At Computex 2024, Foxconn subsidiary Ingrasys showcased a bare NVIDIA HGX B200 motherboard, revealing one of the biggest structural updates in the NVLink architecture since HGX H100.
The number of NVLink Switch chips has been reduced from four (H100 era) to two (B200 era), and both switches have been repositioned to the center of the board.
🧱 First-Generation NVSwitch (DGX-2) #
The original NVSwitch debuted with the NVIDIA DGX-2 platform.
Key specs:
- 20B transistors
- 18 × 50 GB/s ports
- 900 GB/s total bandwidth
- Control interfaces: PCIe Gen2 x4, I²C, GPIO
A DGX-2 baseboard used six NVSwitches to fully interconnect eight GPUs per board.
Two baseboards interconnected 16 V100 GPUs, with spare ports available for POWER9 CPUs supporting NVLink 2.0.
Below: Inspur NF5488M5 HGX-2 with six NVSwitch heatsinks clearly arranged.
⚙️ Second-Generation NVSwitch (A100 Era) #
With A100, NVSwitch heatsinks grew in size and complexity. NVIDIA delivered fully pre-assembled HGX A100 modules to OEMs.
Below: Inspur NF5488A5 HGX A100, again with six NVSwitches.
🚦 Third-Generation NVSwitch (H100/H200) #
The H100 HGX platform moved to four NVSwitches, placed on one side of the motherboard.
Below: ASUS ESC N8A-E12 HGX H100.
The HGX H200 layout remained mostly identical.
🔄 Fourth-Generation NVSwitch (B100/B200) #
HGX B100 With Heatsinks #
Launch Event Photo #
Notable change: PCIe retimers appear along the board edge, replacing prior NVSwitch positions.
The HGX B200 bare motherboard confirms a major redesign — only two NVLink Switches remain, now positioned dead center. NVIDIA states B100 and B200 share this layout.
Retimers are also visible, each requiring only ~10–15W of cooling.
The biggest architectural shift:
✔ NVSwitch count reduced from 4 → 2 #
✔ Both switches relocated to the center of the board #
📌 Why NVIDIA Moved to Two Central NVSwitches #
The new fourth-generation NVSwitch is larger, faster, and optimized for shorter signal paths. Moving the switches inward:
- Shortens NVLink trace lengths for all GPUs
- Improves signal integrity at extremely high speeds
- Reduces power loss and board complexity
- Aligns with GPU groupings (4 GPUs per side on HGX B200)
📝 Key Takeaways #
- HGX B200 cuts NVSwitch chips from 4 down to 2
- Switches moved to the center for optimal routing
- PCIe retimers now occupy edge positions
- Design improves bandwidth efficiency, scaling, and signal quality for next-gen B200 GPUs