英伟达GPU，在这个市场吃瘪

Core Viewpoint - Nvidia is shifting its focus towards the low-end market in the telecom sector, promoting its ARC-Compact chip for distributed RAN, which is less powerful than its previous offerings but is marketed as cost-effective and energy-efficient for low-latency AI workloads [1][2]. Summary by Sections Nvidia's Strategy - Nvidia has not abandoned its efforts to sell AI chips to the telecom industry, despite limited interest so far [1]. - The ARC-Compact is designed for installation at cell sites, contrasting with the previous ARC servers aimed at centralized RAN [1]. Technical Specifications - The main components of ARC-Compact include the Grace CPU and L4 Tensor Core GPU, which are lightweight and suitable for edge video processing but lack the capability for large language model training [2]. - Nvidia describes ARC-Compact as an "economical and energy-efficient" option for low-latency AI workloads and RAN acceleration [2]. Market Competition - Major RAN suppliers like Ericsson, Nokia, and Samsung have invested in virtual RAN technology but show limited interest in adopting Nvidia's CUDA for RAN development [4]. - These suppliers prefer a "lookaside" virtual RAN model to maintain hardware independence, keeping most software on the CPU [4]. Supplier Insights - Ericsson has successfully migrated software for Intel x86 CPUs to Grace with minimal changes, indicating potential for GPU use only in specific tasks like forward error correction (FEC) [5]. - Samsung has tested its software on Grace but denies the need for inline accelerators, suggesting that CPU capacity will suffice as technology advances [5]. Nokia's Position - Unlike Ericsson and Samsung, Nokia has invested all its virtual RAN resources into inline acceleration but acknowledges that its first-layer accelerator comes from Marvell Technology, not Nvidia [6]. Industry Perception - A survey by Omdia revealed that only 17% of respondents believe most AI processing will occur at base stations, with 43% favoring end-user devices [8]. - The telecom industry appears to be in a challenging position between device capabilities and large-scale cloud platforms, with low demand for ultra-low latency services in medium-sized countries [9]. Future Outlook - The emergence of Grace is timely as doubts about Intel's future as a virtual RAN CPU provider grow, allowing RAN suppliers to demonstrate independence from underlying hardware [9]. - There is a potential shift in AI processing focus from GPUs to more powerful CPUs, as model sizes decrease and machines handle critical AI workloads [10].