- Published on
Korean Firm HyperAccel Begins Mass Production of 4nm AI Chip 'Bertha' for Data Center Inference
- Authors
- Name
- DeepHorus
South Korean AI chip startup HyperAccel has achieved a major breakthrough in its design efforts. The company's data center AI inference accelerator chip, named Bertha 500, has entered mass production at global leading foundry Samsung Electronics' 4-nanometer process node. This announcement was officially made on September 8 local time by South Korean chip design service company SEMIFIVE, marking a substantial advance in South Korea's capabilities in high-end AI chip design and manufacturing.
The Bertha 500 chip is HyperAccel's core product targeted at data center inference tasks. Its core computing unit area reaches 500 square millimeters, integrating powerful computational power with a high-bandwidth memory system. The chip supports multiple data formats including FP16, FP8, FP4, as well as INT8 and INT4, and a single chip can deliver up to 768 TFLOPS of FP8 computing power, designed to efficiently handle large-scale AI model inference workloads.
In terms of memory architecture, the Bertha 500 integrates 256MB of on-chip SRAM cache and can connect to external LPDDR5X memory with a capacity of 128GB or 256GB. Reports indicate that its memory bandwidth may reach 546GB/s, providing ample bandwidth assurance for handling data-intensive AI inference tasks. The chip has a thermal design power of 250 watts and adopts a dual-slot PCIe accelerator card (AIC) form factor, facilitating deployment in existing data center servers.
HyperAccel claims that compared to the currently mainstream NVIDIA H100 GPU on the market, the Bertha 500 holds significant advantages in several key metrics. Official data shows its throughput can reach twice that of the H100, with cost-effectiveness being 19 times higher, and energy efficiency reaching 12 times that of the latter. If these figures are verified in practice, it would imply a significant reduction in total cost of ownership and energy consumption in AI inference deployments.
The mass production of this chip is the result of collaboration within South Korea's semiconductor industry ecosystem. Design service company SEMIFIVE provided critical support throughout the process, helping startup HyperAccel successfully implement its chip design on Samsung's advanced 4-nanometer manufacturing process. This "design service + advanced foundry" collaboration model provides a viable path for emerging AI chip companies to enter the market quickly.
The launch of the Bertha 500 directly targets the fastest-growing inference market in current AI computing. As AI applications like large language models transition from training phases to large-scale deployment, the demand for high-performance, high-energy-efficiency inference chips is increasingly urgent. The mass production of this chip will provide data center operators with a new hardware option, potentially intensifying competition in the AI accelerator market.
Looking ahead, HyperAccel plans to further expand its product line. In addition to the data center-focused Bertha 500, the company also has a chip named Bertha 100, which is dedicated to AI deployment on edge devices. This indicates the company's strategic intention to cover the complete AI computing market from cloud to edge.
Analysts point out that although the performance data presented by HyperAccel is highly attractive, the chip's actual market performance still requires independent testing and large-scale deployment verification by customers. The competition in the AI chip market is exceptionally fierce, and new entrants need to prove their product's stability, software ecosystem compatibility, and long-term supply capability. The mass production of the Bertha 500 is an important step, but its true success will depend on the subsequent level of market acceptance.
