Statements (28)
| Predicate | Object |
|---|---|
| gptkbp:instanceOf |
Neural Engine
|
| gptkbp:announced |
2021
|
| gptkbp:architecture |
ASIC
|
| gptkbp:availableOn |
gptkb:Google_Cloud
|
| gptkbp:designedFor |
large-scale AI workloads
|
| gptkbp:energyEfficiency |
improved over TPU v3
|
| https://www.w3.org/2000/01/rdf-schema#label |
TPU v4
|
| gptkbp:interconnectBandwidth |
800 Gbps
|
| gptkbp:location |
data centers
|
| gptkbp:manufacturer |
gptkb:Google
|
| gptkbp:memoryPerChip |
32 GB HBM
|
| gptkbp:nodesPerPod |
4096
|
| gptkbp:openToPublic |
2022
|
| gptkbp:performanceContext |
up to 275 teraflops (BF16) per chip
|
| gptkbp:podSize |
up to 4096 chips
|
| gptkbp:predecessor |
gptkb:TPU_v3
|
| gptkbp:roadType |
gptkb:butter
|
| gptkbp:successor |
gptkb:TPU_v5e
|
| gptkbp:supports |
bfloat16
float32 int8 |
| gptkbp:usedBy |
gptkb:DeepMind
gptkb:Google_Research Google Cloud customers |
| gptkbp:usedFor |
gptkb:artificial_intelligence
gptkb:machine_learning |
| gptkbp:bfsParent |
gptkb:PaLM
|
| gptkbp:bfsLayer |
5
|