Live DB — analyst tables
Test
Neocloud Giant 2048 MI350X CX-10 Neocloud Giant Infiniband Cluster Configuration
NG-MI350X-CX7-IB-2L
Neocloud Giant · MI350X · CX-10
Rack type
—
Cluster / location
MI350X - Infiniband Cluster
Buyer profile
Neocloud Giant
Networking type
Infiniband
Number of XPUs
2,048
NIC used
CX-10
NICs / XPU
3
Scale-out BW / XPU
400G
XPU / server
8
Scale-up network
Assigned blockDefault
MI350X
Scale-UpXPU: MI350X0 linesshared by 8 SKUs
Scale-up Units are based on the scale-up world size (8 XPUs).
Note: Scale-up fabric (AMD Infinity Fabric): 7 links per GPU × 153.6 GB/s bidirectional = 1,075.2 GB/s bidirectional, i.e. 537.6 GB/s unidirectional per GPU. The 448 GB/s unidirectional figure is the prior-generation MI300X value (7 × 128 GB/s bidirectional); MI350X (CDNA4) upgraded the Infinity Fabric links to 153.6 GB/s. (AMD Instinct MI350X platform brief)
XPU World Size
2
WS| Scale-Up Network Layer | Part / Model | Units# | Unit BWGbit/s uni-d | Port Count# | Port BWGbit/s uni-d |
|---|---|---|---|---|---|
| XPUs per Node | · | 8 | 537.6 | 14 | 38.4 |
| Scale-Up World Size | · | 8 | 4,300.8 | · | · |
Switches
1
SW| Scale-Up Network Layer | Part / Model | Units# | Unit BWGbit/s uni-d | Port Count# | Port BWGbit/s uni-d |
|---|---|---|---|---|---|
| Scale-Up Switch Total | · | · | 0 | · | · |
Connections & cabling
3
CN| Scale-Up Network Layer | Part / Model | Units# | BW / PortGbit/s uni-d | Unit DP# | DP / Port# | Total DPs# | BW / Connector | Cable / Model | Cable Units# |
|---|---|---|---|---|---|---|---|---|---|
| Male Connectors | · | · | · | · | · | · | · | Already Counted Below | · |
| Female Connectors | · | · | · | · | · | · | · | Already Counted Below | · |
| Switch Total | |||||||||
| Male Connectors | · | · | · | · | · | · | · | Already Counted Below | · |
Notes
- XPUs per Node — >>Mesh, 7x153.6Gb/s bidirectional links per chip due to improved CDNA PHY from MI300X (Fabric-only upgrade)
Scale-Up Parts Summary
0
| Item | Qty/Cluster | Qty/Server | Attach/XPU |
|---|
Backend network
5 rows
Assigned blockDefault
MI350X - CX7 - 2L - IB - QM9700 - Generic
BackendXPU: MI350X5 linesshared by 2 SKUs
Backend Units are based on the cluster size (2,048 XPUs).
Switches
4
SW Network Layer | Switch and Server Spec | Aggregate Port Utilization | Network Bandwidth Aggregate | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
SwitchPart Name / Model | Units# | Unit BWGbit/s | Port BWGbit/s | Downlink Ports / SW# | Uplink Ports / SW# | Max Ports# | Downlink Ports# | Uplink Ports# | Total Ports# | Downlink BWTbit/s | Uplink BWTbit/s | OversubRatio | |
| XPUs | · | 2,048 | 400 | 400 | 0 | 2,048 | 2,048 | 0 | 2,048 | 2,048 | 0 | 819.2 | · |
| NICs per Node | ConnectX-7 NIC - 400G | 8 | 400 | 400 | · | · | 8 | · | · | · | · | · | · |
| Leaf | Quantum-2 InfiniBand (QM9700) | 64 | 25,600 | 800 | 16 | 16 | 32 | 1,024 | 1,024 | 2,048 | 819.2 | 819.2 | 1 |
| Spine | Quantum-2 InfiniBand (QM9700) | 32 | 25,600 | 800 | 32 | 0 | 32 | 1,024 | 0 | 1,024 | 819.2 | 0 | · |
Transceivers
6
CN Network Layer | Transceiver / CablePart Name / Model | Total Units# | Units per SW/GPU# | Unit BWGbit/s | Lanes# | BW per LaneGbit/s | Agg. BWTbit/s |
|---|---|---|---|---|---|---|---|
| XPU Nodes | Nvidia LinkX 400G SR4 | 2,048 | 2,048 | 400 | 2 | 200 | 819.2 |
| Leaf | |||||||
| Downlink | Nvidia LinkX 800G SR8 | 1,024 | 16 | 800 | 8 | 100 | 819.2 |
| Uplink | Nvidia LinkX 800G SR8 | 1,024 | 16 | 800 | 8 | 100 | 819.2 |
| Spine | |||||||
| Downlink | Nvidia LinkX 800G SR8 | 1,024 | 32 | 800 | 8 | 100 | 819.2 |
| Uplink | Nvidia LinkX 800G SR8 | 0 | · | · | · | · | · |
| Core | |||||||
| Downlink | · | · | · | · | · | · | · |
| Uplink | · | · | · | · | · | · | · |
| Super-Core | |||||||
Cables
6
CB Network Layer | Cable/ FiberType | Total Units# | Units per Transceiver# | Unit BWGbit/s | Lane Count# | SerDes BWGbit/s |
