- Phison’s SSD technique slashes AI training prices from $3 million to $100,000
- aiDAPTIV+ software program shifts AI workloads from GPUs to SSDs effectively
- SSDs may change expensive GPUs in huge AI model training
The event of AI fashions has develop into more and more expensive as their measurement and complexity develop, requiring huge computational sources with GPUs taking part in a central function in dealing with the workload.
Phison, a key participant in transportable SSDs, has unveiled a new resolution that goals to drastically scale back the price of training a 1 trillion parameter model by shifting a few of the processing load from GPUs to SSDs, bringing the estimated $3 million operational expense down to only $100,000.
Phison’s technique includes integrating its aiDAPTIV+ software program with high-performance SSDs to deal with some AI software processing duties historically managed by GPUs whereas additionally incorporating NVIDIA’s GH200 Superchip to reinforce efficiency and maintain prices manageable.
AI model development and the trillion-parameter milestone
Phison expects the AI business to succeed in the 1 trillion parameter milestone earlier than 2026.
In accordance with the firm, model sizes have expanded quickly, shifting from 69 billion parameters in Llama 2 (2023) to 405 billion with Llama 3.1 (2024), adopted by DeepSeek R3’s 671 billion parameters (2025).
If this sample continues, a trillion-parameter model may very well be unveiled earlier than the finish of 2025, marking a vital leap in AI capabilities.
As well as, it believes that its resolution can considerably scale back the variety of GPUs wanted to run large-scale AI fashions by shifting a few of the processing duties away from GPUs to the largest SSDs and this method may convey down training prices to only 3% of present projections (97% financial savings), or lower than 1/25 of the normal working bills.
Phison has already collaborated with Maingear to launch AI workstations powered by Intel Xeon W7-3455 CPUs, signaling its dedication to reshaping AI {hardware}.
As corporations search cost-effective methods to coach huge AI fashions, improvements in SSD know-how may play a essential function in driving effectivity features whereas exterior HDD choices stay related for long-term information storage.
The push for cheaper AI training options gained momentum after DeepSeek made headlines earlier this 12 months when its DeepSeek R1 model demonstrated that cutting-edge AI may very well be developed at a fraction of the normal price, with 95% fewer chips and reportedly requiring solely $6 million for training.
Through Tweaktown
You might also like
Source link
#saving #OPEX #training #1trillion #parameter #model #sounds #ambitious #Phison #claims
