Kathy Gibson reports from Pinnacle TechScape, Durban – AI is not a silver bullet for all applications – it is a layered approach and there are many nuances to the technology.
This is the word from JP De Villiers, brand lead of SuperMicro, who believes AI technology opens up new opportunities for South African companies.
“For instance, data centres are changing, and we are seeing the emergence of native GPU clouds offering AI as a service, GPU as a service or AI factory services.
“These are not necessarily traditional ICT companies, but have seen and grasped the opportunity from GPU.”
However, De Villiers points out that there are no cloud GPU providers in South Africa. “All of the companies offering these services are international – that revenue is going overseas.
“We would like to see people buying the equipment and hosting it in South Africa.”
Such a venture would most likely attract interest from local companies, De Villiers adds. “If you build it, they will come.”
SuperMicro is benefiting from the AI market growth and has become the top platform for Generative AI (GenAI) and LLM platforms, with 500% year on year growth in accelerated computing.
“In our DNA, we want to build the best hardware,” De Villiers says.
The company focuses on building block solutions to enable customer time-to-market advantage. It delivers rack-scale systems that let customers plug and play for onsite deployment and service. “We aim to be the one stop shop for total IT solutions incuding hardware, software and services,” according to De Villers.
Green computing is top of mind for SuperMicro, which offers high efficiency system designs with free air and liquid cooling that can save up to 40% TCO.
Globally, SuperMicro has identified a total addressable market of $120-billion in AI infrastructure, compared to a cloud market of $75-billion, storage infrastructure of $30-billion, and an IoT/5G and telco market of $30-billion.
SuperMicro has developed AI frameworks that can help accelerate AI adoption.
These frameworks include large scale AI training workloads, HPC/AI workloads, enterprise AI inference and training, visualisation, content delivery, virtualisation and AI edge.
“The framework for each is different, so the hardware needs to be tailored for the environment,” De Villiers explains.
SuperMicro places a large focus on its Nvidia GPUs, but also offers servers based on GPUs from Intel and AMD.
Intel offerings include the Flex, Max and Habana Gaudi 2 GPUs; from Nividia, SuperMicro uses Hopper, MGX Grace-Hopper and MGX Grace-Grace chips; and AMD GPUs include the MI300X and MI300A.
Data training is a big stumbling block for many AI implementations, so SuperMicro offers very large servers to perform their data processing and in training it.
Smaller 2u or 4u servers can be employed to fine-tune and optimise models for the target hardware accelerator.
The final step of inference, deployment and running the models, can be done on a variety of edge solutions.