Longsys Redefines On-Device AI with Groundbreaking Edge Memory Solutions

TAIPEI, Taiwan – June 04, 2026 – The hum of servers in distant data centers has long been the soundtrack of the artificial intelligence revolution. But here at COMPUTEX 2026, against a backdrop themed "AI Together," semiconductor memory firm Longsys just unveiled a suite of technologies that could shift the industry’s center of gravity, moving powerful AI from the cloud to the device in your hand. With the global debut of its custom-built AIDIMM™ and AILPBGA™ memory solutions, the company is making a bold play to solve the most significant bottlenecks—memory, cost, and power—that have constrained the dream of truly potent, localized AI.

This isn't just another incremental hardware update. Longsys is presenting a vertically integrated vision for "Edge AI Storage," combining novel memory modules, sophisticated storage processing units, and intelligent software to enable everything from next-generation AI PCs to smarter mobile devices to run massive Large Language Models (LLMs) without a constant connection to the cloud. For business leaders and consumers alike, this signals a future where AI is faster, more private, and more seamlessly integrated into our daily lives.

The Memory Bottleneck Breached

At the heart of the announcement are two purpose-built memory products designed specifically for edge AI inference. The first, AIDIMM™, is a direct assault on the pain points plaguing today’s AI agent hosts and high-end AI PCs. These devices often struggle with insufficient memory, computational stuttering, and thermal throttling when tasked with running complex LLMs.

Longsys’s solution is a single memory module delivering up to 128GB of capacity and an industry-leading 307.2GB/s of single-channel bandwidth. To put that in perspective, the company demonstrated live on the show floor that a single AIDIMM™ module could reliably run an LLM with over 70 billion parameters—a task typically reserved for powerful server racks. More impressively, a live demonstration on a host powered by AMD's Ryzen AI Max+ 395 processor showed a staggering 397-billion-parameter model running locally with 128GB of DRAM. This leap is made possible by a native 256-bit bus and an architecture that radically reduces data transmission latency.

For device manufacturers (OEMs), the design offers a crucial advantage: backward compatibility. The module's high-density connectors are compatible with mainstream AI host motherboards, allowing for upgrades without costly and time-consuming redesigns. Coupled with proprietary power-tuning technology that dynamically adjusts voltage based on AI workload, AIDIMM™ promises a high-performance, low-power, and easy-to-upgrade path for the burgeoning AI PC market, which Gartner forecasts will exceed 143 million units in 2026.

The second innovation, AILPBGA™, targets the other end of the edge spectrum: compact, power-sensitive embedded devices. This tiny 22x22mm chip delivers up to 64GB of capacity and the same 307GB/s bandwidth, but it's engineered for cost-effectiveness and ultra-low power consumption. By offering multiples of the performance of standard LPDDR5x memory while remaining compatible with existing platforms, AILPBGA™ allows developers to integrate powerful AI into smaller devices without a complete system overhaul, shortening development cycles and significantly lowering adaptation costs.

Democratizing AI Through Software-Hardware Synergy

While the new memory modules provide the raw horsepower, Longsys’s true strategic advantage may lie in its integrated hardware-software stack. The company understands that simply throwing more memory at the problem is an inefficient and expensive solution. Instead, it has developed intelligent systems to optimize how that memory is used.

The cornerstone of this strategy is the SPU™ (Storage Processing Unit) paired with the iSA™ (Intelligence Storage Agent). The SPU acts as a dedicated processor for storage tasks, offloading the main CPU. Its integrated HLCache™ (High Level Cache) technology, according to company data, can reduce a device's reliance on expensive DRAM by nearly 40%. The iSA, meanwhile, serves as a "professional scheduling engine" for AI inference. It uses predictive algorithms and intelligent cache management to tackle the massive I/O demands of modern Mixture-of-Experts (MoE) LLMs, ensuring data is where it needs to be before the processor asks for it, dramatically improving smoothness and efficiency.

This combination effectively lowers the barrier to entry for advanced AI. A demonstration of HLCache™ technology in UFS storage for smartphones showed devices with lower-spec memory smoothly running 13B and 20B parameter AI models, delivering an experience comparable to more expensive phones with larger DRAM. By reducing the overall bill of materials (BOM) for OEMs, Longsys is positioning its technology to democratize on-device AI, making it a standard feature rather than a premium add-on.

This strategy extends to the consumer market through its globally recognized brand, Lexar. Celebrating its 30th anniversary, Lexar is set to be the primary vehicle for commercializing these technologies worldwide. The new AI-Grade Gen5 professional storage lineup, powered by Longsys's core architecture, will leverage Lexar’s established global distribution channels, including major retailers like Costco and Best Buy, to bring these advanced capabilities to AI PCs, smart cameras, and consumer robotics.

Reshaping the Ecosystem for a Localized AI Future

Longsys's announcements at COMPUTEX are more than just a product launch; they represent a strategic push to reshape the AI hardware ecosystem. For years, the industry has been defined by a dependency on centralized cloud infrastructure. By providing a viable, high-performance path for edge AI, Longsys is enabling a paradigm shift where storage is no longer a passive data repository but an active participant in the computational process.

The company's focus on high-speed storage media, like its PCIe Gen5 mSSDs, underscores this vision. These drives boast blistering read/write speeds of up to 11GB/s and 10GB/s, respectively, and integrate sophisticated thermal management like VC phase-change liquid cooling to sustain peak performance during intense AI workloads. This is critical for AI PCs, where the drive will be constantly accessed for model parameters and KV Cache data.

Early ecosystem adoption validates this approach. Longsys has already collaborated with AMD to optimize its host platforms and with Unisoc to improve performance on mobile chipsets. Its mature PCIe Gen4 mSSDs are already integrated into AI PCs and laptops from multiple major OEMs, earning strong market recognition. This full-stack approach, from the foundational memory chip to the consumer-facing SSD, provides a cohesive and powerful solution that competitors offering piecemeal components will find difficult to match.

As the industry pivots toward a hybrid model where cloud and edge AI coexist and complement each other, the companies building the foundational hardware for the edge will hold immense influence. With its comprehensive and deeply integrated portfolio, Longsys is positioning itself not just as a supplier, but as a core architect of the localized AI future.