Home · Technology · Broadcom Silicon
OcNOS x Broadcom merchant silicon

The Broadcom silicon behind your OcNOS switches and routers

Two Broadcom families power open networking. IP Infusion turns them into complete, supported systems: OcNOS-DC on the Tomahawk and Trident Ethernet fabrics, and OcNOS-SP on the deep-buffer Jericho and Qumran routers. Here is how the families differ, the generations and their advantages, and the qualified boxes you can buy today.

See qualified hardware (HCL) How the families differ

How Broadcom's silicon families differ

StrataXGS (Tomahawk, Trident) uses a shallow on-chip shared buffer for speed and high radix; StrataDNX (Jericho, Qumran) uses a deep buffer with VOQ scheduling for scale and lossless routing. One choice drives almost everything: where the packet buffer lives.

StrataXGSOcNOS-DC
Tomahawk & Trident

Shallow buffer, high radix, low latency. A shared on-chip SRAM pool absorbs microbursts while a programmable pipeline forwards at line rate. This is the switch line for data-center and AI scale-out fabrics.

Packet buffer
Shared on-chip, no HBM
Fabric
Clos with ECMP + dynamic LB
Route scale
On-chip tables
Best for
AI scale-out, DC leaf/spine
Tomahawk - speed, radixTrident - programmable
StrataDNXOcNOS-SP
Jericho, Qumran & Ramon

Deep buffer, VOQ scheduling, cell fabric. Deep packet buffering plus a credit-scheduled fabric give near-perfect load balancing and internet-scale routing. This is the router line for carrier core, edge and DCI.

Packet buffer
Deep, HBM or external DRAM
Fabric
Scheduled cell fabric or single-chip
Route scale
External TCAM, millions of routes
Best for
Carrier core, edge, DCI
Jericho - deep-buffer routerQumran - single-chip edgeRamon - cell fabric

Choose XGS when you need speed, radix and low latency for a switching fabric. Choose DNX when you need deep buffers, huge route tables and scheduled routing. No Tomahawk or Trident has HBM, and on DNX the route table lives in external TCAM, separate from the HBM buffer.

Each generation, and what it buys you

Each track climbs in bandwidth from left to right. Every Broadcom switch and router ASIC we track is listed, tagged with what OcNOS runs today and what is coming next.

TomahawkAI and data-center switching
Tomahawk
3.2T

First StrataXGS with 25G SerDes; set the high-radix, low-latency template.

Earlier generation
Tomahawk 2
6.4T

Doubled the radix to 64x100G for denser spine and leaf.

On OcNOS today
Tomahawk 3
12.8T

First 12.8T merchant switch; set the RoCEv2 AI scale-out pattern.

On OcNOS today
Tomahawk 4
25.6T

Mainstream 400G AI leaf and spine.

On OcNOS today
Tomahawk 5
51.2T

The 800G AI workhorse: 64x800G with hardware Cognitive Routing.

On OcNOS today
Tomahawk Ultra
51.2T

Scale-up Ethernet: 250ns latency and in-network collectives for XPU-to-XPU.

Tomahawk 6
102.4T

World-first single-chip 102.4T; one part for scale-out and scale-up.

Coming to OcNOS
Tomahawk 6-Davisson
102.4T

Co-packaged optics: about 70% less interconnect power than pluggables.

TridentProgrammable DC leaf
Trident 2
1.28T

Pre-AI 40G top-of-rack. Historical baseline.

Earlier generation
Trident 2+
1.28T

Virtualization-era leaf; single-pass routing for VXLAN overlays.

Earlier generation
Trident 3
3.2T

Programmable 25G/100G DC leaf workhorse.

On OcNOS today
Trident 4
12.8T

400G leaf with IPoDWDM (400G ZR+) headroom.

On OcNOS today
Trident 5-X12
16T

On-chip neural net (NetGNT) spots AI congestion in real time.

