Korea's HyperAccel Launches 4nm AI Chip Bertha 500

Korea-based AI chip startup HyperAccel's data center AI inference accelerator chip Bertha 500 has entered mass production using Samsung's 4nm process, as announced by chip design service firm SEMIFIVE on September 8. The chip, measuring 500 square millimeters, integrates 256MB of on-chip SRAM cache, supports data formats like FP8, FP16, and INT8, and delivers 768 TFLOPS of computing power. It connects to 128GB or 256GB of LPDDR5X memory with 546GB/s bandwidth. Bertha 500 comes in a dual-slot PCIe AIC form factor with a 250W TDP. HyperAccel claims its throughput, cost efficiency, and energy efficiency are 2x, 19x, and 12x that of Nvidia's H100, targeting high-performance, cost-effective AI inference for data centers.