NVIDIA HGX B300 servers: eight systems and how an order runs.

What HGX B300 is

NVIDIA lists HGX B300 as eight Blackwell Ultra GPUs in the SXM form on one baseboard, connected by fifth-generation NVLink through an NVLink Switch. The board before it, HGX B200, carries eight Blackwell GPUs on the same generation of NVLink.

NVIDIA makes the board; the server around it comes from a system maker. Each one chooses the host processors, the memory, the drives, the cooling and how the cluster network leaves the chassis, so the choice between B300 systems is a choice between those, not between GPUs.

The B300 systems in our catalogue

Each of these is built around an HGX B300 board:

• Cisco UCS C880A M8 — Cisco’s air-cooled rack server with two Intel Xeon 6 processors. Cisco’s data sheet lists it as a set of complete configurations, each under its own part number.
• Supermicro SYS-822GS-NB3RT — two Intel Xeon 6700 processors with P-cores, air-cooled, with NVIDIA ConnectX-8 SuperNICs built in. Supermicro marks it as needing onsite service.
• Supermicro AS-8126GS-NB3RT — Supermicro’s AMD counterpart, with two EPYC 9005 processors, also air-cooled and also with ConnectX-8 SuperNICs.
• ASRock Rack 8U16X-GNR2 B300 — two Intel Xeon 6 processors (6700P, 6500P or 6700E series), with the cluster ports on ConnectX-8 SuperNICs.
• ASRock Rack 8U16X-TURIN2 B300 — the AMD counterpart of the GNR2 in the same size of chassis, for EPYC 9005 or 9004 processors.
• Gigabyte G4L4-SD3-LAX7 — two Intel Xeon 6700 or 6500 processors, built for direct liquid cooling with a manifold and leak detection inside.
• Gigabyte G4L4-ZD3-LAX7 — the AMD EPYC 9005 or 9004 version of the same liquid-cooled design.
• ASUS XA NB3I-E12 — two Intel Xeon 6700P processors, with ConnectX-8 InfiniBand embedded on the GPU board.

What separates them

The host platform first. The Cisco, the Supermicro SYS-822GS, the ASRock Rack GNR2, the Gigabyte SD3 and the ASUS run Intel Xeon 6; the Supermicro AS-8126GS, the ASRock Rack TURIN2 and the Gigabyte ZD3 run AMD EPYC. If the rest of the estate already runs one of the two, that usually decides it.

Then cooling. The two Gigabyte systems are built for direct liquid cooling; the Cisco and both Supermicro systems are air-cooled. Check that the room can supply what the system is designed for before choosing a model, not after.

Then the cluster network. In these systems the GPU network ports come from NVIDIA ConnectX-8 on the system itself rather than from add-in cards: Cisco calls them board-integrated, ASUS embedded on the GPU board. Gigabyte lists its ports as XDR InfiniBand or Ethernet, so say which fabric the cluster uses when you ask for a quote.

How a B300 order runs

A B300 system is supplied on request: the configuration and the allocation come from the manufacturer, and an end-user certificate is required.

We are not an authorised partner of NVIDIA or of the system makers, and we do not say we are. A B300 system is supplied through the manufacturer’s authorised partners.

To start, send:

• The system or systems you are considering, or the workload if the choice is still open.
• How many systems, and whether the cluster network is InfiniBand or Ethernet.
• The end user of the systems, for the end-user certificate.
• The power and cooling available per rack, and whether the room has a liquid-cooling loop.

What this page does not promise

No price, stock level or delivery date is published for a B300 system. For the wider choice between servers, which generation is that server and the Cisco UCS and Supermicro line pages compare the models around these, and the rack server catalogue lists the rest.