JerichoDeep-buffer routing
Jericho2
10T

The original deep-buffer DDC line card; pairs with Ramon.

Earlier generation
Jericho2c
4.8T

Cost-per-port-optimized routing edge and DDC line card.

Earlier generation
Jericho2c+
14.4T

Carrier core: deep HBM buffer and internet-scale FIB via external TCAM.

On OcNOS today
Jericho3
28.8T

800G deep-buffer routing with line-rate MACsec and IPsec.

Jericho3-AI
28.8T

Scheduled DDC AI fabric scaling to 32,768 GPUs at 800G.

Jericho4
51.2T

Scale-across: lossless RoCE over 100km-plus between sites.

QumranAccess, aggregation, edge and DCI
Qumran-UX
120G

Compact, fanless, temperature-hardened cell-site and access silicon.

On OcNOS today
Qumran-AX
300G

Access and aggregation with 25G ports and 100G uplinks.

On OcNOS today
Qumran2u
360G

Compact, low-power cell-site and access router silicon.

On OcNOS today
Qumran-MX
800G

Dense 10G access aggregation with 100G uplinks.

On OcNOS today
Qumran2a
800G

Deep-buffer SP edge and cell-site router with PTP/SyncE timing.

On OcNOS today
Qumran2c
2.4T

Single-chip edge, aggregation and DCI router.

On OcNOS today
Qumran2c+
7.2T

Deep-buffer aggregation and DCI in one chip, 6x400G class.

On OcNOS today
Qumran3U
1.2T

Qumran3 access and aggregation tier at 1.2T.

Coming to OcNOS
Qumran3D
25.6T

Single-chip 25.6T deep-buffer router with 800G ZR.

Coming to OcNOS
RamonCell fabric elements for DDC
Ramon
9.6T

Cell-based spine element for Jericho2-class DDC fabrics.

Earlier generation
Ramon 3
51.2T

Spine element of the scheduled Jericho3-AI DDC fabric.

Ramon 4
Not disclosed

Scale-across spine element for the Jericho4 cross-site fabric.

The silicon we support, and the boxes you can buy

Every ASIC in the cards below runs OcNOS today, with the IP Infusion qualified open platforms built on it. The full spec table also lists the silicon we are bringing to OcNOS next, marked Planned. Filter by OcNOS variant or use case.

OcNOS
Use
Trident 3
StrataXGS · Trident
OcNOS-DC
BCM56870 / BCM56873 (TD3-X7) / BCM56770 (TD3-X5) / BCM56771 (TD3-X4) / BCM56277 (TD3-X2)
Bandwidth2.0-3.2 Tbps (by SKU)
Radixup to 32x100GbE
BufferShared on-chip (SmartBuffer) (32 MB)

Programmable DC leaf workhorse. IPI ships it in four sub-variants (X2/X4/X5/X7).

IP Infusion qualified boxes 10
Edgecore AS7726-32X, 32x100G, on Broadcom Trident 3 AS7726-32X Edgecore
UfiSpace S9110-32X, 32x100G, on Broadcom Trident 3 S9110-32X UfiSpace
Edgecore AS7326-56X, 8x100G, 48x25G, on Broadcom Trident 3 AS7326-56X Edgecore
UfiSpace S8901-54XC, 6x100G, 48x25G, on Broadcom Trident 3 S8901-54XC UfiSpace
UfiSpace S8901-54XC-16G, 6x100G, 48x25G, on Broadcom Trident 3 S8901-54XC-16G UfiSpace
Edgecore AS5835-54X, 6x100G, 48x10G, on Broadcom Trident 3 AS5835-54X Edgecore
Edgecore AS5835-54T, 6x100G, 48x10G, on Broadcom Trident 3 AS5835-54T Edgecore
Edgecore AS4625-54T, 6x10G, 48x1G, on Broadcom Trident 3 AS4625-54T Edgecore
Celestica DS1000, 8x10G, 48x1G, on Broadcom Trident 3 DS1000 Celestica
UfiSpace S6301-56ST, 8x10G, 48x1G, on Broadcom Trident 3 S6301-56ST UfiSpace
Full Trident 3 deep-dive →
Trident 4
StrataXGS · Trident
OcNOS-DC
BCM56880 (Trident4-X11, 12.8T) / BCM56780 (separate 8.0T MACsec/IPsec variant)
Bandwidthup to 12.8 Tbps (X11)
Radix32x400GbE / 128x100GbE
BufferShared on-chip (single pool) (132 MB)

