Analyst(s): Alastair Cooke
Publication Date: May 30, 2025
What is Covered in this Article:
- Phison aiDAPTIV+ enables cost-effective running of large AI models on modest GPUs
- The combination of the aiDAPTIVCache hardware and aiDAPTIVLink drivers runs unmodified PyTorch applications
- Training is vital to the successful adoption of AI into an organization
The Event – Major Themes & Vendor Moves: AI Infrastructure Field Day is a semiannual, invitation-only event held in Santa Clara, organized by Tech Field Day. Independent industry experts join presenting companies to learn about product innovations. The event is live-streamed, and then videos are published on the Tech Field Day YouTube channel.
Phison aiDAPTIV+ is Cost-Effective With Large Models to Enable Generative AI
Analyst Take: Using a fast SSD as a tier to expand usable RAM is not a new concept; most operating systems have swap files to enable more system memory than the installed RAM. Phison aiDAPTIV+ is a set of hardware and software that enables the same concept for GPU memory, making AI cost-effective with large models. The high-bandwidth memory in a GPU is a significant part of the GPU’s price. As AI models become larger, the need for more memory follows. The fastest option is still to keep the entire model and data in GPU memory, which requires high-end GPUs for large models and often involves using cluster GPUs to accommodate the whole model. Phison aiDAPTIV+ is not intended to address use-cases that require the highest possible training throughput or the lowest inference latency, where the performance justifies the cost of the GPUs. Phison aiDAPTIV+ will deliver cost-effective results in use cases where lower throughput or longer response latency is acceptable for a significant cost saving. For example, for a 66% cost reduction, a weekly task of fine-tuning a model might be completed overnight, rather than in two hours. The aiDAPTIV+ solution comprises the aiDAPTIVcache hardware and aiDAPTIVlink drivers, with the aiDAPTIVProSuite as an optional AI software development environment.
aiDAPTIVcache
The SSDs Phison uses in AI acceleration have custom firmware designed to extend the SSD’s lifespan, exceeding that of the PC or embedded device where it is installed. Available in M.2 form at up to 320 GB for laptops and PCs, and in U.2 form at up to 8 TB for workstations and servers, the SSDs are part of the Pascari range, which Phison recently introduced directly to the market. The optimized firmware provides 100 DWPD (drive writes per day), allowing both high performance and high endurance for the most demanding generative AI workloads. These SSDs can be installed in embedded devices, such as the NVIDIA Jetson Nano, or laptops and desktops with desktop-class GPUs, making these devices cost-effective with large models
aiDAPTIVlink
The real magic lies in the aiDAPTIVlink software layer, which sits between the standard PyTorch library and the combination of SSD and GPU. The software manages moving blocks of data between GPU memory and the SSD as needed, enabling unmodified PyTorch applications to utilize larger models and more contextual data. Not requiring application changes makes aiDAPTIV+ a simple option to deploy compared to reducing model size through quantization or buying more expensive GPUs.
aiDAPTIVPro Suite
The whole aiDAPTIV+ solution originated from Phison’s internal challenges with adopting generative AI within the business. The aiDAPTIVPro Suite is a graphical tool for building LLM-based training and inference; it is an optional component in aiDAPTIV+. It reflects Phison’s understanding that training is crucial for enabling AI adoption. Phison has a training program for bringing staff up to speed with AI, delivered by select partners for their end-customers who will deploy aiDAPTIV+.
What to Watch:
- Phison aiDAPTIV+ is not intended as a general solution for running large generative AI models, only for situations where the lower performance is justified by lower cost.
- Other SSD vendors have demonstrated similar techniques, although not yet as integrated as the Phison solution. We expect to see more solutions that address the high cost of GPUs and the disparity between the price of GPU memory and SSD storage.
- Moves towards smaller language models for specific purposes may reduce the need for this memory tiering with GPUs during inference. The need for more memory during training is unlikely to change significantly, as quantization to reduce model size typically occurs after training.
The Phison presentations at AI Infrastructure Field Day are available on their appearance page. You can watch all the presentations from the four days of AI Infrastructure Field Day on the Tech Field Day website.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Other insights from Futurum:
Organizations Face Triple Threat in Generative AI: Privacy, Talent & Compliance
DeepSeek Disrupts AI Market with Low-Cost Training and Open Source, Yet Many Questions Loom
Channel Partner Ecosystem is Ready to Capitalize on Generative AI
Author Information
Alastair has made a twenty-year career out of helping people understand complex IT infrastructure and how to build solutions that fulfil business needs. Much of his career has included teaching official training courses for vendors, including HPE, VMware, and AWS. Alastair has written hundreds of analyst articles and papers exploring products and topics around on-premises infrastructure and virtualization and getting the most out of public cloud and hybrid infrastructure. Alastair has also been involved in community-driven, practitioner-led education through the vBrownBag podcast and the vBrownBag TechTalks.
