Nvidia’s Arm-based servers match x86 at handling AI workloads in data centers

With improvements in energy efficiency, performance, and an expanding software ecosystem, Nvidia’s Arm architecture finally appears set to handle AI inference workloads in data centres around the world. 

According to the latest MLPerf benchmark results shared by the company, Arm-based servers using Ampere Altra CPUs and NVIDIA A100 Tensor Core GPUs delivered a near-equal performance to similarly configured x86-based servers.  

Arm is a family of reduced instruction set computing (RISC) architectures for processors, configured for various environments. Meanwhile, x86 follows complex instruction set computing (CISC) architectures. Arm chips are relatively simple and most instructions execute in one clock cycle, while x86 ones are mostly complex, taking up multiple CPU cycles. They lead to better performance but with more power consumption. 

While handling AI inference jobs across six tests, the Arm systems came within a few percentage points of those running on x86, Nvidia said. In fact, in one test, dubbed 3D-Unet, Arm even outperformed the latter. AI inference, as Nvidia explains, is the process when a computer runs AI software to recognize an object or make a prediction. It uses a deep learning model to filter data, finding results no human could capture.  

MLPerf inference tests, from industry benchmarking group MLCommons, were based on the most popular AI workloads and scenarios such as computer vision, medical imaging, natural language processing, recommendation systems, and reinforcement learning. They give enterprises the confidence to make informed buying decisions for hardware and continue to be backed by executives from Alibaba, Arm, Baidu, Google, Intel and NVIDIA. 

“Arm, as a founding member of MLCommons, is committed to the process of creating standards and benchmarks to better address challenges and inspire innovation in the accelerated computing industry,” David Lecomber, a senior director of HPC and tools at Arm, said in a statement. 

“The latest inference results demonstrate the readiness of Arm-based systems powered by Arm-based CPUs and NVIDIA GPUs for tackling a broad array of AI workloads in the data center,” he added. 

In September last year, Nvidia had announced the $40 billion deal to acquire Arm from Softbank. The two companies are expected to work together in future with Nvidia adapting its current x86-based GPUs for ARM chips.