Compiler-programmable (NPL) 400G DC leaf with IPoDWDM headroom. The 12.8T part is BCM56880, not BCM56780.

IP Infusion qualified boxes 2
Edgecore AS9726-32DB, 32x400G, on Broadcom Trident 4 AS9726-32DB Edgecore
UfiSpace S9300-32D, 32x400G, on Broadcom Trident 4 S9300-32D UfiSpace
Full Trident 4 deep-dive →
Tomahawk 2
StrataXGS · Tomahawk
OcNOS-DC
BCM56970
Bandwidth6.4 Tbps
Radix64x100GbE
BufferShared on-chip (low latency)

Doubled radix to 64x100G for denser spine/leaf; superseded for AI by TH3/4/5.

IP Infusion qualified boxes 1
Edgecore AS7816-64X, 64x100G, on Broadcom Tomahawk 2 AS7816-64X Edgecore
Tomahawk 3
StrataXGS · Tomahawk
OcNOS-DC
BCM56980
Bandwidth12.8 Tbps
Radix32x400GbE / 128x100GbE
BufferShared on-chip (~450ns L3) (64 MB)

First 12.8T merchant switch. High radix + shallow buffer + RoCEv2 set the Ethernet AI scale-out pattern.

IP Infusion qualified boxes 1
Edgecore AS9716-32D, 32x400G, on Broadcom Tomahawk 3 AS9716-32D Edgecore
Tomahawk 4
StrataXGS · Tomahawk
OcNOS-DC
BCM56990
Bandwidth25.6 Tbps
Radix64x400GbE / 256x100GbE
Buffer~114 MB shallow shared, no HBM (~114 MB)

First 25.6T switch; mainstream 400G AI scale-out leaf/spine. Shallow shared buffer, part BCM56990 (not BCM56996).

IP Infusion qualified boxes 1
Edgecore AS9736-64D, 64x400G, on Broadcom Tomahawk 4 AS9736-64D Edgecore
Full Tomahawk 4 deep-dive →
Tomahawk 5
StrataXGS · Tomahawk
OcNOS-DC
BCM78900
Bandwidth51.2 Tbps
Radix64x800GbE / 128x400GbE
Buffershallow shared, size not disclosed

800G AI scale-out leaf/spine ASIC. 64x800G radix, hardware Cognitive Routing, RoCEv2 burst absorption.

IP Infusion qualified boxes 3
Edgecore AIS800-64D, 64x800G QSFP-DD, on Broadcom Tomahawk 5 AIS800-64D Edgecore
UfiSpace S9321-64E, 64x800G QSFP-DD, on Broadcom Tomahawk 5 S9321-64E UfiSpace
UfiSpace S9321-64EO, 64x800G, on Broadcom Tomahawk 5 S9321-64EO UfiSpace
Full Tomahawk 5 deep-dive →
Jericho2c+
StrataDNX · Jericho
OcNOS-SP
BCM88850 / BCM88851 (platform-specific)
Bandwidth14.4 Tbps
Radixup to 36x400GbE-class
BufferDeep HBM (VOQ); FIB in external TCAM (8 GB on AS9947-36XKB, 16 GB on S9610-36D)

Higher-capacity carrier-core / DDC line-card silicon. Internet-scale FIB via external TCAM (separate from the HBM packet buffer).

