OCP Global Summit 2026 | Meta | Rongchun Zhou, Reza Niazmand | Inside Meta's 102.4T IceCube Switch
1. From Minipack3 to IceCube: What Changed at 102.4T
This talk by Meta network hardware engineer Rongchun Zhou and network TPM Reza Niazmand is about the switch itself inside the IcePack rack. We have covered IcePack's network architecture separately; this piece is hardware only.
IceCube's baseline specs:
ASIC: Broadcom Tomahawk 6C (102.4T); Meta stresses the architecture can take other ASICs too.
Ports: 64 OSFP, configurable as 128x 800GE or 512x 200GE, non-blocking.
Optics: Meta-defined 2x800G-DR4/FR4 OSFP modules.
Form factor: 2OU, hybrid liquid/air cooled, deployed in ORv3-HPRv1 racks, powered from a 48V bus bar at the rear.
Control plane: reuses the Minipack3 reference design, with an Intel Icelake-D CPU (NetLake card) and an ASPEED AST2620 BMC.
Versus Minipack3, only three things are truly new: an ASIC with twice the switching capacity, high-speed signals moved onto flyover cables, and leak-detection circuits added for liquid cooling. Every other control module is a design already proven in Meta's data centers, which is the key to NPI (new product introduction) speed.
Why OSFP instead of QSFP-DD? The Q&A answer was direct: a single-row connector has a signal integrity (SI) advantage, and the larger OSFP body is easier to cool. Meta moved to OSFP back in the 100G-per-lane Minipack3 generation.
2. Inside the Box: 340 Meters of Flyover Cable
512 lanes of 212G PAM4, one pair each for transmit and receive, add up to 1,024 differential pairs. Routing all of them from the ASIC to the front panel on PCB would eat most of the channel loss budget, so IceCube goes PHY-less with flyover cables: 16 UPXR connectors sit around the ASIC, and copper cables "fly over" the main board straight to two stacked OSFP carrier cards.

▲ IceCube's 212G PAM4 flyover copper cables are routed in four layers, with 340 m of total differential-pair length, close to the 346 m perimeter of a football pitch (Image source: Meta | 2026 OCP Global Summit presentation (page 5))
The cables are routed in four layers (blue, red, green, yellow from bottom to top), 340 m in total, or about 33 cm per differential pair on average. Meta said its mechanical and SI teams spent a great deal of time tuning the "cable recipe"; the slides show seven iterations of the cabling during NPI before it became production-ready.
That work gave Meta end-to-end knowledge of flyover cable technology: electrical SI modeling, mechanical integration, manufacturing process and vendor management. Once a hyperscaler understands things down to the process level, what an ODM sells is no longer design but execution.

▲ IceCube's signal path from ASIC to front panel: 512 lanes, 1,024 differential pairs, 16 UPXRs and 64 OSFPs, with 340 m of copper cable inside the box (Image source: Simple Tech Trend original, data from Meta's 2026 OCP Global Summit presentation)
3. The Main Board: 44 Layers, 5.62 mm, One Board for Four Platforms
The switch main board (SMB) is the other centerpiece:
Size 370 × 265 mm, thickness 5.62 mm (±7%)
26 + 18 = 44 layers in an N+M stack-up
Material mix EM892K2 + MC24PV + EM370Z + MC24P, with Class A material for high-speed signal layers
ENEPIG surface finish, chosen to meet the compression-contact requirement of the flyover cable connectors

▲ The IceCube switch main board: 370×265 mm, 44 layers, 5.62 mm thick, ENEPIG surface finish, shared with other Meta platforms (Image source: Meta | 2026 OCP Global Summit presentation (page 7))
Note the "small and thin" part: because most high-speed signals ride on cables, the board does not have to max out area and layer count for long 212G traces. Flyover cable shifts part of the PCB challenge into a cable challenge, and that is what makes board reuse possible later.
4. Power and Thermal: 85% of 3,820 W Goes to Liquid
Item | IceCube (Tomahawk 6C) | Notes |
Max typical power | 3,820 W | Fully populated 5nm DSP DR4 modules |
Efficiency | ~37 pJ/bit | STT estimate: 3,820 W ÷ 102.4 Tb/s |
vs Minipack3 | ~7% better efficiency | Another ~7% with 3nm DSP modules |
Liquid-cooled share | ~85% | ASIC and OSFP cold plates, OCP BMQC connector |
Air-cooled share | ~15% | Micro-server, power stages, BMC, 4 fan trays |
Coolant inlet | 30°C / 40°C, ~6 LPM per blade | Supports AALC or FLC |
Acoustics | 63.3 dBA (30°C / 6000 ft) | Minipack3 about 71 dB |

