Computing power fp16
It is reported that a tlas 900 is composed of thousands of ascend processors, which can improve AI training efficiency by hundreds of thousands of times. Suppose a training takes 5 months, while using atlas 900 only takes tens of seconds! In the resnet-50 model training, the gold standard of AI computing ability, Atlas 900 completed the training in 59.8 seconds, 10 seconds faster than the original world record . For example, it's equivalent to the sprint champion finishing at the end of the race, drinking a bottle of water and waiting for the second place
In addition, Atlas 90 can also play an important role in some specific fields, such as autonomous driving, weather forecasting, oil exploration and so on. In fact, AI intelligent technology has been combined with all walks of life for a long time, and has developed like a raging firein July this year, the world's first drug completely developed by artificial intelligence (AI) has entered the human trial stage. The original development of a vaccine required thousands of people from pharmaceutical companies to work continuously for five years and cost hundreds of millions of dollars. With the help of AI technology, a small scientific research team only developed it in about two years. Ai not only simplifies the process of vaccine development, greatly reces the cost, but also improves the effectiveness of the vaccine
In terms of single performance, Xiaolong 855 is ahead of Xiaolong 855. It didn't need Apple's hand. It can be an engineering film with purple light
slashleaks has revealed a technical data that the unpublished engineering sample of Spreadtrum ud710 has outstripped other giants such as Xiaolong 855 / 845 and Qilin 970 in AI floating-point scoring
it's just that we should see Ziguang zhanrui's determination to attack high-end chips. We hope that the future Huben SOC and chunteng baseband proct lines can set a new example for local chips after Huawei
1. Different nature: ascend 910 is a new generation of AI open source computing framework, mindspre. Kirin chip adopts Hisilicon k3v2, which makes it one of the top smart phone processors
2. Different characteristics: the super-high computing power of shengteng 910's main cloud scene reaches 256 tflops. Kirin 9905g adopts 7Nm + EUV process and integrates 5g modem into SOC for the first time
3. Different principles: the computing power of shengteng 910 is twice that of the world's top AI chips, which is equivalent to 50 of the latest and strongest CPUs; Its training speed is also 50% - 100% higher than the latest strongest chip. The Hisilicon Kirin 990 processor will be manufactured by the 7Nm process of tsmc-2. With the use of V-Ray recorder, the overall performance of Hisilicon Kirin 990 processor will be about 10% higher than that of the previous generation of Hisilicon Kirin 980
extended data:
precautions:
shengteng 910ai chip belongs to ascend max series. The actual test results show that, in terms of calculation force, shengteng 910 fully meets the design specifications, that is, half precision (fp16) calculation force reaches 256 Tera flops, integer precision (int8) calculation force reaches 512 Tera ops. The important thing is that the power consumption required to reach the specification calculation force is only 310W, which is significantly lower than 350W of the design specification
through the technological innovation of mindspree framework and its collaborative optimization with shengteng processor, it effectively overcomes the complexity of AI computing and the diversity of computing power, realizes the high efficiency of running state, and greatly improves the computing performance. In addition to shengteng processor, mindspore also supports GPU, CPU and other processors