IP Infusion qualified boxes 2
Edgecore AS9947-36XKB, 12x400G, 24x100G, 4x10G, on Broadcom Jericho2c+ AS9947-36XKB Edgecore
UfiSpace S9610-36D, 36x400G, on Broadcom Jericho2c+ S9610-36D UfiSpace
Full Jericho2c+ deep-dive →
Qumran-UX
StrataDNX · Qumran
OcNOS-SP
BCM88273 / BCM88272 / BCM88270
Bandwidth32 - 120 Gbps
Radix1G / 10G compact CSR
BufferDeep - off-chip DRAM (single-chip, VOQ)

Compact cell-site and access router silicon for the far edge, in fanless and temperature-hardened platforms.

IP Infusion qualified boxes 7
UfiSpace S9501-28SMT, 8x10G, 20x1G, on Broadcom Qumran-UX S9501-28SMT UfiSpace
UfiSpace S9501-18SMT, 4x10G, 12x1G, on Broadcom Qumran-UX S9501-18SMT UfiSpace
UfiSpace S9502-16SMT, 4x10G, 12x1G, on Broadcom Qumran-UX S9502-16SMT UfiSpace
Edgecore AS5915-16X, 4x10G, 12x1G, on Broadcom Qumran-UX AS5915-16X Edgecore
UfiSpace S9502-12SM, 4x10G, 8x1G, on Broadcom Qumran-UX S9502-12SM UfiSpace
UfiSpace S9502-12SM-NT, 4x10G, 8x1G, on Broadcom Qumran-UX S9502-12SM-NT UfiSpace
Edgecore AS5915-18X, 6x10G, 12x1G, on Broadcom Qumran-UX AS5915-18X Edgecore
Qumran-AX
StrataDNX · Qumran
OcNOS-SP
BCM88470
Bandwidth300 Gbps
Radix10G / 25G access, 100G uplinks
BufferDeep - off-chip DRAM (single-chip, VOQ)

Deep-buffer access and aggregation router with 25G ports and 100G uplinks for the SP edge.

IP Infusion qualified boxes 3
UfiSpace S9500-22XST, 4x1G, 8x10G, 8x25G, 2x100G, on Broadcom Qumran-AX S9500-22XST UfiSpace
UfiSpace S9500-30XS, 20x10G, 8x25G, 2x100G, on Broadcom Qumran-AX S9500-30XS UfiSpace
Edgecore AS7315-27X, 20x10G, 4x25G, 3x100G, on Broadcom Qumran-AX AS7315-27X Edgecore
Qumran2u
StrataDNX · Qumran
OcNOS-SP
BCM88280 / BCM88284
Bandwidth~360 Gbps class
Radixno 400G; compact
BufferDeep off-chip DRAM (single-chip)

Compact / low-power cell-site or access router silicon. Two port-mix designs.

IP Infusion qualified boxes 2
Edgecore AS7515-24X, 4x100G, 20x25G, on Broadcom Qumran2u AS7515-24X Edgecore
UfiSpace S9510-30XC, 2x100G, 28x25G, on Broadcom Qumran2u S9510-30XC UfiSpace
Full Qumran2u deep-dive →
Qumran-MX
StrataDNX · Qumran
OcNOS-SP
BCM88375 / BCM88370
Bandwidth800 Gbps
Radix48x10G access, 6x100G uplinks
BufferDeep - off-chip DRAM (single-chip, VOQ)

Deep-buffer access and aggregation router: dense 10G access with 100G uplinks.

IP Infusion qualified boxes 2
Edgecore AS5916-54XKS, 48x10G, 6x100G, on Broadcom Qumran-MX AS5916-54XKS Edgecore
Edgecore AS5916-54XL, 48x10G, 6x100G, on Broadcom Qumran-MX AS5916-54XL Edgecore
Qumran2a
StrataDNX · Qumran
OcNOS-SP
BCM88480 / BCM88483
Bandwidth800 Gbps
Radix10-100GbE compact edge
BufferDeep external DRAM (single-chip, VOQ)

Standalone deep-buffer SP edge / cell-site router with PTP/SyncE timing and 400G ZR+ headroom. Runs OcNOS-SP.

