GB300

02 / SYSTEM ARCHITECTURE

A closer
perspective.

A compute tray slides forward from its service bay. Inside, different kinds of memory serve different parts of the workload.

FROM THE SERVICE FRONT

03 / BENEATH THE SURFACE

Performance starts
with proximity.

Lift the enclosure. Beneath the cooling assembly, memory sits close to the processors it serves. Short distances matter when data moves continuously.

SCROLL TO REVEAL THE NEXT LAYER

04 / THERMAL DESIGN

Move heat.
Sustain work.

Liquid-cooled plates sit over the CPUs and GPUs. Airflow supports the other electronics. Both paths help the system sustain demanding workloads.

05 / THE MEMORY HIERARCHY

Different layers.
One data path.

HBM sits beside the GPUs. LPDDR5X packages serve CPU-side work. NVMe drives retain the larger dataset. Follow each layer as the system opens into an exploded view.

Illustrative separation shows assembly relationships.

06 / MICRON HBM3E

Keep the
accelerator fed.

Weights and intermediate results must reach the GPU again and again. HBM3E places high-bandwidth memory beside the accelerator, helping keep its compute resources supplied.

Exploded view · soldered in the physical system

07 / MICRON LPDDR5X

Give the CPU
room to work.

The CPU coordinates another part of the work. LPDDR5X provides its working memory with an emphasis on power efficiency. These mounted packages stay seated as we move closer.

Exploded view · soldered in the physical system

08 / MICRON STORAGE

Bring the data
into play.

Before data enters memory, it has to come from somewhere. The 9550 SSD provides a fast, persistent source for the next dataset, cache or checkpoint.

09 / THE CONNECTED SYSTEM

Beyond a
single server.

ConnectX-8 carries cluster traffic. BlueField-3 supports storage and management. At the rear, NVLink connects compute trays through the rack’s switch fabric.

10 / THE COMPLETE PICTURE

One system.
Many possibilities.

Storage keeps it. System memory prepares it.
HBM brings it close to compute.