Arm Custom Chips Get a Boost from Nvidia Partnership

Arm announced on Monday that its central processing units (CPUs), based on its proprietary technology, will now be able to seamlessly integrate with Nvidia’s AI chips utilizing NVLink Fusion technology. This development is poised to streamline custom infrastructure solutions, particularly for hyperscalers seeking to leverage the combined power of Arm-based Neoverse CPUs and Nvidia’s industry-leading graphics processing units (GPUs).

This move underscores Nvidia’s strategic embrace of partnerships across the technology landscape, cementing its position at the heart of the rapidly evolving AI sector. The announcement signifies Nvidia’s willingness to broaden the NVLink ecosystem, fostering integration with a diverse range of custom chips rather than mandating the exclusive use of its own CPUs.

Currently, Nvidia offers the Grace Blackwell platform, a high-performance solution that pairs multiple GPUs with a specialized Nvidia-branded Arm-based CPU. Existing server configurations also support CPUs from Intel and Advanced Micro Devices (AMD), offering customers a degree of flexibility. However, the increasing demand for customized solutions is driving a shift towards greater control over hardware and associated costs.

Companies like Microsoft, Amazon, and Google are actively developing and deploying Arm-based CPUs within their cloud infrastructure. This strategic move enables finer-grained control over system architecture and optimizes operational expenses.

Arm, a key player in the semiconductor industry, focuses on licensing its instruction set architecture (ISA), which forms the foundation for CPU design. The company also provides pre-designed cores, allowing partners to accelerate the development and deployment of Arm-based processors. The company’s Neoverse platform is specifically targeted at datacentre and cloud applications, emphasizing power efficiency and scalability.

According to Monday’s announcement, future custom Neoverse chips will incorporate a novel protocol engineered to facilitate high-speed, low-latency data transfer with Nvidia GPUs. This enhanced communication layer is crucial for maximizing the performance of AI workloads, where data movement between the CPU and GPU is a critical bottleneck.

Traditionally, the CPU was considered the central component of a server architecture. However, generative AI applications have shifted the focus towards specialized AI accelerator chips, predominantly Nvidia GPUs. Modern AI servers can incorporate as many as eight GPUs paired with a single CPU, underscoring the GPU’s pivotal role in these compute-intensive environments.

This announcement follows Nvidia’s recent investment of $5 billion into Intel, highlighting the strategic importance of CPU integration within AI server infrastructure. This investment is aimed at fostering greater interoperability between Intel CPUs and Nvidia’s NVLink technology, further expanding the ecosystem and providing customers with more flexible solution options.

While Nvidia’s previous attempt to acquire Arm for $40 billion in 2020 fell through due to regulatory hurdles, the company maintains a keen interest in the Arm ecosystem. Softbank, the majority owner of Arm, recently liquidated its stake in Nvidia. Softbank is also backing the OpenAI Stargate project that plans to use Arm technology alongside Nvidia and AMD chips, showing a diverse ecosystem approach. This partnership underscores the enduring importance of Arm’s technology within the broader AI infrastructure landscape and signals a commitment to open innovation.