Intelligence Processing Unit
E660957
The Intelligence Processing Unit is a specialized processor architecture designed by Graphcore to accelerate artificial intelligence and machine learning workloads with highly parallel, memory-rich compute.
Statements (48)
| Predicate | Object |
|---|---|
| instanceOf |
AI accelerator
ⓘ
processor architecture ⓘ |
| architectureFeature | memory-rich compute ⓘ |
| architectureType | highly parallel processor ⓘ |
| benefit |
efficient memory bandwidth utilization
ⓘ
high compute density ⓘ scalability across many chips ⓘ |
| category |
machine learning hardware accelerator
ⓘ
parallel computing hardware ⓘ |
| competesWith |
graphics processing unit
ⓘ
tensor processing unit NERFINISHED ⓘ |
| countryOfOrigin | United Kingdom ⓘ |
| designedFor |
high-throughput training
ⓘ
large-scale machine intelligence models ⓘ low-latency inference ⓘ |
| designer | Graphcore NERFINISHED ⓘ |
| focusesOn |
graph-based machine learning
ⓘ
sparse computation efficiency ⓘ |
| hardwareParadigm | manycore architecture ⓘ |
| introducedBy | Graphcore IPU product line NERFINISHED ⓘ |
| keyConcept |
distributed on-chip memory
ⓘ
fine-grained parallelism ⓘ low-precision arithmetic support ⓘ |
| manufacturer | Graphcore NERFINISHED ⓘ |
| marketSegment |
AI infrastructure
ⓘ
data center accelerators ⓘ |
| memoryModel | distributed local memory per core ⓘ |
| optimizationTarget |
deep learning
ⓘ
neural network inference ⓘ neural network training ⓘ |
| primaryUse |
artificial intelligence workloads
ⓘ
machine learning workloads ⓘ |
| programmedWith |
Graphcore software stack
NERFINISHED
ⓘ
Poplar SDK NERFINISHED ⓘ |
| relatedCompany | Microsoft NERFINISHED ⓘ |
| supports |
data-parallel computation
ⓘ
model-parallel computation ⓘ |
| supportsFramework |
ONNX
NERFINISHED
ⓘ
PyTorch NERFINISHED ⓘ TensorFlow NERFINISHED ⓘ |
| typicalDeployment |
cloud data centers
ⓘ
on-premises AI clusters ⓘ |
| typicalUsers |
AI researchers
ⓘ
machine learning engineers ⓘ |
| usedBy | Microsoft Azure NERFINISHED ⓘ |
| vendorPlatform |
Graphcore IPU systems
NERFINISHED
ⓘ
Graphcore IPU-M2000 NERFINISHED ⓘ Graphcore IPU-POD NERFINISHED ⓘ |
Referenced by (1)
Full triples — surface form annotated when it differs from this entity's canonical label.