NVIDIA GB300 NVL72 — Liquid-Cooled Data Center Rack
48U MGX rack · 18 compute trays · 9 NVLink switch trays · CDU cooling loop · 72× B300 GPUs · 135 kW TDP · 1.1 ExaFLOPS FP4
Raised floor · Reinforced concrete · 21 kN/m² rated · ~3,000 lb rack weight
NVIDIA GB300 NVL72 · 48U MGX Rack
1U
5U
9U
13U
17U
21U
25U
29U
33U
37U
41U
45U
48U
Coolant Distribution Manifold Header
DI water supply / return · QD fittings · leak detection sensors
NVLink Switch Tray 1 — 5th-gen NVLink · 1.8 TB/s per GPU
B300
B300
B300
B300
cold plate
CX8
CX8
CX8
CX8
BF3
Grace×2
Tray 1
B300
B300
B300
B300
CX8
CX8
CX8
CX8
BF3
Grace×2
Tray 2
NVLink Switch Tray 2
B300
B300
B300
B300
CX8
CX8
CX8
CX8
BF3
Grace×2 Tray 3
B300
B300
B300
B300
Grace×2 Tray 4
NVLink Switch Tray 3
B300
B300
B300
B300
Tray 5
B300
B300
B300
B300
Tray 6
NVLink Switch Tray 4
B300
B300
B300
B300
Tray 7
B300
B300
B300
B300
Tray 8
NVLink Switch Tray 5
B300
B300
B300
B300
Tray 9
B300
B300
B300
B300
Tray 10
NVLink Switch Tray 6
B300
B300
B300
B300
Tray 11
B300
B300
B300
B300
Tray 12
NVLink Switch Tray 7
B300
B300
B300
B300
Tray 13
B300
B300
B300
B300
Tray 14
NVLink Switch Tray 8
B300
B300
B300
B300
Tray 15
B300
B300
B300
B300
Tray 16
NVLink Switch Tray 9 (final)
B300
B300
B300
B300
Tray 17
B300
B300
B300
B300
Tray 18
Power Shelves × 8 (33 kW each = 264 kW capacity)
PSU 1
PSU 2
PSU 3
PSU 4
PSU 5
PSU 6
PSU 7
PSU 8
DC Bus Bar — 48V DC distribution to all trays
IEC 60309 5-wire whips · 135 kW TDP · 155 kW peak · provision to 192 kW
In-Rack CDU — Coolant Distribution Unit
Liquid-to-liquid HX · 250 kW capacity
Facility water IN 18–22°C / OUT 40–45°C · 50 µm filtration
SUPPLY
RETURN
Cable management · 800G OSFP LPO fiber → Arista 7060XE7
LPO optics: no DSP retiming · ~60% lower power · MRC plane wiring
COLD SUPPLY ←
→ HOT RETURN
Arista 7060XE7
MRC Fabric Switches
128×800G · 102.4 Tbps · 267MB buffer
Leaf P1 — 9× 800G server downlinks
Leaf P2 — 9× 800G server downlinks
Leaf P3 — 9× 800G server downlinks
Leaf P4 — 9× 800G server downlinks
Leaf P5 — 9× 800G server downlinks
Leaf P6 — 9× 800G server downlinks
Leaf P7 — 9× 800G server downlinks
Leaf P8 — 9× 800G server downlinks
↑ uplinks to 8× spine switches
No cross-plane links — fully independent
MRC 32-bit EV / SRv6 uN SID · 128–256 EVs per QP · zero ECMP · no dynamic routing
LPO optics · 840 ns latency
7060XE7 64×1600G / 128×800G
800G OSFP LPO
Power & Thermal
135 kW TDP / 155 kW peak
1,400 W per B300 GPU
~3,000 lb / ~1,360 kg total weight
GPU Memory
288 GB HBM3e per B300
21 TB total HBM per rack
8 TB/s memory bandwidth / GPU
Compute
15 PFLOPS FP4 per B300
1.1 ExaFLOPS FP4 / rack
130 TB/s NVLink / rack
MRC Networking
8× CX-8 SuperNIC @ 800G per tray
57.6 Tbps aggregate fabric
2-hop any-to-any · zero ECMP · no dynamic routing
In-Row CDU
1.8 MW capacity
for 8-rack SuperPOD cluster
576 GPUs · 11.5 ExaFLOPS
Facility chilled water loop
Hot / cold aisle containment
Row busway: Starline / Legrand