Sunmmio Tech Raises $200M for 3D AI Inference Chips

Sunmmio Technology, a Chinese AI chip innovator, has successfully closed a significant funding round, amassing nearly 1 billion yuan (approximately $200 million USD). This substantial capital infusion, achieved through two distinct pre-Series A tranches, underscores strong investor confidence in the company's disruptive approach to AI processing. The latest round was spearheaded by Xiang He Capital, with participation from government-backed entities like CDB Capital. The preceding pre-Series A funding saw lead investments from Source Code Capital and Stony Creek Capital, alongside contributions from Lenovo Capital.

The company is channeling these funds into the development and mass production of its proprietary AI chips, aiming to establish a fully domestic supply chain. At the heart of Sunmmio's strategy is the ambitious goal to overcome the critical "memory wall" bottleneck that hampers current AI acceleration. By leveraging advanced 3D integration technology, Sunmmio seeks to redefine the performance benchmarks for AI inference, a segment projected to consume the vast majority of AI computing resources.

Sunmmio's founder, Dr. Wang Fuquan, a distinguished figure from the Chinese Academy of Sciences, identified a fundamental limitation in existing AI hardware: the disparity between computational power growth and memory bandwidth expansion. While processing cores have seen exponential improvements over two decades, memory access speeds have lagged significantly. This imbalance leads to substantial idle time for computational units, as highlighted by the fact that up to 70% of processing power in high-end chips like NVIDIA's H100 can be spent waiting for data during inference tasks.

The company's innovative 3D stacking approach vertically integrates memory modules directly atop processing cores. This architecture, utilizing hundreds of thousands of Through-Silicon Vias (TSVs), enables ultra-high-speed data transfer, achieving memory bandwidths of 16-32 TB/s. This represents a fourfold increase compared to the bandwidth offered by chips like NVIDIA's B200, promising a dramatic reduction in data latency for AI operations.

Sunmmio is strategically focusing on AI inference, differentiating itself from the general-purpose GPU market dominated by giants like NVIDIA. This focus is driven by market dynamics, with Dr. Wang predicting that inference workloads will account for over 90% of future AI computation. By specializing, Sunmmio can optimize its Application-Specific Integrated Circuits (ASICs) for power efficiency and cost-effectiveness, crucial factors for widespread inference deployment, particularly in edge computing scenarios.

The company's inference chip, codenamed 'A4', manufactured on a 12nm process, has demonstrated impressive performance. Benchmarks show it achieving 1.26 to 2.19 times the throughput of NVIDIA's H200 chip when running large language models such as Llama and Mixtral, despite the latter being built on a more advanced 4nm process. This performance leap is rooted in Sunmmio's prior expertise in developing highly efficient cryptocurrency mining hardware, where its 3D integration techniques in the JASMINER X4 server delivered exceptional power efficiency, outperforming competitors and generating significant revenue.

With a team comprising experts from prestigious institutions like the Chinese Academy of Sciences and Tsinghua University, Sunmmio emphasizes a global competitive vision over a purely domestic focus. The company aims to capture market share from established players and address the growing demand for on-premise AI solutions. Their emphasis on low power consumption, enabled by 3D integration, positions them favorably for the edge AI market, where energy efficiency is paramount.