Hot on the heels of other cloud giants, Tencent on Wednesday introduced three system-on-chips designed in house that will first complement and then replace various processors designed by western companies like Intel or Nvidia. The list of SoCs includes an AI accelerator, a video transcoding processor, and a network processing unit.
Chinese Internet conglomerate Tencent offers dozens of various cloud services and operates massive datacenters in China. Just like other cloud giants, Tencent has been developing its own datacenter infrastructure for a while. Now, in in a bid to increase performance, cut power consumption, reduce costs, and get a complete control over its software and hardware, Tencent has ventured into chip design.
Tencent's initial lineup of SoCs will include three devices that will be used in its own datacenters, reports South China Morning Post:
Zixiao — an AI accelerator that can process images, videos and natural language. The chip is already in 'trial production' and performance is 100% better than competing offerings, according a Tencent executive who did not elaborate.
Canghai — a video transcoding unit that is said to be 30% better than existing offerings when it is ready.
Xuanling — a high-performance network interface controller for cloud datacenters than will be four times better than similar products, according to Tencent, though it is unclear which criteria were used for the comparison.
Just like other Internet conglomerates, Tencent has been historically using processing hardware from companies like AMD, Broadcom, Cisco, Intel, and Nvidia for the majority of its datacenter needs. Once it gets its own AI, video processing, and network processing into its datacenters, it will naturally reduce (and eventually cease) buying chips from the aforementioned U.S.-based companies.
For now, Tencent's ambitions do not officially include general-purpose processors or graphics processing units. Meanwhile, Chinese companies like Alibaba and Huawei already run homegrown general-purpose SoCs in their datacenters.