Fixstars Corporation, a leading company in performance engineering technology, is pleased to announce that it has conducted operational testing of its performance engineering platform “Fixstars AI Booster” (hereinafter “AI Booster”) on a server equipped with the latest GPU – NVIDIA H200 SXM 141GB (hereinafter “H200”) – for Sakura Internet Inc.’s bare metal GPU cloud service “High Power PHY,” and achieved a 2.5 times speed increase over the previous model (H100).
Maximizing the performance of “high-powered PHY” with AI Booster
AI Booster is software that aims to maintain high performance at all times by optimizing the utilization efficiency of computing resources such as GPUs. It provides two functions, Performance Observability (PO) and Performance Intelligence (PI), that support the cycle of performance observation and improvement.
Fixstars, in collaboration with Sakura Internet, has confirmed that AI Booster’s performance monitoring and improvements work effectively on the H200 architecture GPU, which is newly offered as part of the company’s “High Power” cloud service for generative AI, the first in the bare metal series “High Power PHY.” This accelerates speeds.
Also Read: NTT Group, Getworks team up on AI-powered ICT platform
Fixstars AI Booster and H200 performance test results
It is known that when processing large-scale generative AI models, the problem of insufficient GPU memory often occurs, especially during training. For example, with the previous model, H100, in order to run pre-training of a 70B-class model on two nodes, it was necessary to reduce processing speed and accuracy to conserve memory consumption.
Operation verification details
Learning method Pre-learning
Applicable model: Llama 3.1 70B
Training data RedPajama-Data-1T arXiv
Framework used: Megatron-LM
Overview of H200 pre-training speed-up measures using AI Booster
Use of high-precision optimizer (SGD-SaI -> Adam)
Use of high-speed arithmetic unit (fp16->fp8)
Reducing the amount of recomputation (recompute-granularity full -> selective, etc.)
SOURCE: PRTimes
