SoftBank (SFTBY) has enhanced its AI computing platform in Japan by installing approximately 4,000 NVIDIA (NVDA, Financial) Hopper GPUs. Initially launched in September 2023, the platform operated with around 2,000 NVIDIA Ampere GPUs, delivering a performance of 0.7 Exaflops. With the new upgrades, the platform's capacity has surged to about 4.7 Exaflops.
Looking ahead, SoftBank aims to further boost its platform's performance between the fiscal years of 2024 and 2025. By deploying NVIDIA's DGX B200 systems within the DGX SuperPOD, the company intends to increase the total number of installed GPUs to approximately 10,000. Ultimately, SoftBank's goal is to achieve a top computing capacity of 25.7 Exaflops in Japan through additional GPU installations.
Initially, this AI platform will be utilized by SB Intuition, a SoftBank subsidiary dedicated to developing native large language models specifically for the Japanese language. SB Intuition plans to build a large language model with around 390 billion parameters and support multiple modalities by the 2024 fiscal year. The enhancements in the AI platform will expedite advancements in these language models.