IP Infusion qualified boxes 2
UfiSpace S9510-28DC, 2x400G, 2x100G, 24x25G, on Broadcom Qumran2a S9510-28DC UfiSpace
Edgecore AS7535-28XB, 2x400G, 2x100G, 24x25G, on Broadcom Qumran2a AS7535-28XB Edgecore
Full Qumran2a deep-dive →
Qumran2c
StrataDNX · Qumran
OcNOS-SP
BCM88820
Bandwidth2.4 Tbps
Radix10-400GbE
BufferOn-chip SRAM + off-chip HBM (VOQ) (16 MB + 4 GB HBM)

Single-chip deep-buffer edge / aggregation / DCI router. Distinct from Jericho2c (BCM88800).

IP Infusion qualified boxes 7
UfiSpace S9600-72XC, 8x100G, 64x25G, on Broadcom Qumran2c S9600-72XC UfiSpace
Edgecore AS7946-30XB, 4x400G, 22x100G, 4x25G, on Broadcom Qumran2c AS7946-30XB Edgecore
UfiSpace S9600-64X, 64x100G (dual ASIC, 4.8T), on Broadcom Qumran2c S9600-64X UfiSpace
UfiSpace S9600-56DX, 8x400G, 48x100G, 4x25G (dual, 4.8T), on Broadcom Qumran2c S9600-56DX UfiSpace
UfiSpace S9600-32X, 32x100G, on Broadcom Qumran2c S9600-32X UfiSpace
UfiSpace S9600-28DX, 4x400G, 24x100G, 4x25G, on Broadcom Qumran2c S9600-28DX UfiSpace
Edgecore AS7946-74XKSB, 10x100G, 64x25G, on Broadcom Qumran2c AS7946-74XKSB Edgecore
Qumran2c+
StrataDNX · Qumran
OcNOS-SP
BCM88840
Bandwidth7.2 Tbps
Radix6x400G, 40x100G class
BufferDeep (VOQ, single-chip)

Deep-buffer aggregation / DCI single-chip router.

