Huawei shakes up AI with new server architecture

Kathy Gibson reports from AI Expo Africa in Sandton – Huawei is launching a new AI infrastructure the will offer a scalable platform for organisations of all sizes.

Clinton George, solution architect at Huawei SAT computing, marketing and solutions sales department, says Atlas AI will be launched towards the end of this year.

“It is a new option,” he explains. “This is not a replacement for anything. It is an alternative, and a different method of computing and running AI infrastructure.”

South African users are expected to be particularly keen on Atlas AI, since if offers than additional choice, with a range of new strengths.

AI has seen big changes over the last 14 years, George says.

In 2012, perception AI was involved in deep learning; the era of generative AI (GenAI) and large language models (LLMs), Ai was able to start creating content; the current, agentic AI, era is where AI can reason, plan and take autonomous actions.

“But this is not just a tech story – it’s a business story,” George says. “And AI is starting to deliver measurable business value.

“Today, AI is no longer experimental. It has moved from the lab to the balance sheet. And this is why the focus on the technology underneath it is so integral.”

As token usage continues to grow, organisations are looking to cost-effective options. Meanwhile, the performance gap between open source and closed source models has essentially disappeared, so open source is quickly becoming a mainstream global choice.

“This means the ecosystem is more open and more accessible,” George explains.

This underpins Huawei’s decision to launch a new offering based on the SuperPoD architecture.

SuperPoD is a computing system in which multiple computing nodes are tightly coupled through high-speed interconnect protocols featuring unified memory addressing across physical nodes but functioning like a single logical computer.

George explains that there are three forces driving SuperPoD adoption:

The first is post-training compute demand. After training, compute use can grow by as much as 10-times.

The second is multi-modal workloads, which are responsible for ever-higher compute and memory use.

The third is latency, where new use cases like AI earphones, glasses and phones are starting to demand sub-10 millisecond response times. “You cannot achieve this with traditional compute,” George points out.

The SuperPoD architecture addresses these issues with its high compute density, high interconnectivity bandwidth and low latency.

Importantly, the SuperPoD architecture is not just a blueprint: it has had real market validation, George says.

The new Atlas product matrix coming from Huawei scales from individual AI cards to enterprise data centres to large data centres.

This means users can scale up as needed without having to forklift their entire system, and with no forced migration to proprietary systems.

The flagship in Huawei’s new line-up is the Atlas 950 SuperPoD. This system empowers large-scale AI training and inference, enabling scenario-specific models with a good user experience. And it scales easily, so it can be sized to the actual workload.

The Atlas 950 is available in either water-cooled or cableless configurations.

Also coming in the next weeks is the Atlas 850E SuperPoD server, and enterprise-ready offering with flexible scaling for post-training and multi-tier inference.

“This is a serious machine in its own right, available in a conventional form factor,” George says. “The Atlas 850E is air-cooled, so can be deployed in standard data centres. And it can also scale as needed.”

Huawei will also launch an all in one appliance solution preconfigured for more than 60 scenarios. “This offers the fastest path from delivery to running AI workloads,” George says.

The Atlas AI offerings are backed by Huwaei’s global ecosystem that includes more than 2 000 ISVs (independent software vendors) in more than 20 industries, with more than 6 700 solutions.

The post Huawei shakes up AI with new server architecture appeared first on IT-Online.