China's Growing Shift from HBM to DDR5 in AI Semiconductors

China's Growing Shift from HBM to DDR5 in AI Semiconductors

China's Growing Shift from HBM to DDR5 in AI Semiconductors

China's Growing Shift from HBM to DDR5 in AI Semiconductors

China is rapidly intensifying its semiconductor ambitions by transitioning from High Bandwidth Memory (HBM) to DDR5 memory in artificial intelligence (AI) chip applications. This strategic move addresses both the challenge of U.S. export restrictions on advanced HBM and the surging demand for high-performance memory in AI workloads, stimulating strong domestic DDR5 production growth.

Context: China's Semiconductor Ambitions Amid Export Restrictions

China’s semiconductor industry has seen brisk growth in foundational chip manufacturing capacity, with local production rising from 19% of global wafer capacity in 2015 to 33% in 2023. However, China still lags in producing advanced chips and the sophisticated equipment required for cutting-edge semiconductor fabrication, primarily due to export controls and technology restrictions from the U.S. and its allies.

Specifically, China faces significant hurdles importing advanced GPUs and HBM memory modules, which are critical for high-end AI computations. As a result, Beijing has enforced policies directing data centers to replace or cancel foreign chip purchases, pushing the domestic industry to innovate around these constraints.

The Shift from HBM to DDR5 in AI Chips

HBM is known for its exceptionally high bandwidth, energy efficiency, and 3D-stacked architecture, making it ideal for AI training workloads that require massive data throughput. However, HBM is also expensive and subject to export restrictions. DDR5, meanwhile, offers a cost-effective alternative with growing performance capabilities, including speeds up to 6.4 GT/s and wider adoption potential.

According to industry experts, China is increasingly replacing HBM with DDR5 in AI chips to circumvent supply limitations. For example, Huawei unveiled its Ascend 910B AI chip using DDR5 instead of HBM, achieving performance comparable to Nvidia's A100 chip with HBM. This substitution reflects a pragmatic approach amid U.S. export controls on advanced memory components.

Technical Comparison: HBM vs. DDR5

  • Bandwidth: HBM offers significantly higher memory bandwidth (6.4–9.8 Gbps per pin) than DDR5 (up to 6.4 GT/s), crucial for AI training workloads.
  • Architecture: HBM uses vertically stacked DRAM dies with ultra-wide data channels, optimizing bandwidth and energy efficiency. DDR5 has narrower interfaces but benefits from larger capacity and lower cost.
  • Cost and Availability: HBM is more expensive and harder to source due to export restrictions, while DDR5 is more widely produced domestically in China, supporting large-scale deployment.

Implications for China's Semiconductor Industry

This pivot to DDR5 aligns with China’s broader strategy to boost self-reliance in semiconductor technology by expanding domestic production capacity and reducing dependency on foreign suppliers. The surge in DDR5 demand has fueled a memory supercycle, intensifying China's participation in the global memory market.

China’s leading foundries, such as SMIC and HHGrace, continue to grow steadily, with significant focus on AI and automotive semiconductors, analog devices, image sensors, and power management chips. Despite challenges in advanced-node fabrication (e.g., 7nm chips produced with older deep ultraviolet lithography), China is leveraging mature-node capacity expansion to support growing AI chip demand.

However, while DDR5 can compensate for some limitations in memory bandwidth compared to HBM, it may not fully match HBM's efficiency for the most demanding AI training tasks. Industry analysts expect that for inference workloads and cost-sensitive applications, DDR5 and other memory types like GDDR6 and LPDDR will gain prominence, whereas HBM will remain relevant in specialized, high-performance segments where bandwidth is paramount.

Future Outlook

China’s semiconductor market is projected to grow robustly, reaching an estimated US$206.7 billion in 2025 with steady growth expected through 2030. The memory supercycle driven by AI and data center expansion is a key factor. China’s approach of replacing restricted foreign components with domestically produced DDR5 memory illustrates a pragmatic adaptation to geopolitical constraints and market demands.

While HBM5 and future memory technologies promise even higher performance, China’s current focus on DDR5 for AI chips highlights the ongoing trade-offs and innovation efforts within its semiconductor ecosystem.

Written by Deepak Periyasamy.

Call to action: Visit TNT PCB for PCB fabrication and instant quotes and TNT E-Comp to order electronic components online (a THANSIV group).

ब्लॉग पर वापस जाएँ