IP Infusion qualified boxes 1
UfiSpace S9610-46DX, 6x400G, 40x100G, 4x25G, on Broadcom Qumran2c+ S9610-46DX UfiSpace
No silicon matches that filter. Reset to see the full list.
Show the full spec table, including silicon planned for OcNOS
ASICOcNOSPart(s)BandwidthNodeSerDesRadixBuffer
Trident 3 OcNOS-DC BCM56870 / BCM56873 (TD3-X7) / BCM56770 (TD3-X5) / BCM56771 (TD3-X4) / BCM56277 (TD3-X2) 2.0-3.2 Tbps (by SKU) 25G SerDes (up to 128 lanes) up to 32x100GbE Shared on-chip (SmartBuffer) (32 MB)
Trident 4 OcNOS-DC BCM56880 (Trident4-X11, 12.8T) / BCM56780 (separate 8.0T MACsec/IPsec variant) up to 12.8 Tbps (X11) 256x 50G PAM4 32x400GbE / 128x100GbE Shared on-chip (single pool) (132 MB)
Tomahawk 2 OcNOS-DC BCM56970 6.4 Tbps 256x 25G (NRZ) 64x100GbE Shared on-chip (low latency)
Tomahawk 3 OcNOS-DC BCM56980 12.8 Tbps 256x 50G PAM4 32x400GbE / 128x100GbE Shared on-chip (~450ns L3) (64 MB)
Tomahawk 4 OcNOS-DC BCM56990 25.6 Tbps 512x 50G PAM4 64x400GbE / 256x100GbE ~114 MB shallow shared, no HBM (~114 MB)
Tomahawk 5 OcNOS-DC BCM78900 51.2 Tbps 5 nm (monolithic) 512x 100G PAM4 64x800GbE / 128x400GbE shallow shared, size not disclosed
Jericho2c+ OcNOS-SP BCM88850 / BCM88851 (platform-specific) 14.4 Tbps 144x 50G net + 192x 50G fabric up to 36x400GbE-class Deep HBM (VOQ); FIB in external TCAM (8 GB on AS9947-36XKB, 16 GB on S9610-36D)
Qumran-UX OcNOS-SP BCM88273 / BCM88272 / BCM88270 32 - 120 Gbps Not disclosed 1G / 10G compact CSR Deep - off-chip DRAM (single-chip, VOQ)
Qumran-AX OcNOS-SP BCM88470 300 Gbps Not disclosed 10G / 25G access, 100G uplinks Deep - off-chip DRAM (single-chip, VOQ)
Qumran2u OcNOS-SP BCM88280 / BCM88284 ~360 Gbps class integrated 25G/50G no 400G; compact Deep off-chip DRAM (single-chip)
Qumran-MX OcNOS-SP BCM88375 / BCM88370 800 Gbps Not disclosed 48x10G access, 6x100G uplinks Deep - off-chip DRAM (single-chip, VOQ)
Qumran2a OcNOS-SP BCM88480 / BCM88483 800 Gbps integrated 25G/50G 10-100GbE compact edge Deep external DRAM (single-chip, VOQ)
Qumran2c OcNOS-SP BCM88820 2.4 Tbps 32x 50G PAM4 + 96x 25G NRZ 10-400GbE On-chip SRAM + off-chip HBM (VOQ) (16 MB + 4 GB HBM)
Qumran2c+ OcNOS-SP BCM88840 7.2 Tbps Not disclosed 6x400G, 40x100G class Deep (VOQ, single-chip)
Tomahawk 6 Planned (OcNOS-DC) BCM78910 (1024x100G) / BCM78914 (512x200G) 102.4 Tbps 3 nm 1024x 100G PAM4 (Peregrine) or 512x 200G PAM4 (Condor); CPO-capable 64x1.6TbE / 128x800GbE On-chip packet buffer
Qumran3U Planned (OcNOS-SP) BCM88400 (vendor-identified) 1.2 Tbps Not disclosed by Broadcom Not disclosed by Broadcom Not disclosed by Broadcom
Qumran3D Planned (OcNOS-SP) BCM88870 25.6 Tbps 256x 100G PAM4 32x800GbE / 64x400GbE class Built-in HBM (single-chip, VOQ)

Coming to OcNOS

Silicon we are bringing to OcNOS next. Qualified platforms follow; talk to us if you are planning around these.

Tomahawk 6 Coming soon
BCM78910 (1024x100G) / BCM78914 (512x200G)

World's first single-chip 102.4T switch. Dual-role scale-out (Clos) and scale-up (SUE); Cognitive Routing 2.0. Targets 1M+ XPU clusters.

102.4 Tbps
Full Tomahawk 6 deep-dive →
Qumran3U Coming soon
BCM88400 (vendor-identified)

The 1.2 Tbps access and aggregation tier of the Qumran3 family. UfiSpace and Edgecore identify the silicon in their 1RU aggregation routers as Qumran3U (BCM88400).

1.2 Tbps
Full Qumran3U deep-dive →
Qumran3D Coming soon
BCM88870

Single-chip 25.6T deep-buffer DCI / core / edge / peering router; 800G ZR coherent.

25.6 Tbps
Full Qumran3D deep-dive →

Scale-up, scale-out, scale-across

AI clusters stress the network in three planes. Broadcom answers each with different silicon.

Across the data center

Tomahawk 5 on OcNOS-DC

Leaf and spine in one building. RoCEv2, hardware dynamic load balancing and shallow-buffer burst absorption for GPU back-end traffic.

