NVIDIA HGX B300 NVL16 shifts from package count to die count
NVIDIA's HGX B300 NVL16 carries up to 2.3TB of HBM3e across sixteen dual Blackwell GPU package modules, while eight ConnectX-8 NICs sit on the UBB.

| NVIDIA HGX B300 NVL16 | |
|---|---|
| memory | up to 2.3TB HBM3e |
| GPU package modules | sixteen dual Blackwell GPU package modules |
| heatsink towers | eight large air cooling heatsink towers |
| ConnectX-8 NICs | eight |
| UBB connectors | OCP UBB style connectors with orange covers |
NVIDIA's HGX B300 NVL16 carries up to 2.3TB of HBM3e memory. Wire headline calls the platform "Massively Different", but the document we trust is the ServeTheHome article, and the number that decides it is 2.3TB. That figure is not a marketing adjective; it is still the memory capacity that changes the node's data budget.

It also shows sixteen dual Blackwell GPU package modules, eight large air cooling heatsink towers, and eight ConnectX-8 NICs on the UBB. The rename moves the spec sheet from package count to die count, so buyers should read NVL16 as the NVLink-connected die count, not as a simple HGX B200 swap. Built-in PCIe switch in the ConnectX-8 still replaces functions that previously needed separate chips, which constrains the PCIe retimer, PCIe switch, and AI NIC vendors.

Desk judgment is that NVIDIA is not just adding memory; it is moving east-west GPU networking into the platform. Buyers face a constraint of choice in practice: the HGX B300 NVL16 makes NVIDIA the default for in-node GPU traffic, while north-south networks remain more open for now. Fab or component makers see a limit: the document does not prove that the ConnectX-8 PCIe switch will replace every external part, because the page today still leaves out per-die HBM capacity, package power, and the NVLink switch count for the B300 NVL16.
NVIDIA is moving east-west GPU networking into the HGX B300 NVL16, which narrows buyer choice for in-node traffic. The document does not prove that the ConnectX-8 PCIe switch replaces every external part, because per-die HBM capacity, package power, and the B300 NVL16 NVLink switch count are missing.

The document leaves out per-die HBM capacity, package power, and the NVLink switch count for the HGX B300 NVL16.
The platform integrates eight ConnectX-8 NICs and uses their built-in PCIe switch to absorb functions previously handled by separate retimer or switch parts. That shifts east-west GPU traffic to NVIDIA and leaves north-south traffic more open, but the missing per-die HBM, package power, and NVLink switch count limit a full BOM impact assessment.
After servethehome.com. We did not report this. The pictures, if any, are theirs.


