Enflame, which this summer represented accelerators based on the second generation of their DTU AI chips, has released a new inference accelerator Cloudblazer Yunsui i20 with a Suixi 2.5 chip. It is manufactured using the 12nm FinFET process of GlobalFoundries and has an updated high-performance architecture of computing cores GCU-CARE 2.0, thanks to which, according to the creators, it was possible to achieve efficiency comparable to the mainstream 7nm GPU.
Among the key features of the new product, the company notes the increased computing power, the ability to perform tensor, vector and scalar calculations, API for C ++ and Python, as well as support for major frameworks and model formats (TensorFlow, PyTorch, ONNX). The bundled software provides flexible migration options with support for virtualization technologies, as well as multi-user and multitasking environments with secure process isolation.
The Yunsui i20 features 16GB HBM2e memory with a bandwidth of up to 819GB / s. The new product supports all key formats and provides a universal inference platform, including for the clouds. Peak computing FP32 performance reaches 32 Teraflops, TF32 (it is not specified if it is about compatibility with NVIDIA) – 128 Teraflops, FP16 / BF16 – 128 Teraflops, and INT8 reaches 256 Tops. Compared to the first generation of products, Yunsui i20 increased FP performance by 1.8 times and INT computing by 3.6 times.
Optimization of the proprietary TopsRider software stack made a significant contribution to improving performance, due to which the load on the memory subsystem was reduced. As a result, the average performance of the models’ execution increased by 3.5 times, and the efficiency of using the computing power – on average, 2 times. In addition, the new programming model and automation technologies can accelerate development efficiency and reduce the cost of model migration. The company is convinced that all this will make the Yunsui i20 a more competitive solution.
Source: Oceanpine Capital
