72-GPU Blackwell Ultra MRC Rack — Arista 7060XE7 + Nvidia HGX B300
9 × HGX B300 (72 B300 GPUs) · 8× CX-8 SuperNIC per node @ 800G · 8 MRC planes · 16× 7060XE7 · 57.6 Tbps aggregate
Reference architecture aligned with OCP MRC Rev 1.0 and the
OpenAI MRC paper .
ConnectX-8 SuperNIC is Nvidia's reference platform (paper Clusters A/B on GB200).
This rack uses HGX B300 and Arista 7060XE7 — GPU and switch SKUs differ from the paper testbeds.
Server Rack
9× HGX B300 · 72 Blackwell Ultra GPUs
MRC Fabric — 8 independent planes
Arista 7060XE7 · 800G per uplink · no cross-plane links
P1
P2
P3
P4
P5
P6
P7
P8
Spine P1
Spine P2
Spine P3
Spine P4
Spine P5
Spine P6
Spine P7
Spine P8
7060XE7
7060XE7
7060XE7
7060XE7
7060XE7
7060XE7
7060XE7
7060XE7
← fully independent planes — no cross-plane links →
Leaf P1
Leaf P2
Leaf P3
Leaf P4
Leaf P5
Leaf P6
Leaf P7
Leaf P8
7060XE7
7060XE7
7060XE7
7060XE7
7060XE7
7060XE7
7060XE7
7060XE7
HGX B300 #9 — GPU 65–72
8× B300 · NVSwitch5 1.8TB/s · 8× CX-8 SuperNIC @ 800G
800G×8
HGX B300 #1 — GPU 1–8
8× B300 · NVSwitch5 1.8TB/s · 8× CX-8 SuperNIC @ 800G
800G×8
ConnectX-8 SuperNIC · Nvidia OCP MRC Rev 1.0 reference NIC · 800G per plane
72
B300 GPUs
9× HGX B300
72
CX-8 SuperNICs
800G each · 1:1 GPU:NIC
16
7060XE7 switches
8 leaf + 8 spine
57.6
Tbps total fabric
9 nodes × 6.4 Tbps (8×800G)
2
hops any-to-any
leaf → spine → leaf
1.0
OCP MRC Rev
validated on CX-8 SuperNIC
Cable = plane:
P1
P2
P3
P4
P5
P6
P7
P8
· CX-8 NIC n → leaf of plane n · 32-bit EV in SRv6 uN SID · 128–256 EVs/QP · no ECMP · no dynamic routing · liquid cooling req · 1,400W TDP