PDF 原檔：報告_OCP2026_Meta_Tahan乙太網路ScaleUp交換器_20261015_original.pdf
講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，
### p<n>為頁碼；圖表內容見原檔。
圖片清單（已驗證 2026-10-19）
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_Meta_Tahan乙太網路ScaleUp交換器_20261015_p7.png |
p7 | 真資料圖 | Tahan 系統方塊圖：Netlake 微伺服器（Intel Xeon、FBOSS）、RunBMC（Aspeed AST2620）、TH5（BCM78900）；32 個 800GE 埠經 4 組 backplane 連接器通往 16 顆加速器（每顆 2 埠），另 32 個 OSFP 保留、1 個 QSFP 接 ToR；頁面文字：1OU 混合液冷／氣冷、基於 51.2Tbps Tomahawk5 |
報告_OCP2026_Meta_Tahan乙太網路ScaleUp交換器_20261015_p10.png |
p10 | 真資料圖 | 每加速器頻寬圖：16 顆加速器各以 2×800G 連 Tahan 1~4 共 4 台交換器，6.4Tbps／加速器／單向，non-blocking，註記交換晶片為 Broadcom Tomahawk5 51.2Tbps |
報告_OCP2026_Meta_Tahan乙太網路ScaleUp交換器_20261015_p11.png |
p11 | 真資料圖 | 延遲圖：加速器—1.5m twinax（約 7ns）—Tahan 交換器（store-and-forward <1µs）—1.5m twinax—加速器；1 hop；cut-through 可再降 25%（探索中） |
報告_OCP2026_Meta_Tahan乙太網路ScaleUp交換器_20261015_p14.png |
p14 | 真資料圖 | 次世代 scale-up fabric：72 顆加速器各 24×400G、8 台 Tomahawk6 102.4Tbps 交換器（每台 216×400G），9.6Tbps／加速器／單向，單跳 non-blocking，為 Tahan 的 1.5 倍 |
原始內容
p1
Vimal Vasudevan Technical Program Manager, Meta Ashutosh Grewal Software Engineer, Meta Tahan: An Open- Ethernet-Based Scale- Up Switch for Meta's MTIA Accelerators AI CLUSTERS
p2
Agenda 01 Industry Problem 02 Inflection Point 03 Solution - Introducing Tahan 04 Bandwidth and Latency 05 Scale Up Features 06 Next Gen Scale Up Fabric 07 Open Standards for Scale Up 08 Call to action
p3
• Proprietary and vendor-locked • Closed, Bundled ecosystems from a single vendor • Limited flexibility • Constrained roadmaps • Little leverage over cost or supply. Industry Problem
p4
• MTIA - Proprietary scale-up model no longer fit • Giving up Control and Efficiency • Vendor-locked scale-up solutions were not a viable path for an open, purpose-built platform. Inflection Point - Meta Custom Silicon
p5
• Open, Ethernet-based scale-up switch, Tahan designed specifically for MTIA300 Training Accelerators. • Vendor independence • A mature, open ecosystem • Full-stack control • Scale and Cost Efficiency Solution - Introducing Tahan
p6
Tahan - System Overview
p7
Tahan - System Overview • 1OU, Hybrid liquid/air cooled • Based on 51.2 Tbps Broadcom Tomahawk5 ASIC
p8
Tahan - System Overview • PHY-less design with flyover cables • Features 32x 800GE towards backplanes, 32x OSFP ETH(800GE) on front panel
p9
Tahan - System Overview • Deployed in ORv3-HPRv1 racks
p10
Bandwidth per Accelerator Tahan 1 Tahan 2 Tahan 3 Tahan 4 2 × 800G 2 × 800G 2 × 800G 2 × 800G 32 × 800G per switch 16 accelerators × 2 links * Broadcom Tomahawk5, 51.2 Tbps merchant Ethernet switch 6.4 Tbps per accelerator, per direction 16 accelerators, 8 × 800GbE each · 4 switches Non-blocking All of it on merchant Ethernet silicon Accelerator each of 16
p11
1 hop always every accelerator pair - no second tier < 1 µs per switch hop today store-and-forward −25% latency, cut-through egress on the header, not the full frame - under exploration Latency: One Hop, Always Accelerator Tahan switch Accelerator 1.5 m twinax ~7 ns < 1 µs store-and-forward 1.5 m twinax ~7 ns One switch traversal, always. Across the fabric path the cables are ~1.5% - the switch is the budget. Fabric path only - excludes NIC, PCIe and host software
p12
●FBOSS already underpins Meta's scale-out fabrics - the same NOS evolves to run the scale-up fabric ●Same repair and upgrade philosophy ●Same monitoring and telemetry One NOS, Two Fabrics Scale-out fabric (today) Scale-up fabric (Tahan) FBOSS Agent · QSFP Service · FSDB on the switch's own x86 CPU
p13
A Subset of the Scale-Out Feature Set Reused Turned off Forwarding IPv6 dataplane, interface routes Lossless PFC lossless classes, AI-tuned QoS Multipathing ECMP, Dynamic load balancing Routing Protocols BGP routes
p14
Next-Gen Scale-Up Fabric Switch 1 Switch 2 Switch 3 Switch 4 Switch 5 Switch 6 Switch 7 Switch 8 3 × 400G to each switch 9.6 Tbps per accelerator, per direction 72 accelerators, 24 × 400G each · 8 switches Non-blocking · Single hop 1.5× Tahan on merchant Ethernet silicon * Broadcom Tomahawk6, 102.4 Tbps merchant Ethernet switch Accelerator each of 72 216 × 400G per switch
p15
Open Standards for Scale-Up ESUN: OCP base spec Rev 1.0 approved · LLR and CBFC: UEC specs, required by ESUN ESUN Ethernet, minus the header tax 48 B of IPv6 + UDP → ~4 B; more goodput per frame LLR Link Layer Retry Replay at the link, one link RTT; transport never sees the loss CBFC Credit-Based Flow Control Send only against receiver credits; lossless, less buffer than PFC
p16
Call to Action Take it The complete Tahan hardware spec Will be contributed to OCP Join it ESUN at OCP; LLR and CBFC at UEC Rev 1.0 approved, driven with Microsoft Build it Scale-up on merchant Ethernet One NOS, one operational model
p17
Thank You!