ALAB: How Large is the GB200 Content Opportunity?
Key Points
- Deep supply chain analysis of PCIe retimer and switch content in NVIDIA servers
- Examines differences between DGX (current) and GB200 (next-gen) configurations
- DGX servers equipped with 8 PCIe Gen5 retimers on UBB + 8 on head node
- Standard DGX includes two 144-lane PCIe Gen5 switches for CPU/GPU/network connectivity
- Some MGX customers optimize board layout to reduce retimer count (potential headwind)
DGX Configuration Details
- Retimers: 8 on Universal Base Board (GPU side) + 8 on CPU head node = 16 total
- PCIe Switches: Two 144-lane switches connecting CPU (32 lanes), CX7 NICs (32 lanes), GPUs (64 lanes)
- Flexible lanes: 16 remaining lanes for customer/manufacturer customization
- Design variability: Customer customization vs. NVIDIA reference design critical to content modeling
Bull vs. Bear Debate
- Bears: Concerned about reduced retimer usage in GB200 vs. DGX due to different form factor/optimization
- Bulls: Optimistic that PCIe switch product launch significantly increases company content dollar
- Supply chain complexity makes consensus difficult; variant perception on system architecture design
Content Opportunity Modeling
- Detailed breakdown required to assess true content opportunity in newer platforms
- Mix of standard deployments vs. customized configurations affects aggregate content
- Product diversification (switches + retimers) provides upside if adoption broadens
Source
- File:
(4) Astera Labs – How Large is the GB200 Content Opportunity_.pdf - Location: Dropbox/2. Semi/Networking/ALAB/
Related
- ALAB | _MOC-networking
Pages that link here (1)