The race to trillion-parameter model training in AI is on, and this company thinks it can manage it for less than $100,000

Date:
Wed, 26 Mar 2025 21:24:00 +0000

Description:
Phison slashes AI training costs shifting workloads from GPUs to SSDs could cut trillion-parameter model expenses from $3 million to just $100,000.

FULL STORY ======================================================================Phisons SSD strategy slashes AI training costs from $3 million to $100,000 aiDAPTIV+ software shifts AI workloads from GPUs to SSDs efficiently SSDs could replace costly GPUs in massive AI model training

The development of AI models has become increasingly costly as their size and complexity grow, requiring massive computational resources with GPUs playing
a central role in handling the workload.

Phison, a key player in portable SSDs , has unveiled a new solution that aims to drastically reduce the cost of training a 1 trillion parameter model by shifting some of the processing load from GPUs to SSDs, bringing the
estimated $3 million operational expense down to just $100,000.

Phison's strategy involves integrating its aiDAPTIV+ software with high-performance SSDs to handle some AI tool processing tasks traditionally managed by GPUs while also incorporating NVIDIAs GH200 Superchip to enhance performance and keep costs manageable. AI model growth and the trillion-parameter milestone

Phison expects the AI industry to reach the 1 trillion parameter milestone before 2026.

According to the company, model sizes have expanded rapidly, moving from 69 billion parameters in Llama 2 (2023) to 405 billion with Llama 3.1 (2024), followed by DeepSeek R3s 671 billion parameters (2025).

If this pattern continues, a trillion-parameter model could be unveiled
before the end of 2025, marking a significant leap in AI capabilities.

In addition, it believes that its solution can significantly reduce the
number of GPUs needed to run large-scale AI models by shifting some of the processing tasks away from GPUs to the largest SSDs and this approach could bring down training costs to just 3% of current projections (97% savings), or less than 1/25 of the usual operating expenses.

Phison has already collaborated with Maingear to launch AI workstations powered by Intel Xeon W7-3455 CPUs, signaling its commitment to reshaping AI hardware.

As companies seek cost-effective ways to train massive AI models, innovations in SSD technology could play a crucial role in driving efficiency gains while external HDD options remain relevant for long-term data storage.

The push for cheaper AI training solutions gained momentum after DeepSeek
made headlines earlier this year when its DeepSeek R1 model demonstrated that cutting-edge AI could be developed at a fraction of the usual cost, with 95% fewer chips and reportedly requiring only $6 million for training.

Via Tweaktown You may also like These are the best AI website builders around Take a look at the best firewalls on offer right now Seagate teams with
Nvidia to build an NVMe hard drive proof of concept, more than 3 years after its last effort



======================================================================
Link to news story: https://www.techradar.com/pro/the-race-to-trillion-parameter-model-training-in -ai-is-on-and-this-company-think-it-can-manage-it-for-less-than-usd100-000


--- Mystic BBS v1.12 A47 (Linux/64)
* Origin: tqwNet Technology News (1337:1/100)