At Computex 2024, Foxconn subsidiary Ingrasys displayed a bare Nvidia HGX B200 motherboard without heatsinks, revealing a major hardware update.
The NVLink Switch chip count has been reduced from four (H100 generation) to two (B200 generation), and their placement on the board has shifted.
First-Generation NVSwitch #
The first NVSwitch was introduced with the Nvidia DGX-2 supercomputer.
- 20 billion transistors
- 18 ports (50 GB/s each)
- Total bandwidth: 900 GB/s
This switch enabled up to 9 devices to communicate with any other 9 devices.
It also included control I/O interfaces:
- PCIe Gen 2 x4 management port
- I²C and GPIO
In DGX-2, each baseboard used 6 NVSwitches to create a fully connected GPU network.
The DGX-2 system had two baseboards to connect all 16 V100 GPUs. Each NVSwitch had 2 unused ports, which could be used with POWER9 CPUs that natively supported NVLink 2.0.
Below: Inspur NF5488M5 HGX-2, clearly showing 6 NVSwitch heatsinks.
Second-Generation NVSwitch (A100 Era) #
With the Nvidia A100, NVSwitch heatsinks became larger. The HGX A100 platform was entirely designed and pre-assembled by Nvidia before shipping to OEMs.
Below: Inspur NF5488A5 HGX A100 with 6 NVSwitch heatsinks visible.
Third-Generation NVSwitch (H100 Era) #
In the H100 generation, 4 NVSwitches were placed on one side of the baseboard.
Below: ASUS ESC N8A E12 HGX H100 system.
The HGX H200 design looks almost identical to HGX H100.
Fourth-Generation NVLink Switch (B100/B200 Era) #
HGX B100 with Heatsinks Installed #
Launch Event Photo of HGX B100 #
Notice: near the edge connector, PCIe Retimers appear instead of NVSwitch chips.
With the HGX B200 bare motherboard, the NVLink Switch positions are clear. Nvidia confirmed that the B100 and B200 share the same NVSwitch layout.
Here, PCIe Retimers are exposed. These typically use smaller heatsinks since their TDP is only ~10–15W, depending on vendor (Astera Labs, Broadcom, Marvell, etc.).
Key update: Only two NVLink Switches remain, now placed in the center of the motherboard, instead of four near the edges as before.
Conclusion: Why Nvidia Changed NVSwitch Design #
The latest NVLink Switch chips are larger and moving them to the center of the board helps:
- Reduce trace length → improves high-speed signaling
- Optimize GPU-to-switch communication paths
On the HGX B200, GPUs are grouped in sets of four on each side, minimizing signal distance to the central NVLink Switches.
Key Takeaways #
- Nvidia reduced NVSwitch count from 4 to 2 on HGX B200.
- Switches are now centrally located for shorter trace length.
- PCIe Retimers are used along the edges instead of switches.
- This redesign improves signal integrity, efficiency, and scalability for the next-gen B200 GPUs.