Generative AI is becoming more common across industries as organizations seek to benefit from the technology. However, AI applications at the edge often face memory constraints, limiting the quality of the output
To address these challenges, Phison has leveraged its expertise in high-performance NAND flash technology with its aiDAPTIV+ solution, enabling a groundbreaking approach to expanding AI capabilities. By integrating SSDs as extended memory for Large Language Models (LLMs), Phison’s technology unlocks new possibilities for edge IoT and robotic devices, PCs, engineering workstations, and data center servers. These advancements empower devices to load larger models and achieve both fine-tuning and better inferencing results—capabilities previously unattainable.
For instance, Phison’s measurements reveal that PCs and newly release IoT devices from NVIDIA and others can now fine-tune models in the 1-8 billion parameter range, beyond merely performing inference. Moreover, using NVIDIA’s Jetson Orin Nano Super Developer Kit, aiDAPTIV+ extended token lengths by 16x, enabling better context comprehension and more accurate outputs. Specifically, it accelerated the time to the first token by 24x, enhancing user experiences by improving response time.
Tackling key AI challenges
Phison’s aiDAPTIV+ is also designed to resolve critical hurdles in AI inference applications:
-
-
- Breaking memory constraints: aiDAPTIV+ integrates NAND flash storage as extended memory, enabling training bigger Large Language Models (LLM), while significantly increasing token processing capacity and better enabling long-form content generation and understanding.
-
-
-
- Boosting inference efficiency: By optimizing resources, the solution enhances inference speeds, reducing response times while maintaining cost-effectiveness.
-
Partnering for innovation
To bring these advancements to real-world workloads, Phison has partnered with ADLINK Technology. Together, they’ve developed the ADLINK Technology DLAP Supreme Series, an edge generative AI platform powered by NVIDIA’s Jetson Orin Nano Super Developer Kit and Phison’s aiDAPTIV+ technology. This platform delivers high memory capacity and computing power at the edge without incurring significant hardware costs, making AI more accessible for various industries and applications.
From data centers to edge computing
Phison’s aiDAPTIV+ exemplifies the potential of innovative software and hardware integration to disseminate AI capabilities across computing levels—from edge IoT devices to data centers. Leveraging software from industry leaders like NVIDIA, Meta and Hugging Face, Phison is enabling practical, cost-effective solutions that enhance AI adoption across industries, government sectors, and academia.
The future of AI lies in intelligent automation of mundane tasks, freeing humans for more strategic and creative pursuits. Phison’s aiDAPTIV+ solution is a significant step forward in realizing this vision, paving the way for advanced, widespread generative AI capabilities.
Additional AI Content from Phison