Announced details
Up to 43% reduction in AI training costs
Search time reduced to one-sixteenth of the previous level
After approximately 2 hours of searching, throughput increased from 78.2 to 118.6 TFLOP/s/GPU
Training time per epoch accelerated by approximately 1.52x
A similar approximately 1.7x speedup achieved in approximately one-150th of the previous time, down from approximately 300 hours
Throughput of 140.0 TFLOP/s/GPU after approximately 18 hours of searching
Training time accelerated by approximately 1.79x
In this announcement the company addresses none of Contract Value / Offering price / Impact on Business Performance.
| Item | Details |
|---|---|
| Names of Products / Services Covered | Fixstars AIBooster, Hyperparameter auto-optimization functionality, Heuristic search, Staged BlackBox search |
| Applicable area | Distributed training of large language models (LLMs), AI Training and Inference, AI service providers, Embedded applications using AI |
| Quantitative Effects | Up to 43% reduction in AI training costs, Search time reduced to one-sixteenth of the previous level, After approximately 2 hours of searching, throughput increased from 78.2 to 118.6 TFLOP/s/GPU, Training time per epoch accelerated by approximately 1.52x, A similar approximately 1.7x speedup achieved in approximately one-150th of the previous time, down from approximately 300 hours, Throughput of 140.0 TFLOP/s/GPU after approximately 18 hours of searching, Training time accelerated by approximately 1.79x |
Items this document does not state (15)
These are items the company does not state in this document. We do not leave them blank, nor fill them in from other sources or by inference.
The reasons the company gives
The company's own explanation, not something inferred from the numbers.Implemented two proprietary algorithms, heuristic search and Staged BlackBox search, by leveraging domain knowledge of Megatron Core's parallelization strategies. The company expects to reduce AI-related costs, including GPU acquisition costs and power consumption, by operating GPU assets more efficiently.