switching capacity. The 2U model in the same family arrives with 64 ports and 51.2 Tb/s.
In this class of hardware, performance is determined by congestion management rather than raw capacity. RoCEv2 running lossless depends on PFC and ECN being tuned to the workload; a fabric left at defaults keeps GPUs waiting on packet loss. Mevasis applies those settings during installation and reports them with NCCL measurements.
Specifications
| Specification | N8650-32OD 32 x 800G, 1U | 64-port variant 64 x 800G, 2U |
|---|---|---|
| Switching | ||
| Ports | 32 x 800G OSFP | 64 x 800G OSFP |
| Switch chip | Broadcom BCM78902 Tomahawk 5 | Broadcom BCM78900 Tomahawk 5 |
| Switching capacity | 25.6 Tb/s | 51.2 Tb/s |
| Lossless transport and L3 | ||
| RoCEv2 | Yes | Yes |
| PFC (priority flow control) | Yes | Yes |
| ECN (explicit congestion notification) | Yes | Yes |
| MLAG | Yes | Yes |
| OSPFv2 / v3 | Yes | Yes |
| Physical and electrical | ||
| Dimensions (H x W x D) | 44 x 438.4 x 589 mm (1U) | 86.1 x 440 x 760 mm (2U) |
| Weight | 14.53 kg | 26 kg |
| Power supplies | 2 (1+1 redundant), hot-swap, AC | 2 (1+1 redundant), hot-swap, AC |
| Fans | 7 (6+1 redundant), front-to-back | 4 (3+1 redundant), front-to-back |
| Max. power consumption | 2,400 W | 2,980 W |
Specifications follow the manufacturer datasheet. The exact configuration is confirmed in your quote.
Connectivity Solutions
AI compute fabric spine
The 800G ports serve the spine layer where 400G links from GPU nodes are aggregated through breakout. The 32-port 1U model acts as spine in mid-sized clusters and as leaf in large ones; the 64-port 2U model collapses the spine into a single layer at 51.2 Tb/s.
Lossless RoCEv2 configuration
Unless PFC priority classes and ECN thresholds are tuned to the workload, RoCE falls well short of expected performance. Tomahawk 5's buffer architecture amplifies the effect of those settings; we validate them with NCCL tests during commissioning.
Alongside the storage plane
Separating storage traffic from the compute fabric removes unpredictable latency from training jobs. A 100G-class storage plane is sufficient in most deployments and leaves the 800G spine to GPU traffic alone.
Features
51.2 Tb/s in a single box with Tomahawk 5
High capacity lowers the switch count at the spine layer and with it the number of hops. Fewer hops mean more predictable latency in distributed training.
Flexible port planning through breakout
One 800G port splits into two 400G or four 200G links, which makes it possible to aggregate nodes of different speed tiers onto the same spine.
Power and cooling planning is mandatory
Draw approaching 2,400 W directly affects the per-rack power budget. Front-to-back airflow suits hot-aisle / cold-aisle layouts; rack placement is planned together before ordering.
Open network operating system
PicOS decouples hardware from software; the switch is managed with the CLI and automation tooling your team already knows, with no single-vendor lock-in.
Resources
Documents open on the manufacturer's or supplier's own site.
Request a quote for this product
Tell us the quantity and the cluster it will join; we send a compatible part list with lead times — usually the same business day.
- Authorised dealer supply, manufacturer warranty
- Compatibility check for switch, adapter and optics
- Lead time per line item
Related products
Let's Build Future Together.