|---|---|---|---|---|---|---|
| XPU Nodes | Nvidia LinkX 400G Multimode | 2,048 | 4 | 400 | 4 | 100 |
| Leaf | ||||||
| Downlink | Already counted above | · | · | · | · | · |
| Uplink | Nvidia LinkX 400G Multimode | 2,048 | 2 | 400 | 4 | 100 |
| Spine | ||||||
| Downlink | Already counted above | · | · | · | · | · |
| Uplink | · | · | · | · | · | · |
| Core | ||||||
| Downlink | Already counted above | · | · | · | · | · |
| Uplink | · | · | · | · | · | · |
| Super-Core | ||||||
Backend Parts Summary
5
| Item | Qty/Cluster | Qty/Server | Attach/XPU |
|---|---|---|---|
| Nvidia LinkX 400G SR4 | 2,048 | 8.00 | 1.00 |
| Nvidia LinkX 400G Multimode | 4,096 | 16.00 | 2.00 |
| Nvidia LinkX 800G SR8 | 3,072 | 12.00 | 1.50 |
| ConnectX-7 NIC - 400G | 2,048 | 8.00 | 1.00 |
| Quantum-2 InfiniBand (QM9700) | 96 | 0.38 | 0.05 |
Frontend network
11 rows
Assigned blockDefault
MI350X - SN5600 - Generic
FrontendXPU: MI350X11 linesshared by 3 SKUs
Converged Frontend + Storage Fabric
11
| Layer | Part / Model | Units | Attach / XPU |
|---|---|---|---|
| · | Spectrum-4 51.2T (SN5600) | 25 | 0.0120 |
| · | Spectrum-3 12.8T (SN4700) | 0 | 0.0000 |
| · | Spectrum-2 6.4T (SN3700) | 4 | 0.0018 |
| · | Spectrum-2 1GbE (SN2201) | 30 | 0.0147 |
| · | Nvidia LinkX 800G SR8 | 1,189 | 0.5807 |
| · | Nvidia LinkX 400G SR4 | 1,183 | 0.5777 |
| · | Nvidia LinkX 100G SR4 QSFP28 | 525 | 0.2563 |
| · | Nvidia LinkX 800G Multimode | 1,628 | 0.7947 |
| · | Nvidia LinkX 400G Multimode | 8 | 0.0041 |
| · | Nvidia LinkX 100G Multimode | 234 | 0.1142 |
| · | RJ45 Cables | 1,443 | 0.7048 |
Frontend Parts Summary
11
| Item | Qty/Cluster | Qty/Server | Attach/XPU |
|---|---|---|---|
| Spectrum-4 51.2T (SN5600) | 25 | 0 | 0.0120 |
| Spectrum-3 12.8T (SN4700) | 0 | 0 | 0.0001 |
| Spectrum-2 6.4T (SN3700) | 4 | 0 | 0.0018 |
| Spectrum-2 1GbE (SN2201) | 30 | 0 | 0.0147 |
| Nvidia LinkX 800G SR8 | 1,189 | 5 | 0.5807 |
| Nvidia LinkX 400G SR4 | 1,183 | 5 | 0.5777 |
| Nvidia LinkX 100G SR4 QSFP28 | 525 | 2 | 0.2563 |
| Nvidia LinkX 800G Multimode | 1,628 | 6 | 0.7947 |
| Nvidia LinkX 400G Multimode | 8 | 0 | 0.0041 |
| Nvidia LinkX 100G Multimode | 234 | 1 | 0.1142 |
| RJ45 Cables | 1,443 | 6 | 0.7048 |
Compute Tray Networking Content
4 rows
Networking components inside the compute tray — accelerators, board-to-board connectors, retimers, and their per-tray bandwidth.
Assigned blockDefault
MI350X - Generic
Compute TrayXPU: MI350X4 linesshared by 8 SKUs
| Tray layer | Part / Model | Units per Tray | Unit BWGb/s | Lanes | Lane BWGb/s | Per switch |
|---|---|---|---|---|---|---|
| XPUs per Tray | Rubin 200 | 4 | · | · | · | · |
| PCIe 6.0 Connector (Male) | Paladin Board to Board 32 Pin 4p8c (Male) | 2 | 512 | · | · | · |
| PCIe 6.0 Connector (Male) | Paladin Board to Board 64 Pin 4p16c (Male) | 8 | 1,024 | · | · | · |
| PCIe 6.0 Connector (Female) | Paladin Board to Board 32 Pin 4p8c (Female) | 4 | 512 | · | · | · |
| PCIe 6.0 Connector (Female) | Paladin Board to Board 64 Pin 4p16c (Female) | 16 | 1,024 | · | · | · |
| PCIe 6.0 Connector (Male) | Paladin Board to Board 32 Pin 4p8c (Male) | 2 | 512 | · | · | · |
| PCIe 6.0 Connector (Male) | Paladin Board to Board 64 Pin 4p16c (Male) | 8 | 1,024 | · | · | · |
Notes
- PCIe 6.0 Connector (Male) · Paladin Board to Board 32 Pin 4p8c (Male) — 2 pins = 1DP
- PCIe 6.0 Connector (Male) · Paladin Board to Board 64 Pin 4p16c (Male) — 2DP = 1Lane
Compute Tray Parts Summary
4
| Item | Qty/Cluster | Qty/Server | Attach/XPU |
|---|---|---|---|
| Paladin Board to Board 32 Pin 4p8c (Male) | 2,048 | 8.00 | 1.00 |
| Paladin Board to Board 64 Pin 4p16c (Male) | 8,192 | 32.00 | 4.00 |
| Paladin Board to Board 32 Pin 4p8c (Female) | 2,048 | 8.00 | 1.00 |
| Paladin Board to Board 64 Pin 4p16c (Female) | 8,192 | 32.00 | 4.00 |