Skip to main content

Iluvatar TianGai 150 (BI-V150)

Product Overview

TianGai 150 (model BI-V150) is Iluvatar's second-generation cloud general-purpose GPU training accelerator released in 2023, an upgraded version of TianGai 100 (BI-V100). Using general-purpose GPU architecture, compatible with international mainstream GPU general computing models, supporting domestic and international mainstream AI ecosystems and deep learning frameworks and native operators, with significant performance improvements compared to the previous generation.

Key Features: As a next-generation product, TianGai 150 has improvements in FP16 compute and memory bandwidth, capable of handling large-scale AI training and complex computing requirements.

Core Specifications

ItemParameter
ArchitectureIluvatar self-developed general-purpose GPU architecture (TianGai 100 upgrade)
Process7nm (estimated, same as TianGai 100)
FP16Not disclosed (estimated 150–200 TFLOPS)
INT8Not disclosed (estimated 300–400 TOPS)
Memory CapacityEstimated 32–64 GB HBM2/HBM3
Memory BandwidthNot disclosed
TDP275 W (estimated)
InterfacePCIe Gen4 (estimated)
Release2023
Mass ProductionSince 2023
Software StackIluvatar computing software stack (PyTorch/TensorFlow compatible)

⚠️ Specification Note: Detailed specifications of TianGai 150 are not fully disclosed by official sources. Some values above are estimates, subject to Iluvatar's subsequent official data sheet.

TianGai 100 vs TianGai 150

MetricTianGai 100 (BI-V100)TianGai 150 (BI-V150)Improvement
Release20212023Next generation
FP16128 TFLOPSEstimated 150–200 TFLOPS+20–55%
INT8256 TOPSEstimated 300–400 TOPS+20–55%
Software EcosystemInitialMore matureIterative optimization

Hybrid Training Case

TianGai 150 has participated in actual large-scale training deployments:

  • Zhiyuan Research Institute Aquila2-70B-Expr large model hybrid training: Using 120-node BI-V100 cluster + 8-node BI-V150 cluster for hybrid training, hybrid cluster performance reached 85.3% of theoretical limit (ideal pool theoretical peak), demonstrating excellent heterogeneous computing power hybrid training capability.

Application Scenarios

  • Large-scale AI training (below 100 billion parameter models)
  • Heterogeneous computing clusters (hybrid deployment with TianGai 100)
  • Scientific computing (FP32/INT32 support)
  • Smart city (traffic flow analysis, environmental monitoring)
  • Data center (cloud computing, big data acceleration)
  • Ultra-high compute requirements (FP16 still lower than H100/A100)
  • Emerging FP8 precision (FP8 not supported)

Product Evolution

ProductReleaseFP16 TFLOPSStatus
TianGai 100 (BI-V100)2021128 TFLOPSOn sale
TianGai 150 (BI-V150)2023Estimated 150–200 TFLOPSCurrent mainstream
TianGai 200 (BI-V200)TBDNot disclosedNext generation (estimated)

References