01 / INSIDE VERA RUBIN
Vera
Rubin
An AI system must hold context as well as perform calculations. Follow the hardware behind longer conversations and more demanding reasoning.
Follow the data inside ↓02 / SYSTEM ARCHITECTURE
A closer
perspective.
The service front faces you. One compute tray slides out of its bay, leaving the rear interconnects behind. Now the architecture comes into view.
FROM THE SERVICE FRONT
03 / BENEATH THE SURFACE
A system built
around data.
The enclosure, retention covers and cooling assembly reveal the working layers beneath. Each has its own material, scale and purpose.
SCROLL TO REVEAL THE NEXT LAYER
04 / THERMAL DESIGN
Make room
for performance.
Liquid cooling carries heat away from the compute tray. The cold-plate assemblies lift clear to reveal the processors and memory beneath.
05 / THE MEMORY HIERARCHY
Different layers.
One data path.
HBM sits beside the GPUs. SOCAMM2 modules 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 HBM4
Keep reasoning
in motion.
The GPU repeatedly accesses model weights and working data. HBM4 widens the path beside the accelerator, connecting stacked memory through a 2,048-bit interface.
Exploded view · soldered in the physical system
07 / MICRON SOCAMM2
Make room
for context.
Longer conversations leave more context to retain. SOCAMM2 adds low-power memory capacity beside the CPU, supporting a broader memory hierarchy when workloads use it.
08 / MICRON STORAGE
Open a faster
path to data.
A dataset is only useful when the system can reach it. The 9650 SSD brings PCIe Gen6 bandwidth to the persistent tier—the start of the next movement of data.
09 / THE CONNECTED SYSTEM
A system works
together.
Four cache drives keep data near compute. ConnectX-9 handles the cluster network, while BlueField-4 supports storage and management. Rear NVLink connections join the rack.
10 / THE COMPLETE PICTURE
One system.
Many possibilities.
An answer depends on more than compute.
It depends on what the system can keep close.