NetApp and SK Telecom have completed a joint proof of concept (PoC) to achieve AI performance close to that of physical servers even in a virtualized environment, thereby reducing the inevitable performance degradation in virtual machines and resolving the constraints on operating large-scale AI workloads.

NetApp announced on the 24th that it has completed a PoC based on the NetApp AFX system and SK Telecom's 'Petasus AI Cloud'.

By utilizing NVIDIA GPU Direct Storage (GDS), over 99% of the performance of an environment using physical servers directly was achieved in a virtual machine environment.

Previously, running AI workloads on virtual machines resulted in reduced performance due to the additional processing burden of resource allocation, acting as a limiting factor in industries requiring ultra-low latency, such as Electronic Design Automation (EDA), finance, manufacturing, and telecommunications.

The two companies reduced the performance gap between virtualized environments and physical servers by optimizing the software stack and infrastructure design.

In the PoC, both the virtual machine and the physical server achieved 32.7GB/s, and CPU usage was reduced by 40–50%.

Both companies explained that they increased computational efficiency by reducing unnecessary processing burdens, allowing GPUs to focus on learning and inference.

NetApp AFX is structured to independently scale performance and capacity, and provides data management, security, and resilience features based on the operating system ONTAP.

In this PoCSeo confirmed that AFX can serve as a high-performance, expandable storage space that complements GPU memory by storing the KV cache, which is data temporarily referenced by AI while generating answers, in NetApp AFX.

Both companies stated that they rapidly processed large-scale data I/O during the inference process using a high-bandwidth, ultra-low-latency separated storage structure.

NetApp has been collaborating with SK Telecom as a storage partner for over 10 years, and based on this achievement, plans to continue cooperation, including the integration of AI data center solutions and joint response to enterprise AI customers.

PD Prasad, Executive Vice President of AI Data Infrastructure at NetApp, said, “By bridging the performance gap between physical servers and virtual machines, enterprises can perform faster and more efficient AI training and inference in a cloud environment.”

Jeong Min-young, Head of AI DC Solution at SK Telecom, said, “We succeeded in significantly reducing factors that degrade performance in virtualization environments and verified Petasus AI Cloud as a commercial-grade platform for next-generation AI workloads.”