Skip to main content

Enflame CloudBlaze T10

Enflame's first AI training accelerator, based on self-developed Suisi 1.0 chip, released in December 2019, one of China's earliest mass-produced high-performance AI training cards.

Core Specifications​

SpecificationValue
ArchitectureSuisi 1.0 (DTU dataflow architecture)
ProcessGF 12nm FinFET
Transistor Count14.1 billion
TDP225 W
Memory32 GB HBM2
Memory Bandwidth512 GB/s
FP32 Compute20 TFLOPS
FP16 / BF16 Compute80 TFLOPS
INT8 Compute160 TOPS (estimated)
InterfacePCIe 4.0 x16, dual-slot
InterconnectEnflame ESL 200GB/s (bidirectional)
Launch Date2019-12-11
PriceNot disclosed (estimated ¥40,000)

Technical Highlights​

  • Suisi 1.0 chip: Enflame's first-generation self-developed cloud AI training chip, using GF 12nm process, 2.5D advanced packaging, integrating 14.1 billion transistors
  • DTU dataflow architecture: Reconfigurable dataflow architecture designed for AI training, supporting dynamic task scheduling
  • ESL high-speed interconnect: Self-developed 200GB/s bidirectional interconnect technology, supporting 1024-node cluster training, linear acceleration ratio up to 86%
  • TopsRider software stack: Companion TopsRider 1.0 software platform, supporting TensorFlow, PyTorch and other mainstream frameworks
  • Training scenario coverage: Supports CNN, RNN, LSTM, BERT and other common training models, covering image, streaming data, speech and other scenarios

Product Positioning​

CloudBlaze T10 is Enflame's founding work, targeting NVIDIA V100, focusing on cloud AI training scenarios. As the earliest mass-produced high-performance AI training card in China, T10 filled the gap in domestic AI training chips when released in 2019.

Application Scenarios​

  • Large-scale deep learning model training (GPT, BERT, ResNet, etc.)
  • Internet recommendation system training
  • Financial risk control model training
  • Medical imaging AI training
  • Educational and research AI training clusters

Reference Price​

ChannelPriceDescription
Official pricingNot disclosedPlanned price ~$6000/card in 2019
Channel estimate≈ ¥40,000Estimated based on concurrent V100 pricing ratio

References​