Broadcom Jericho 3-AI (AI Ethernet Fabric Switch・2024年)（ブロードコムジェリコ）
2024年Q1 Broadcom (米San Jose・1991年Henry Samueli + Henry Nicholas創業・World Ethernet switch chip Number 1 share 80%+) 発表のJericho 3-AI・AI workload optimized Ethernet Fabric switch chip + 28.8 Tbps capacity + 800G port + Ethernet RoCEv2 + Adaptive Routing + Cell-based packet switching + 5nm TSMC process + AI Cluster scale-out向け fabric switching・Arista 7800R3 / NVIDIA Spectrum-X800 / Cisco Nexus 9800 / Juniper PTX10008 OEM主要switches採用・AI training cluster Ethernet fabric defacto standard。
概要
Broadcom Jericho 3-AI は、2024年Q1 Broadcom (米San Jose・1991年Henry Samueli + Henry Nicholas創業・World Ethernet switch chip Number 1 share 80%+・累計売上$50B+/year) が発表したAI workload optimized Ethernet Fabric switch chip。Broadcom Jericho series (Service Provider routers向け Higher capacity switches series・Tomahawk = ToR switches向け / Jericho = Service Provider routers向け の Broadcom switch chip product line)・Jericho 1/2系 = Service Provider routing向け Traditional design → Jericho 3-AI 2024年Q1 = AI workload optimization Pivot後継 (AI Ethernet Fabric for AI training cluster Specific design)。28.8 Tbps switch capacity + 800G port × 36 ports + Standard Ethernet + RoCEv2 (RDMA over Converged Ethernet v2) + Adaptive Routing (Congestion-aware path selection) + Cell-based packet switching (Higher latency optimization vs Standard packet switching) + Programmable pipeline + 5nm TSMC process + HBM3 onboard memory (Larger packet buffer for AI workload bursty traffic absorption)・AI Cluster scale-out向け fabric switching (NVIDIA NVLink Switch Scale-Up + Ethernet Fabric Scale-Out hierarchy)・Arista 7800R3 / NVIDIA Spectrum-X800 (Spectrum-4 / Jericho 3-AI dual products in NVIDIA Networking) / Cisco Nexus 9800 / Juniper PTX10008 等のOEM主要switches採用・AI training cluster Ethernet fabric defacto standard chip。
主な特徴・仕組み
- Architecture: AI workload optimized Ethernet Fabric switch chip
- Switch Capacity: 28.8 Tbps (Tomahawk 5 51.2 Tbps to同レベル range・Jericho service provider higher capacity tier)
- Port Configuration: 36 ports × 800G (HBM3 buffer per port)
- Process: 5nm TSMC
- RoCEv2: RDMA over Converged Ethernet v2 (Industry-leading AI fabric support)
- Adaptive Routing: Congestion-aware dynamic path selection
- Cell-based Switching: Higher latency optimization vs Standard packet switching (AI workload micro-bursty traffic better absorption)
- HBM3 Buffer: Large packet buffer onboard (AI workload bursty traffic absorption)
- OEM Adopters: Arista 7800R3 / Cisco Nexus 9800 / Juniper PTX10008 / Dell PowerSwitch / NVIDIA Spectrum-X800 (mixed deployment)
- Use Cases: AI training cluster Ethernet fabric (NVIDIA NVLink Scale-Up + Ethernet Scale-Out)
- Customer Adopters: Meta + xAI Colossus + Microsoft Azure等のHyperscaler AI training clusters
スペック比較表
| AI Ethernet Fabric Switch Chip (2024) | Capacity | Port | RoCEv2 | Cell-based | Vendor |
|---|---|---|---|---|---|
| Broadcom Jericho 3-AI | 28.8 Tbps | 36 × 800G | ◎ | ◎ | Broadcom |
| Broadcom Tomahawk 5 | 51.2 Tbps | 64 × 800G | ○ | × (Packet-based) | Broadcom |
| NVIDIA Spectrum-X800 (Spectrum-4) | 51.2 Tbps | 64 × 800G | ◎ | △ | NVIDIA |
| Cisco Silicon One G200 | 51.2 Tbps | 64 × 800G | ◎ | × | Cisco |
| Arista 7800R3 (Jericho 3-AI base) | OEM Switch level | 36 × 800G | ◎ | ◎ | Arista (Broadcom chip) |
具体例・対応製品
