Phison’s Secret Weapon: SSDs Reshape AI Training

New Middleware Makes AI Accessible to More

Phison, a hardware manufacturer, is disrupting AI training by leveraging its proprietary middleware to expand the accessibility of high-performance computing. This innovative approach significantly reduces the cost barrier, democratizing access to advanced AI capabilities for a wider audience.

Revolutionizing AI Training with Aidaptiv+

Many organizations now want to develop artificial intelligence within their existing infrastructure. However, the expense and scarcity of GPUs present significant obstacles. Michael Wu, General Manager and President of Phison Technology Inc. (USA), explained that their solution, Aidaptiv+, uses economical flash storage to augment the costly VRAM, broadening AI training’s reach.

Overcoming AI and Memory Limitations

The increasing size and complexity of AI models pose significant challenges. According to Brian Cox from Phison, AI model parameters have been increasing rapidly, while the capacity of high-bandwidth memory (HBM) in professional GPUs has not kept pace. This often forces companies to buy more GPUs to compensate for limited memory, leading to underutilized resources.

“In terms of parameters, the complexity and size of AI models increase a 410 factor every two years. The capacity of HBM (High Bandwidth Memory) and other GDDR Memoirs (Graphics Double Data Rate, an ultra -performing, editor’s note), two extremely fast memories, but very expensive by professional GPU, has only progressed by one factor two over the same period.”

—Brian Cox, Product Marketing Director at Phison

The global AI market is projected to reach $1.81 trillion by 2030, according to a recent report (Fortune Business Insights, 2024), highlighting the pressing need for cost-effective solutions.

Necessity as the Mother of Invention

Phison’s development of Aidaptiv+ stemmed from an internal need to improve their own design and manufacturing processes. When the solutions they sought were too expensive, CEO KS Pua challenged the team to find a more cost-effective approach. The result was Aidaptiv+, a system that leverages SSDs for AI model training.

The SSD as a Lifebuoy

Aidaptiv+ works by storing AI models on larger, more affordable flash storage. Its secret ingredient, AidapTivlink Middleware, intelligently manages data transfer between the SSD and the GPU, keeping the GPU pipeline consistently fed with the necessary model slices.

The specialized hardware component, aidaptivcache, is designed to withstand the high write intensity of AI model training, supporting a high number of drive writes per day (DWPD). This is considerably more robust than standard corporate SSDs.

Cost vs. Speed Trade-Off

Brian Cox acknowledges that Aidaptiv+ involves a technological compromise. The training of a model will take longer compared to using a server farm with expensive GPUs. Phison’s tests show that training a 70-billion-parameter model on a four-GPU workstation fails without Aidaptiv+, but takes just over four hours with it. For many companies, universities, and research and development departments, this compromise is acceptable.

Phison aims to evolve from a consumer controller seller into a full-fledged storage partner. The company’s strategy involves designing and holding intellectual property of the controller, firmware, and hardware, to offer advanced personalization. Their SSDs have been used in the International Space Station ISS and are even heading to the moon.