The card is based on Enflame's Sisi so-called deep thinking unit (DTU). Based on a reconfigurable chip design approach, Enflame’s DTU computing core features 32 scalable intelligent processors, with 8 SIPs combined into 4 scalable intelligent clusters (SICs). On-chip scheduling algorithms move data to maximize SIP utilization and the component is built on a 12nm FinFET process from Globalfoundries Inc. with a total of 14.1 billion transistors in a 2.5D package.
The performance of the card is 20TFLOPS under single precision (FP32); 80TFLOPS under half precision and mixed precision (BF16/FP16), and the maximum power consumption is 225W.
Support is provided for common artificial training models such as CNN, RNN, LSTM, BERT, which can be used for training scenarios such as image, streaming data, and speech.
"At present we have worked closely with Tencent's projects for general artificial intelligence application scenarios, and will also expand to more AI application scenarios in the future," said Zhao Lidong, CEO of Enflame Technology
Enflame’s AI accelerator SoC, DTU on Globalfoundries' 12LP platform has been sampled and production is scheduled for early 2020 in Globalfon Fab 8 in Malta, New York. The Cludblazer T10 will be available in the first quarter of 2020.
Related links and articles: