Huawei details the AI ​​chips for training and inference



[ad_1]

To gain a foothold in an estimated $ 91.18 billion AI chips market by 2025, Huawei has introduced Ascend 910, a new Ascend-Max family chipset optimized for model training. IA, and Ascend 310., an Ascend-Mini series inference chip designed to handle tasks such as image analysis, OCR, and object recognition. He also announced MindSpore, an IT infrastructure designed to support the development of AI applications.

The two chips were detailed for the first time in October 2018, but this week marks their commercial debut. The Ascend 910 focuses primarily on data center workloads, while the Ascend 310 targets Internet-connected devices, such as smartphones, smartwatches, and other Internet of Things (IoT) devices.

"Ascend 910 [in particular] Behaving much better than expected, "said Eric Xu, Huawei's rotating chairman, at a press conference held this morning in Shenzhen, China. "There is no doubt that it has more computing power than any other artificial intelligence processor in the world."

According to Huawei, the Ascend 910 provides 256 teraflops of computing power for semi-precision (FP16) and 512 TOPS (tera-operations per second) operations in full precision calculations (INT8) while consuming approximately 310 watts. Huawei said that in a typical workout with ResNet-50, a benchmark in image recognition, it is about twice as fast as rival AI accelerator chips.

By comparison, the Baidu Kunlun 818-300 AI chip, announced last July, can reach 260 TOPS, while Amazon's Inferentia machine learning processor offers scalable performance ranging from 32 TOPS to 512 TOPS. 'INT8.

The Ascend 310, on the other hand, contains a 16-channel FHD video decoder and a capacity of 16 TOPS in full precision (INT8) and 8 teraflops in half-precision (FP16). Huawei said it was already used by the "leading" automakers in shuttle buses, new energy vehicles and autonomous driving. He indicates that his acceleration card and his Atlas Ascend 310 series server are now part of "dozens" of industrial solutions. developed by dozens of partners.

The performance of the Ascend 310 places it slightly behind the Hailo-8 recently announced by the start-up Hailo, capable of handling 26 operations per second (TOP), more than double that of chips like the 2801 Lightspeeur Gyrfalcon, which reaches a maximum of 9.3 TOP, and the CEVA NeuPro. , which can reach 12.5 TOP. And it is slower than the Nvidia test chip, called RC-18, which can reach 128 TOP to 10 TOP per watt.

MindSpore

Alongside the Ascend 910 and Ascend 310 systems, Huawei developed MindSpore, a scalable end-to-end computing infrastructure similar to Google's TensorFlow and Facebook's PyTorch. It is suitable for all devices, peripheral environments and cloud environments, and is designed to be lightweight. MindSpore generates 20% fewer lines of code compared to "leading" frameworks when it comes to a typical AI model for natural language processing, which results in an efficiency gain of 50% on average.

Huawei stated that MindSpore did not process the data itself, but instead used template protection technology to integrate only the already-processed template and gradient information. In this way, it apparently preserves sensitive data even in "multi-scenario" environments, while ensuring the security and reliability of artificial intelligence algorithms.

MindSpore will be available in open source in the first quarter of 2020, with support for graphics processors, processors, and other processor types, in addition to Huawei's Ascend chips. It will complement Huawei's fully managed ModelArts platform, which provides full model production services, including data collection and model development.

"Any advance as planned. We promised a full AI portfolio, across all scenarios. And today we have delivered, "said Xu. This launch is a new step in Huawei's roadmap on artificial intelligence; it's also a new start … We want to promote the wider adoption of AI and help developers do what they do best. "

The unveiling comes weeks after the announcement of HarmonyOS, Huawei's internal operating system for mobile devices, embedded systems, smart speakers and apparel devices. In addition, the United States is considering easing trade restrictions on the company and its subsidiaries. early as this year. However, escalating tensions with China threaten to materialize these short-term projects – Bloomberg announced yesterday that the White House would postpone its decision to grant licenses to allow US companies (including Google and Intel) to resume their activities with Huawei. in the light of China's decision to stop purchases of US agricultural products.

Huawei's founder and CEO, Ren Zhengfei, said in June that the US ban would cost Huawei about $ 30 billion in lost revenue.

[ad_2]

Source link