Huawei intends to be a big player in the public cloud space and is starting out by building a new intelligent cloud hardware platform powered by its with own FusionServer gear.
The cloud platform, known as Atlas, was announced last week in Shanghai at Huawei Connect, the company’s annual technology conference. The focus of the event was cloud computing, or more precisely, “intelligent clouds” that would serve AI and HPC users. Although late to the utility computing game, the company intends to use Atlas as a jumping off point to eventually challenge companies like Amazon, Microsoft, Google, Alibaba, and IBM in the public cloud space.
But rather than just fielding zillions of generic servers, the company revealed it was basing the Atlas cloud on a new set of FusionServer boxes that combined x86 processors with GPUs. For example, Huawei will be using the FusionServer G5500, a 4U server that can house up to eight NVIDIA P100 or P40 cards that can be used to accelerate deep learning or more traditional high performance computing workloads. It also will employ the new FusionServer G2500, another 4U servers that can house up to 16 NVIDIA P4 GPUs and 240 TB of local storage. The P4 accelerators are specifically geared for inferencing neural networks, and in the case of the G2500, it will be tasked to do things like video analytics, using data gathered from vehicles and IoT devices.
Huawei will also be offering an FPGA-accelerated cloud service but it’s not clear which server will be called upon to house the reconfigurable silicon. The company currently offers a Xilinx FPGA-accelerated E9000 blade, a converged architecture server aimed at cloud computing, HPC, and high-end enterprise use, so that’s one possibility. The target application for the Atlas FPGA service appears to be inferencing, once again, but the reconfigurable aspect could theoretically be employed for a variety of throughput-demanding applications, including HPC.
In any case, the principle rationale behind Atlas is to offer a heterogeneous platform for the types of compute-demanding tasks that Huawei believes will dominate cloud computing from here on out. At the Huawei Connect event, Qiu Long, President of the company’s IT server product line, explained that cloud platforms will have to become more diverse to meet these needs.
"As the AI era is approaching, traditional hardware cannot meet the requirements on development of AI and cloud technologies,” he said. "Positioned as a new-generation intelligent cloud hardware platform, Huawei's Atlas pioneers heterogeneous resource pooling and intelligent orchestration technologies to bring resource utilization and performance to new high levels.”
The reference to resource pooling and intelligent orchestration refers to Atlas’s other big potential draw, namely turning the compute and storage in these servers into virtual resources. So, for example, for a financial analytics application, half a dozen CPUs on multiple servers may require only one GPU or FPGA, but for a deep learning application, a single CPU could tap into eight or more of these accelerators. All of that allocation and provisioning is taken care of by Atlas’s system software.
Of course, catching up to the current cloud computing leaders will be a formidable challenge. Huawei does have the advantage of an existing telco customer base, which, according to the company. encompasses about half of the world’s network traffic. It also has the advantage of manufacturing its own cloud server gear, which no other large cloud provider, save IBM, can claim to do.
Whether this will be enough to vault Huawei into the upper echelons of cloud computing, remains to be seen. For the near-term Amazon will continue to offer the largest and most mature base of GPU-accelerated instances, and Microsoft will offer the same for FPGA-accelerated cloud infrastructure. Nonetheless, a serious competitor like Huawei could press some of its unique advantage and eventually shake up the market. Time will tell.