SK Group and Broadcom Form 'One Team' for AI Chip Co-Design
SK Group, the parent company of memory maker SK hynix, and Broadcom have formalized a "one team" initiative to co-design next-generation AI chips and memory solutions. The partnership aims to accelerate the integration of advanced technologies like HBM4 memory and CXL interconnects, signaling a move toward tighter vertical integration between compute, memory, and networking hardware for AI workloads.
- This partnership is part of a broader industry trend where hyperscalers and chip designers are moving away from general-purpose GPUs toward custom-designed ASICs for specific AI workloads like inference, in order to optimize performance and reduce costs. Broadcom has emerged as a key enabler in this space, co-designing custom AI accelerators for major tech companies including Google, Meta, and OpenAI. - The collaboration aims to incorporate SK hynix's HBM memory technology into the early design stages of Broadcom's AI chips, a move intended to maximize system performance through simultaneous optimization. This follows a series of meetings between SK Group Chairman Chey Tae-won and top executives from major U.S. tech firms, including Broadcom CEO Hock Tan, to solidify AI semiconductor alliances. - SK hynix has accelerated its HBM roadmap, with mass production of HBM4 slated for 2025 and HBM4E for 2026, aligning with the faster one-year release cycles for new AI accelerators from companies like NVIDIA. The company recently announced the completion of HBM4 development, which doubles the I/O to 2,048 and improves power efficiency by over 40% compared to the previous generation. - Broadcom's CEO, Hock Tan, projects that the market opportunity for its custom AI chips and networking components could reach between $60 billion and $90 billion by 2027. The company's AI-related revenue saw a 220% year-over-year growth in its 2024 fiscal year, reaching $12.2 billion. - The partnership will heavily leverage Compute Express Link (CXL), an open-standard interconnect designed for high-speed, low-latency communication between CPUs, GPUs, accelerators, and memory. CXL is critical for creating shared memory pools, which allows for more flexible and efficient resource allocation in AI data centers, moving beyond the limitations of traditional PCIe interconnects. - This alliance positions both companies to better compete in the rapidly growing AI accelerator market, which is projected to grow from over $39 billion in 2025 to nearly $500 billion by 2033. While NVIDIA currently dominates the GPU segment, the market for custom ASICs is expected to see the fastest growth. - SK Group is also diversifying its role in the AI ecosystem beyond memory manufacturing, with plans to invest $10 billion to become an "AI solutions company" and developing South Korea's largest AI data center. This strategy includes backing other South Korean AI chip makers, such as the recently merged Sapeon Korea and Rebellions, to challenge global market leaders. - The co-design strategy reflects a larger trend of vertical integration in the semiconductor industry, where companies control more of the technology stack from silicon to software to optimize system performance and capture more value from the end customer.