Scale-out

Inside the rack

Tomahawk Ultra (reference)

The XPU-to-XPU low-latency plane: 250ns switching, in-network collectives, link-layer retry. Noted for context; no OcNOS platform today.

Scale-up

Between data centers

Jericho-class deep-buffer DDC

A scheduled deep-buffer fabric for lossless routing across sites, with internet-scale FIB and perfect load balancing.

Scale-across

Questions engineers actually ask

StrataXGS (Tomahawk, Trident) uses a shallow on-chip shared buffer with a high-radix, low-latency pipeline: it is the workhorse for data-center and AI scale-out leaf/spine fabrics. StrataDNX (Jericho, Qumran, Ramon) uses a deep buffer with virtual-output-queue (VOQ) scheduling and, in Jericho, a credit-based cell fabric: it is built for carrier core/edge routing, deep-buffer DCI, and lossless AI scale-out via a Distributed Disaggregated Chassis (DDC). Rule of thumb: XGS = shallow buffer, speed and radix; DNX = deep buffer, scale and scheduling.
No. Every Tomahawk and Trident part is a shallow on-chip shared-buffer switch (tens to low-hundreds of MB of SRAM), with no HBM. Deep buffering is a StrataDNX property (Jericho and Qumran). If a design needs to absorb sustained congestion or long cross-DC flows, that is a deep-buffer DNX job, not a Tomahawk job.
Three roles. Scale-out (leaf/spine inside a data center): Tomahawk 5 and Tomahawk 6 on the XGS side, or Jericho3-AI plus Ramon3 (a deep-buffer scheduled DDC) on the DNX side. Scale-up (XPU-to-XPU inside a rack): Tomahawk Ultra and Tomahawk 6 with Broadcom Scale-Up Ethernet. Scale-across (between data centers): Jericho4 with Ramon4. OcNOS-DC runs the Ethernet scale-out fabrics on qualified open hardware.
The 12.8 Tbps, 32x400G Trident 4 is BCM56880 (Trident4-X11). BCM56780 is a separate, lower-capacity 8.0 Tbps Trident 4 variant with integrated MACsec/IPsec and cannot do 32x400G. Keep the two distinct.
Tomahawk 6 (BCM78910 / BCM78914) is the world's first single-chip 102.4 Tbps switch, built on a 3nm process, for both scale-out and scale-up AI fabrics; a co-packaged-optics variant (Tomahawk 6-Davisson, BCM78919) is sampling with integrated silicon photonics. Tomahawk Ultra (BCM78920) is a 51.2 Tbps scale-up chip with 250ns latency, in-network collectives, and link-layer retry for lossless XPU-to-XPU Ethernet, positioned against InfiniBand and NVLink.
OcNOS is qualified across Broadcom StrataXGS (Trident 3 and 4, Tomahawk 2 through 5) and StrataDNX (Qumran-AX, Qumran-MX and Qumran-UX, Qumran2a, 2c, 2c+ and 2u, Jericho2c+) on open hardware from Edgecore, UfiSpace, and others. The authoritative list is the OcNOS Hardware Compatibility List. OcNOS-DC targets the DC/AI Ethernet fabrics; OcNOS-SP targets the deep-buffer routing platforms.
A DDC replaces a traditional chassis router with separate 1RU/2RU boxes: Jericho line cards and Ramon fabric elements, wired as a scheduled cell fabric. Because the fabric is credit-scheduled and cell-based (not ECMP-hashed), it delivers perfect load balancing and no hash-collision hot spots, which is why it is attractive for both carrier scale and lossless AI back-end fabrics.
Tomahawk 6 (BCM78910 / BCM78914) roughly doubles Tomahawk 5: 102.4 Tbps versus 51.2 Tbps, on 3nm versus 5nm, with 1.6T-capable ports and configurable 100G or 200G SerDes. Tomahawk 5 is the current 800G AI scale-out workhorse; Tomahawk 6 is the first single-chip 102.4T switch and handles both scale-out (Clos) and scale-up fabrics. Both hold the packet buffer on chip. Broadcom publishes no buffer size for either, and in-package HBM is a StrataDNX property rather than a Tomahawk one.
Both are StrataXGS shallow shared-buffer switches. Tomahawk is the high-radix, latency-optimized line for hyperscale spine and AI scale-out (Tomahawk 5, Tomahawk 6). Trident is the more deeply programmable, feature-rich line for the data-center leaf and edge (Trident 3, Trident 4). Choose Tomahawk for raw speed and radix, Trident for programmability and richer features at a lower cost per port.
Both are StrataDNX deep-buffer routers with VOQ scheduling. Jericho is the higher-capacity switch-router used as a line card in a Distributed Disaggregated Chassis (DDC) with Ramon fabric elements (Jericho2c+, Jericho3-AI). Qumran is single-chip and needs no separate fabric, covering compact edge, aggregation, DCI, and access routing (Qumran2a, Qumran2c, Qumran3D, and the Qumran-UX, AX and MX access family). Use Jericho for scale-out chassis-class routing, Qumran for self-contained edge and access boxes.
Merchant silicon is off-the-shelf switch and router ASICs, such as Broadcom StrataXGS and StrataDNX, that any hardware maker can build a system on, as opposed to custom in-house chips. It is what makes open networking possible: the same Broadcom chip appears in boxes from Edgecore, UfiSpace, Celestica and others, and IP Infusion runs OcNOS across them, so you choose the hardware without being locked to a single vendor stack.