- Broadcom Jericho 3-AI Chip (BCM88890): Standalone chip pricing $5,000-$10,000 estimated
- Arista 7800R3 (Jericho 3-AI based): $50,000-$200,000・Service provider switch level
- NVIDIA Spectrum-X800 (Jericho 3-AI variant): NVIDIA Networking solutions multi-chip strategy
- Cisco Nexus 9800 (Jericho 3-AI option): Multi-vendor switch chip choice
- 競合 Broadcom Tomahawk 5: 51.2 Tbps + Top-of-Rack ToR・Jericho 3-AI Service provider higher capacity tier
自作PCでの選び方・注意点
Broadcom Jericho 3-AI は「AI training cluster Ethernet fabric (Meta 350,000 H100 + Microsoft Azure + xAI Colossus 100,000 H100等のHyperscaler AI cluster scale-out)」「Service Provider routing networks (Backbone routing higher capacity)」「AI workload optimized Ethernet fabric (Standard Ethernet biased vs AI specific optimization)」用途のEnterprise tier Ethernet switch chip。Consumer / 自作PC範囲外・Service provider routers / AI training cluster fabric deployment専用・Broadcom switch chip Number 1 share (80%+) + Jericho 3-AI = Broadcom Service Provider router heritage + AI workload optimization再Position・AI Cluster scale-out (NVIDIA NVLink Scale-Up + Ethernet Scale-Out hierarchy) で Ethernet portion主要 chip。Tomahawk 5 (Top-of-Rack ToR 51.2 Tbps switches) と Jericho 3-AI (Service Provider / Inter-rack higher capacity 28.8 Tbps + AI fabric optimization) で Broadcom switch chip portfolio position differentiation・Hyperscaler customers (Meta + Microsoft + xAI) AI training cluster Ethernet fabric Broadcom Jericho 3-AI dependent deployment実績。
関連用語との違い
- vs Tomahawk 5 (兄弟Broadcom chip): JerichoはAI fabric + Cell-based + Service provider routing・TomahawkはTop-of-Rack + Packet-based + Higher capacity 51.2 Tbps
- vs NVIDIA Spectrum-X800: BroadcomはMulti-vendor switch chip (Arista/Cisco/Juniper OEM)・NVIDIAはNVIDIA branded switches (Spectrum-X marketing)
- vs Cisco Silicon One G200: Broadcomは80% market share・CiscoはCisco独自 5-10% market share
よくある質問（FAQ）
Q1: Tomahawk 5 vs Jericho 3-AI どちら？ A: Tomahawk 5 = Top-of-Rack ToR (Inside rack switching・Higher capacity 51.2 Tbps・Higher port count・Packet-based) ・Jericho 3-AI = Service provider routing / Inter-rack (Lower capacity 28.8 Tbps + AI fabric optimization + Cell-based)。Use case依存。
Q2: Cell-based switching vs Packet-based switching 違い？ A: Cell-based = 細小cells (64-256 byte cells) packet-divide + Switch + Reassembly at egress・Lower latency optimization (Packet-by-packet vs Cell-by-cell scheduling)・AI workload micro-bursty traffic absorption better・Service provider router heritage (Multi-tenant traffic management)。Tomahawk 5 = Packet-based for ToR・Jericho 3-AI = Cell-based for AI fabric。
Q3: AI Ethernet Fabric の意義は？ A: AI training cluster (NVIDIA H100 1,000+ GPU cluster) で NVLink Scale-Up (Inside rack) + Ethernet Scale-Out (Inter-rack) hierarchy で Ethernet fabric portion = AI workload bursty Heavy traffic (Gradient sync / Model parallel communication) → AI specific optimization (RoCEv2 + Adaptive Routing + Cell-based + HBM3 buffer) で Standard Ethernet vs AI workload performance大幅向上。
まとめ
Broadcom Jericho 3-AI = 2024年Q1 Broadcom発表のAI Ethernet Fabric switch chip。28.8 Tbps + 800G × 36 ports + RoCEv2 + Adaptive Routing + Cell-based switching + 5nm + HBM3 buffer・AI training cluster Ethernet fabric defacto standard + Arista/Cisco/Juniper/NVIDIA OEM主要switches採用 + Hyperscaler customer dominant deployment chip。