IRVINE, California – June 12, 2025 – Fixstars Corporation (TSE Prime: 3687, US Headquarters: Irvine, CA), a leading company in performance engineering technology, is pleased to announce successful verification of its "Fixstars AIBooster" (hereinafter, AIBooster) on Sakura Internet Inc.’s newest GPU-equipped servers, "Koukaryoku PHY," featuring NVIDIA's latest H200 GPUs. The results demonstrated a remarkable 2.5-fold acceleration compared to servers equipped with NVIDIA’s previous H100 GPUs.
Maximizing "Koukaryoku PHY" Performance with AIBooster
AIBooster is designed to maximize GPU utilization efficiency and consistently deliver optimal performance
through two core functionalities: Performance Observability (PO) for detailed monitoring, and Performance
Intelligence (PI) for automatic optimization.
In collaboration with Sakura Internet, Fixstars has confirmed that AIBooster's performance observation and
improvement capabilities function effectively and achieve acceleration even on the newly launched H200
architecture GPUs within "Koukaryoku PHY," Sakura Internet's first bare-metal series for generative AI cloud
services.
Fixstars AIBooster and H200 Operation Verification Results
In the processing of large-scale generative AI models, it is known that GPU memory shortage frequently occurs, especially during training. For example, with the previous H100 model, to perform pre-training of a 70B-class model on 2 nodes, it was necessary to suppress processing speed and accuracy to conserve memory consumption. On the other hand, the H200 has enhanced memory compared to the H100, which was expected to alleviate such problems and improve processing speed and accuracy. Fixstars validated this expectation by applying AIBooster to the H200 GPUs, achieving a remarkable 2.5x performance acceleration compared to the previous H100 model. This corresponds to an approximately 60% reduction in processing time.
Operation Verification Details
- Training Method: Pre-training
- Target Model: Llama 3.1 70B
- Training Data: RedPajama-Data-1T arXiv
- Framework Used: Megatron-LM
Overview of H200 Acceleration Measures for Pre-training Processing by AIBooster
- Use of high-precision optimizer (SGD-SaI -> Adam)
- Use of high-speed arithmetic unit (fp16 -> fp8)
- Reduction of recalculation amount (recompute-granularity full -> selective, etc.)
Provision and Support for Fixstars AIBooster
"Koukaryoku PHY" subscribers of Sakura Internet will receive an AIBooster installation kit when they start
using the server. Since Fixstars provides verified installation and operation procedures on "Koukaryoku
PHY," users can easily start using it.
Additionally, Fixstars engineers provide installation support and GPU server maintenance support upon
request. For customers seeking even higher performance after commencing operation, GPU acceleration support
is also provided.
Fixstars AIBooster is available for free download, and online demos are provided to showcase its functionality in advance.
Fixstars AIBooster Details: https://www.fixstars.com/en/ai/ai-booster
About Fixstars Corporation
Fixstars is a technology company dedicated to accelerating AI inference and training through advanced software optimization solutions. It supports innovation in healthcare, manufacturing, finance, mobility, and other industries. For more information, visit: https://www.fixstars.com/
