Skip to content
EMARQUE AI

NVIDIA

NVIDIA GB300 NVL72

Thirty-six NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs in one liquid-cooled rack, linked by NVLink. We source it through NVIDIA's OEM partners and coordinate the install.
  • 72 × NVIDIA Blackwell Ultra GPUs and 36 × NVIDIA Grace CPUs per rack
  • 20 TB HBM3e, 37 TB of fast memory in the rack (NVIDIA figures)
  • Fifth-generation NVLink Switch, 130 TB/s across the rack
  • ConnectX-8 SuperNICs with Quantum-X800 InfiniBand or Spectrum-X Ethernet
  • Liquid-cooled, plumbed into a coolant distribution unit

The price depends on the rack configuration, the networking between racks and the software licences you need.

NVIDIA GB300 NVL72 rack

Key features

NVIDIA GB300 NVL72 rack

72 GPUs, one memory pool

Every GPU in the rack is linked by fifth-generation NVLink and addresses 20 TB of HBM3e as one pool, or 37 TB of fast memory once the Grace CPUs' memory is counted (NVIDIA figures).
NVIDIA Blackwell Ultra GPU package

What Blackwell Ultra adds

NVIDIA states that Blackwell Ultra delivers 1.5x the dense FP4 throughput and twice the attention performance of NVIDIA Blackwell.
  • 130 TB/s inside the rack

    NVLink Switch carries 130 TB/s between the GPUs, so a model spread across the rack is not waiting on a network between machines (NVIDIA figure).

  • We source it and set it up

    We order the rack through NVIDIA's OEM partners, review your site with them, and coordinate delivery and installation in Malaysia. The warranty is the maker's, and we handle it from here.

Two ways to buy

A GB300 NVL72 rack comes two ways: NVIDIA's own DGX GB300, or the same rack from one of NVIDIA's OEM partners.

NVIDIA DGX GB300 racks
NVIDIA

NVIDIA DGX GB300

NVIDIA's own build of the GB300 NVL72 rack, configured and supported by NVIDIA. It ships with NVIDIA Mission Control, NVIDIA AI Enterprise, DGX OS and three years of NVIDIA support.

  • Configured by NVIDIA, supported by NVIDIA
  • NVIDIA Mission Control, NVIDIA AI Enterprise and DGX OS included
  • The operating model NVIDIA documents for NVIDIA DGX SuperPOD with GB300
  • Three years of NVIDIA business-standard support
A row of GB300 NVL72 racks
OEM GB300 NVL72 racks

GB300 NVL72 from NVIDIA's OEM partners

The same rack platform from NVIDIA's OEM partners, for teams that run their own software stack. Same 72 GPUs, 36 CPUs and NVLink Switch fabric; the warranty and support model are the OEM's.

  • Same rack platform, OEM warranty and support
  • Your own orchestration (Slurm, Kubernetes)
  • NVIDIA AI Enterprise licensed separately
  • We order it through the OEM and coordinate the install

Use cases

  • Training

    Models too large for one server

    A model spread across all 72 GPUs in the rack, trained without crossing a network between separate machines.

  • Inference

    Long context at production scale

    Large models served with long context windows, held in the rack's 20 TB of GPU memory.

  • Clusters

    A building block for NVIDIA DGX SuperPOD

    NVIDIA designs the rack as the unit a larger cluster is built from, joined by NVIDIA Quantum-X800 InfiniBand or NVIDIA Spectrum-X Ethernet.

Specifications

NVIDIA GB300 NVL72 rack
GPUs per rack
72 × NVIDIA Blackwell Ultra, one NVLink domain
CPUs per rack
36 × NVIDIA Grace, 2,592 Arm Neoverse V2 cores
GPU memory
20 TB HBM3e (NVIDIA figure)
Fast memory
37 TB in the rack, GPU and CPU together (NVIDIA figure)
FP4 compute
1,440 PFLOPS with sparsity, 1,080 PFLOPS dense (NVIDIA figure)
FP8 compute
720 PFLOPS with sparsity (NVIDIA figure)
NVLink fabric
Fifth-generation NVLink Switch, 130 TB/s across the rack (NVIDIA figure)
Networking (rack to rack)
NVIDIA ConnectX-8 SuperNICs with NVIDIA Quantum-X800 InfiniBand or NVIDIA Spectrum-X Ethernet
Cooling
Liquid-cooled, plumbed into a coolant distribution unit
Software
NVIDIA DGX GB300 includes NVIDIA Mission Control, NVIDIA AI Enterprise and DGX OS. OEM builds license NVIDIA AI Enterprise separately
Site
Three-phase power, a coolant loop and floor loading, reviewed with the OEM before quotation

Questions about the GB300 NVL72

  • What is NVIDIA GB300 NVL72?
    One liquid-cooled rack holding 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs, linked by NVLink so the GPUs can work on one model together. NVIDIA lists it as available now.
  • How is it different from NVIDIA DGX B300?
    DGX B300 is one air-cooled 10U server with eight GPUs. This is a whole liquid-cooled rack of 72, where the GPUs are joined by NVLink Switch instead of by a network between separate machines.
  • What does the site need?
    A coolant loop to a distribution unit, three-phase power, and floor loading for a full rack. NVIDIA does not publish a power figure per rack, so we take the OEM's figure for your exact configuration and review the room with them before quoting.
  • Who supports it?
    The maker: NVIDIA on a DGX GB300, the OEM on an OEM rack. We arrange the order, coordinate the install and handle the support relationship from Malaysia.
  • What comes after it?
    NVIDIA Vera Rubin NVL72, which NVIDIA also lists as available now. We can quote either.

Get a quote for the GB300 NVL72.

Tell us what it needs to run and who will use it. We reply with a written quote in MYR.