This article is more than 1 year old

Tencent Cloud says it's mass producing custom video chips

Chinese tech giant claims better performance than competing GPUs

Chinese social media, cloud, and entertainment giant Tencent on Monday revealed that it has started mass production of a home brew video transcoding accelerator.

The announcement comes nearly two years after the company unveiled a trio of custom chips designed to accelerate everything from streaming video to networking and artificial intelligence workloads.

In a post published on WeChat, Tencent Cloud revealed that "tens of thousands" of its Canghai chips, which are designed to offload video encode/decode for latency sensitive workloads, have been deployed internally to accelerate cloud gaming and live broadcasting.

Tencent says the Canghai chip can be paired with GPUs from a variety of vendors to support low-latency game streaming. When used for video transcoding, Tencent said a single node equipped with Canghai can deliver up to 1,024 video channels . We'll note that Nvidia, with the launch of its L4 GPUs last month, made similar claims. Without real-world benchmarks, it's hard to say how either firm's claims stack up.

With that said, Tencent says its chip can achieve superior quality at the same bitrate as competing GPU-based accelerators, with latencies for a single 1080p frame of just four milliseconds. But again, it's not stated which GPUs it's comparing its results against.

Whatever the real world results, the promised improvement in compression efficiency should allow Tencent to maintain current streaming quality while reducing the bandwidth consumed — although some of this improvement could be down to support for better codecs.

The Canghai chip reportedly supports H.264, HEVC, and AV1. As we've previously reported, AV1 is a relatively new codec that in addition to being royalty free, is also exceptionally efficient. Testing puts AV1 at somewhere between 20 and 40 percent more efficient compared to popular web streaming codecs, like HEVC.

Custom compute for the Chinese cloud

When it comes to spinning custom chips to improve the efficiency and economics of cloud computing, Amazon Web Services gets a lot of credit. The American e-tail giant and cloud titan has developed everything from custom CPUs, AI training and inference accelerators, and smartNICs to offload many housekeeping workloads.

And while Google has developed an accelerator of its own, called the Tensor Processing Unit (TPU), most US cloud providers have largely stuck with commercially available parts from the likes of Intel, AMD, Ampere, Broadcom, or Nvidia, rather than designing their own.

However, in China, custom chips appear to be more prevalent, with development an imperative accelerated by US sanctions that mean some tech products can't be exported to the Middle Kingdom.

In addition to its Canghai video chips, Tencent also created the Zixiao AI "reasoning" chip which it's deployed internally, plus the Xuanling network processor. The latter is designed to offload and accelerate networking, storage, and compute workloads, which sounds an awful lot like a smartNIC.

Tencent's local peer Alibaba Cloud has also created custom silicon. In April 2022 the cloud provider showed off its custom Arm CPUs and DPUs.

The cloudy business unit of e-commerce giant Alibaba began previewing its Yitian 710 processor, which reportedly features 128 Armv9 cores running at 3.2GHz, eight channels of DDR5, and 96 lanes of PCIe 5.0, about a year ago in its Elastic Compute Service.

Meanwhile, Alibaba's Cloud Infrastructure Processing Unit (CIPU), much like AWS Nitro cards, is designed to offload virtualization functions associated with storage, networking, and security that would otherwise be handled by the main CPU.

And while not fully custom, Baidu has tapped RISC-V chipmaker StarFive to help it develop high-performance products based on the open ISA. However, it remains to be seen when, or whether, these chips will come into play. ®

More about

TIP US OFF

Send us news


Other stories you might like