Pick the silicon. We ship the complete system.

A merchant ASIC is half a router. OcNOS turns Broadcom silicon into a complete, supported switch or router on open hardware you choose, from Edgecore, UfiSpace and Celestica. The Hardware Compatibility List is the source of truth: if a box is not on the HCL, it is not a qualified OcNOS target.

Hardware Compatibility List Feature Matrix Contact us

Silicon to solution

AI fabric
Tomahawk 5 spine/leaf, OcNOS-DC
DC leaf
Trident 3 / 4, OcNOS-DC
Carrier core
Jericho2c+, OcNOS-SP
Edge / DCI
Qumran2a / 2c, OcNOS-SP
Cell site
Qumran2u, OcNOS-SP

Go deeper. Take it with you.

Two short, technical downloads that go further than this page: the full OcNOS-DC datasheet and the lossless 800G AI fabric architecture.

Datasheet

OcNOS-DC Datasheet

Full OcNOS-DC specification: the EVPN-VXLAN and Ethernet for AI feature set, software SKUs, supported hardware platforms, and the solution ordering guide.

Get the datasheet
Solution Brief

OcNOS 800G Lossless AI Fabric

Lossless RoCEv2 fabric on Tomahawk 5 spines: SKU tiers, qualified platforms, and deployment architecture.

Get the brief

Design the whole AI fabric with OcNOS

From the business case to the port-count maths, pick up wherever you are in the build.

Solution Open AI Fabric The complete 800G AI fabric: open switches, OcNOS-DC, and support under one contract. Reference designs AI Fabric Topologies Rail-optimized two-tier and three-tier fabrics, and coherent DCI, sized in real port counts. Size & build AI Fabric Design Suite Size a GPU fabric: leaf, spine, and super-spine counts with a component and power summary.
New to AI fabrics? Start here What is an AI fabric? What is a GPU fabric? What is lossless Ethernet? What is RDMA? AI networking switches
The technology inside RoCEv2 lossless Rail-optimized network DLB GLB (7.1) Ultra Ethernet Tomahawk 6 at 1.6T DCQCN PFC deadlock InfiniBand vs Ethernet RoCE vs InfiniBand AI fabric architecture Coherent DCI