Live DB — analyst tables
Test
Google 65536 TPUv4i NIC - 100G Hyperscaler Ethernet Cluster Configuration

GOOG-TPUv4i-200G-Eth-WB-256

Google · TPUv4i · NIC - 100G

Rack type
Cluster / location
TPUv4i Cluster - 2D Torus
Buyer profile
Hyperscaler
Networking type
Ethernet
Number of XPUs
65,536
NIC used
NIC - 100G
NICs / XPU
0.5
Scale-out BW / XPU
50G
XPU / server
256

Scale-up network

Assigned blockDefault
TPUv4i
Scale-UpXPU: TPUv4i0 linesshared by 1 SKU

Scale-up Units are based on the scale-up world size (256 XPUs).

XPU World Size

2
WS
Scale-Up Network LayerPart / Model
Units#
Unit BWGbit/s uni-d
Lane Count#
Lane BWGbit/s uni-d
XPUs per Node·256···
Scale-Up World Size·256···

Switches

2
SW
Scale-Up Network LayerPart / Model
Units#
Unit BWGbit/s uni-d
Lane Count#
Lane BWGbit/s uni-d
per World64x64 Port OCS48···
per Cluster·341.333···

Google ICI Networking

3D Torus Worksheet
8

Scale-Up Network Worksheet (per Scale-Up World Size) — per-TPU ICI connectivity by connector type, from the config workbook.

ConnectorPart / ModelTPUs
per TPU Units#
Units#
Unit BWGbit/s
Lane Count#
SerDes BWGbit/s
Attach Ratio# per GPU
Cluster Total#
TPU Rack64
Interior TPUs112
Copper Cable200G Copper2112200450
PCBPCB2112
Optical
Corner TPUs144
Copper Cable200G Copper2144400850
PCBPCB2144
Optical
Total for TPU Rack64
Copper Cable200G Copper2564004100
ActiveNo Active Copper for 400G
PassiveGeneric 200G DAC25640041004262,144
PCBPCB144
Optical

Scale-Up Parts Summary

1
ItemQty/ClusterQty/ServerAttach/XPU
Generic 200G DAC262,1441,024.004.00

Backend network

7 rows
Assigned blockDefault
TPUv4i - 100G NIC - 4L - ETH - TH2 - GOOG
BackendXPU: TPUv4i7 linesshared by 1 SKU

Backend Units are based on the cluster size (65,536 TPUs).

Switches

7
SW
 Network Layer
Switch and Server SpecAggregate Port UtilizationNetwork 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
TPUs·65,53650100032,76832,768032,76832,76803,276.8·
TPUs per Node·4····1······
NICs per NodeNIC - 100G2100100··1······
ToRWhiteBox Tomahawk 2 6.4T2,0486,40010016166432,76832,768131,0723,276.83,276.81
LeafWhiteBox Tomahawk 3 12.8T51212,800100646412832,76832,76865,5363,276.83,276.81
SpineWhiteBox Tomahawk 3 12.8T51212,800100646412832,76832,76865,5363,276.83,276.81
4 Aggregation Blocks
Core136x136 Port OCS256·100128128032,76832,76803,276.83,276.81

Transceivers

10
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 NodesNone - AECs are Used ->32,76865,5361004253,276.8
Leaf
DownlinkGeneric 100G DR4 QSFP2832,768641004253,276.8
UplinkGeneric 100G DR4 QSFP2832,768641004253,276.8
Spine
DownlinkGeneric 100G DR4 QSFP2832,76864100#ERR253,276.8
UplinkGeneric 100G FR432,76864100#ERR253,276.8
Core
Downlink·32,768128100··0
Uplink·32,768128100··3,276.8
Super-CoreOCS is Circular, 128 North, 128 South······
ToR
DownlinkNone - AECs are Used ->32,768161004253,276.8
UplinkGeneric 100G DR4 QSFP2832,768161004253,276.8

Cables

10
CB
 Network Layer
Cable/ FiberType
Total Units#
Units per Transceiver#
Unit BWGbit/s
Lane Count#
SerDes BWGbit/s
XPU NodesGeneric 100G AEC32,7681100425
Leaf
DownlinkAlready counted above·····
UplinkGeneric 100G Single Mode32,7681100425
Spine
DownlinkAlready counted above·····
Uplink·32,7681100425
Core
DownlinkAlready counted above·····
Uplink······
Super-Core······
ToR
DownlinkAlready counted above·····
UplinkGeneric 100G Single Mode32,7681100425

Backend Parts Summary

7
ItemQty/ClusterQty/ServerAttach/XPU
Generic 100G AEC32,768128.000.50
Generic 100G DR4 QSFP28131,072512.002.00
Generic 100G Single Mode65,536256.001.00
NIC - 100G5122.000.01
WhiteBox Tomahawk 2 6.4T2,0488.000.03
WhiteBox Tomahawk 3 12.8T1,0244.000.02
Generic 100G FR432,768128.000.50

Frontend network

No converged frontend + storage fabric data for this cluster.

Frontend Parts Summary

0
ItemQty/ClusterQty/ServerAttach/XPU

Compute Tray Networking Content

1 rows

Networking components inside the compute tray — accelerators, board-to-board connectors, retimers, and their per-tray bandwidth.

Assigned blockDefault
TPUv4i - GOOG
Compute TrayXPU: TPUv4i1 lineshared by 1 SKU
Tray layerPart / ModelUnits per Tray
Unit BWGb/s
Lanes
Lane BWGb/s
Per switch
TPU Tray to Host TrayPCIe 3.0 x16 Cable4512···

Compute Tray Parts Summary

1
ItemQty/ClusterQty/ServerAttach/XPU
PCIe 3.0 x16 Cable···