The South Korean startup HyperAccel has launched mass production of its Bertha AI accelerator chip. The hardware is designed for data center inference tasks and is manufactured using Samsung Foundry's 4nm process node. The project was facilitated by the design service provider SEMIFIVE.
The Bertha 500 model features a die size of 500mm² and provides 768 TFLOPS of FP8 computing power. It supports various data formats, including FP16, FP8, FP4, INT8, and INT4. The chip integrates 256MB of on-chip SRAM and supports external connectivity to 128GB or 256GB of LPDDR5X memory.
The device comes in a dual-slot PCIe AIC form factor with a TDP of 250W. HyperAccel claims that the Bertha 500 offers significant performance improvements over the NVIDIA H100, specifically citing a 2x increase in throughput, 19x better cost-efficiency, and 12x higher energy efficiency.
