South korean AI chipmaker HyperAccel launches mass-produced 4nm bertha 500 for data center inference

On September 9, IT Home reported that SEMIFIVE, a South Korean chip design services firm, announced that HyperAccel, a local AI chip startup, has begun mass production of its data center AI inference accelerator chip, Bertha 500, using Samsung's 4nm process. The chip, which measures 500mm², is designed for large-scale data center applications, while its smaller counterpart, Bertha 100, is intended for edge deployment. Bertha 500 delivers 768 TFLOPS of FP8 performance, supporting FP16/8/4 and INT8/4 formats, with 256MB of on-chip SRAM and 128GB or 256GB of LPDDR5X memory. It operates at 250W TDP and uses a dual-slot PCIe AIC form factor. HyperAccel claims it offers twice the throughput, 19 times the cost efficiency, and 12 times the energy efficiency of NVIDIA's H100 chip.