▲ About 85% of IceCube's 3,820 W is removed by liquid cooling; efficiency is about 37 pJ/bit, roughly 7% better than Minipack3, with about another 7% possible from 3nm DSPs (Image source: Simple Tech Trend original, data from Meta's 2026 OCP Global Summit presentation)
Three observations:
The optics are on cold plates too. Each OSFP-RHS cage is sandwiched between top and bottom cold plates fed by a parallel flow network, which means air cooling is no longer the default for 1.6T-class DSP modules.
DSP process node is the next 7%. Meta itself pointed out that moving from 5nm to 3nm DSP saves roughly another 7%, so a meaningful part of system-level efficiency gains now depends on optical-module DSP process scaling.
These numbers are "fully retimed". In other words, Meta's baseline for this generation is full-DSP pluggable modules, not LPO, and not NPO or CPO.
5. One Design, Two ASICs, Two ODMs
Reza Niazmand was clear about the real motive behind flyover cables: design reuse.
IceTea: the same main board with half the flyover cables routed to the rear into a copper cable cartridge, eliminating a large number of fibers and optics. Meta itself calls it a "hack".
Santa Barbara, Anacapa: iterations in other product lines use the same SMB.
IceCube-N: reuses the IceCube OCP design with an NVIDIA Spectrum-6 ASIC; some people call it "Nice Cube".
Item | IceCube | IceCube-N |
Switch ASIC | Broadcom Tomahawk 6C | NVIDIA Spectrum-6 |
212G signal path | UPXR flyover cable | 100% PCB routing |
OSFP layout | 1x8 OSFP-RHS belly-to-belly ×8 | 2x8 stacked OSFP cages and cold plate, belly-to-belly |
ODM partner | Celestica | Accton |

▲ IceCube-N reuses IceCube's OCP design but switches to NVIDIA Spectrum-6, with 100% of 212G signals routed in PCB, co-developed with Accton (Image source: Meta | 2026 OCP Global Summit presentation (page 11))
The most telling row is the second one: in the same 102.4T generation, the Spectrum-6 version routes 100% on PCB, while the Tomahawk 6C version chose flyover cables. The talk did not explain why, but the outcome is clear: Meta now has production experience with both signal paths, plus a second source for both the ASIC and the ODM. Meta's rationale is that GPU generations change so fast that the network must iterate just as quickly.
We covered Celestica's weight at Meta in Earnings Highlights: Celestica (CLS) Q2 FY2026; for its thermal work in the 51.2T generation, see Tech Analysis: 51.2T NPO Switch Cooling. At 102.4T, Meta handed the Spectrum-6 version to Accton, and the ODM role shifts from "sole source" to "replaceable".
6. The Rack Is the Delivery Unit: 8,192 Fibers and the NPO/CPO Option
Meta's other shift is that its smallest deployment unit moves from the L10 blade to the L11 full rack. IcePack has three rack types, GTSW, STSW and ETSW, with the same elevation and different optics and fiber:
A GTSW rack alone holds 8,192 in-rack fibers, with an intricate fiber-shuffle design
Pre-cabled at the integrator so racks go live on arrival, shortening turn-up time
The second Call to Action is blunt: rack-level interconnect technology is at its limit on current 100T platforms, and the next generation needs innovation in L11 fiber management, cable backplanes and optical backplanes. Niazmand also said NPO and CPO are being evaluated for the 200T and 400T generations, mainly for power and cost. For why NPO is more than a transition step, see NPO Is Not a Stepping Stone to CPO.
Specs and the design package were submitted to OCP in September 2026.
7. Risks and Counterarguments
Flyover cable is not the only path: IceCube-N proves the same generation can route 100% on PCB. If PCB materials and SerDes keep improving, the cable approach may only win for certain ASIC packages.
37 pJ/bit is an STT estimate, not an official figure: the slide's bar chart is only relative, and 3,820 W is "max typical power", so real averages will be lower.
The extra 7% from 3nm DSP is a verbal projection: no production timeline was given, and no DSP vendor was named.
NPO/CPO is only "under evaluation": Meta made no NPO or CPO deployment commitment for the 102.4T generation.
The rack fiber bottleneck has no answer yet: Meta only raised the problem; the 8,192-fiber shuffle still relies on pre-cabling at the integrator.
8. What It Means for Taiwan's Supply Chain
Accton won the Spectrum-6 version, a fact stated on the slides. Because IceCube-N reuses the OCP design, Accton is entering a Meta platform rather than a one-off project; watch for when IceCube-N reaches volume production.
Taiwanese laminate part numbers show up on the main board. EM892K2 and EM370Z listed on the slide belong to Elite Material's (EMC) EM series; a 44-layer, 5.62 mm 102.4T board means high-end CCL content and pricing keep rising.
The BMC is ASPEED AST2620. The RunBMC block on the architecture diagram is labeled AST2620, so ASPEED's position in Meta's switch platforms carries into the 102.4T generation.
Optics and liquid cooling get bigger in the numbers. OSFP modules on cold plates, a 30/40°C coolant inlet and about 6 LPM per blade are new specs for pluggable thermal design; but Meta did not disclose optics or cold-plate suppliers, so this should not be over-read.
9. Conclusion
IceCube matters not because of the 102.4T figure but because Meta has reordered the challenges of a switch: the ASIC can be swapped and so can the ODM, while the hard parts are the 340 m of copper in the box, the 44 layers on the board and the 8,192 fibers in the rack. With one shared main board, Meta hedges both ASIC and ODM risk and takes control of interconnect technology back in-house.
Keep the timelines separate: the 102.4T generation ships DSP pluggables plus liquid cooling, and NPO/CPO will only be seriously evaluated from 200T onward. Over the next 12 months, watch three things: IceCube-N's production schedule, when 3nm DSP modules arrive, and what the OCP community proposes for L11 fiber management and optical backplanes.
When both the ASIC and the ODM are replaceable, the only differentiation left in a switch is interconnect; this talk is Meta publicly declaring that the bottleneck has moved into copper and fiber.
This article is for technology and industry trend analysis only and does not constitute investment advice.




Comments