- Phison SSD strategy cuts training costs from $ 3 million to $ 100,000
- Aidaptiv+ software changes the workloads of the GPUs to the SSDs efficiently
- SSD could replace the expensive GPU in training models
The development of AI models has become increasingly expensive as their size and complexity grow, which requires massive computational resources with GPUs that play a central role in the management of the workload.
Phison, a key player in Portable SSDs have presented a new solution that aims to drastically reduce the cost of training a 1 billion parameter model by changing part of the SSD GPU processing load, which reduces the estimated operating expenses from $ 3 million to only $ 100,000.
Phison strategy implies the integration of its Aidaptiv+ software with high -performance SSD to handle some processing tasks of the AI tools traditionally handled by GPU, while incorporating the NVIDIA GH200 surchip to improve performance and maintain manageable costs.
Growth of the AI model and the milestone of billions of parameters
Phison expects the AI industry to reach the milestone of 1 billion parameters before 2026.
According to the company, the model sizes have expanded rapidly, from 69 billion parameters in flame 2 (2023) to 405 billion with flame 3.1 (2024), followed by the 671 billion parameters of Deepseek R3 (2025).
If this pattern continues, a model of billion parameters could be revealed before the end of 2025, marking a significant leap in AI’s capabilities.
In addition, he believes that his solution can significantly reduce the number of GPUs necessary to execute large -scale AI models by changing some of the GPU processing tasks to larger SSDs and this approach could reduce training costs to only 3% of current projections (97% savings), or less than 1/25 of the usual operating expenses.
Phison already has He collaborated with Maingear to launch work stations promoted by the Intel Xeon W7-3455 CPU, indicating his commitment to remodel AI hardware.
As companies look for profitable ways to train massive AI models, innovations in SSD technology could play a crucial role in promoting efficiency gains while External HDD options remain relevant to long -term data storage.
The impulse for the cheapest training solutions gained impulse after Depseek was news earlier this year, when its Deepseek R1 model showed that the Avuardia AI could develop at a fraction of the usual cost, with 95% less chips and, according to reports, requiring only $ 6 million for training.
Via Tweaktown
